In early brainstorming sessions that would later form the basis of Yelpboxd (now in limited beta), Jenny provided empirical evidence of the superiority of Impossible™ Nuggets as an AMC Theatres concession and drew branding parallels between AMC and Chili’s. Alamo Drafthouse, we agreed, might pair better with Applebee’s; Regal Cinemas with Dave & Buster’s; but we soon realized there was no industry standard for chain restaurant-movie theater pairings. These data points, combined with an ill-fated endeavor to get through a complete Benihana experience before a Pillion screening at the AMC Kabuki (we couldn’t say no to framed photographs with our tablemates! we were very late to the movie), motivated our research into a comprehensive recommendation system for the elusive “dinner and a movie” that would withstand the tests of time, traffic, and temperament.

We built Yelpboxd to formulate “dinner and a movie” as a constrained optimization problem and recommend personalized plans addressing user needs across specified film, restaurant, and planning preferences:

Problem statement

We established a toy problem of matching chain restaurants to movie theaters by branding similarity, but we realized the actual gap in the market was much larger: how would we design a recommendation engine that could suggest optimal plans tailored to a user’s restaurant and film preferences while giving them enough time to get from dinner to movie? If Yelp was the people’s restaurant review platform and Letterboxd (“the social network for cinephiles”) the same for film, Yelpboxd begged to be built for the combined restaurant-and-film industry.

Farm-to-table recommender systems

“Recommender systems” encapsulate such a wide range of technologies that it can seem fruitless to try to define them in a few sentences, but one useful categorization is algorithmic- vs. human/socially-driven recommendations. Netflix, YouTube, TikTok, and Spotify lean algorithmic: they build a model of your viewing and listening “taste profile” as you engage with and spend time on their platform and suggest content you might like based on inherent features (content-based filtering) and your activity as compared to other users’ (collaborative filtering). Letterboxd and Yelp fall somewhere closer to organically- and socially-driven. Due to higher time and cost for users to dine at restaurants and watch movies and fewer interactions tracked natively on these platforms (you don’t post reviews on Yelp or log movies on Letterboxd at the rate you might interact with music on Spotify or scroll through TikTok), pair-wise user comparison data is inherently limited.

Still, this shouldn’t be a technical limitation preventing a platform like Letterboxd, which maintains user watch logs, ratings, reviews, and film data (genre, director, cast, etc.), from training and exposing a recommendation system using both content-based and collaborative filtering techniques. This means that Letterboxd is making an active choice not to maintain an algorithmic recommendation system, preferring to limit users to leverage social ratings and browse curated lists. As I previously covered in Scaling laws for personalized recommendation models and Alignment in recommender systems, this style of more human-, less machine-driven recommendation aligns with the ethos of encouraging natural information spread and diversity over isolation and segmentation: “we view the whole of Letterboxd as one big, organic recommendation engine”.

We take a similar “farm-to-table” approach to building the Yelpboxd recommender, prioritizing organic value to our customers (“would you use this? does it help you plan dinner and movie nights that are actually successful?”) over optimizing for monetization or data collection opportunities that would “improve” the recommendation model by e.g. YouTube measuring user engagement via click through rate and watch time metrics, but not necessarily capture enjoyment and logistic success rate of Yelpboxd users.

Prototyping Yelpboxd v0.0

• Seed prompt: “build a website called YELPBOXD that recommends dinner & a movie plans for the user based on their location and time of day (to see what restaurants & showtimes they could reasonably get to locally)”

  • We’d expect a functional UI but no integration with Letterboxd, Yelp, or any recommender model from this.

  • Indeed, Claude built a functional UI informed by Yelp and Letterboxd (“Yelp red + Letterboxd dark/green-orange-blue aesthetic”) without additional style guidance from me. However, because Claude knew I had a “free-first preference”, it made a decision to build a UK-based app because it could find more free cinema showtime APIs in the UK. UK localization was not desirable, because Yelpboxd’s first users were primarily based in the US.

  • Dangerously, Claude listed “simulated showtimes” in this version of Yelpboxd. That meant you could theoretically get dinner and then show up to a theater to a nonexistent showtime. This was an unacceptable user experience.

• Find an alternative to UK movie theater chain-specific showtime APIs with reasonable coverage in the US. Claude was able to find a free Cineworld UK API endpoint to serve UK showtimes, but in the US, the only free solution it could find was SerpApi (a limited Google Search API scraper).

  • Google sued SerpApi last year for “circumventing security measures protecting others’ copyrighted content that appears in Google search results”, but Claude seemed to have no issue suggesting it to me as a “genuine option”. Even so, nothing is truly free in this world: SerpApi limits free plan users to 250 searches per month, which Claude claimed it could stretch via Cloudflare Worker caching. I generated a SerpApi key and deployed it as a Cloudflare Worker Secret.

  • SerpApi only provided showtimes as pulled from Google. (Apparently, Google used to provide If we wanted to add runtimes, Claude suggested TMDB (The Movie Database), for which I similarly generated a developer key and deployed it as a Cloudflare Worker Secret.

• Yelpboxd’s free-first philosophy limits integration with Yelp and Letterboxd. Yelp limits OAuth authorization to contracted partners and its Places API to plans with only a 30-day free trial, so continued access beyond that that would not scale for a free-first Yelpboxd prototype. Letterboxd also limits its API (“not granting access for data-analysis, visualization or recommendation projects, for LLM or GPT-related use, for private or personal projects”) but Claude pointed out that Letterboxd does expose RSS feeds by username.

  • By asking a Yelpboxd user to optionally input their Letterboxd username, we can pull their public RSS feed including their watchlist and ratings. Claude suggested a basic personalization filter which could hide already-watched films, boost films on your watchlist, and bias recommendations to films with genres (cross-checked with TMDB) you’ve historically rated highly.

• Localize distance conventions by country and transit options by city size. Claude starting with UK-based theaters and assuming driving as mode of transit revealed two optimizations required for Yelpboxd’s primary users:

  1. localize distance conventions, e.g. imperial vs. metric,

  2. prioritize walking and public transit in big cities (“transit-aware”).

Yelpboxd v0.0 UX walkthrough

At this point, we had a working v0.0 Yelpboxd prototype we could user test.

Landing page

Yelpboxd uses a standard search landing page reminiscent of Yelp’s, with options to import your film log from Letterboxd and options to set when you’re planning your dinner and movie, how far you’re willing to travel, and cuisine preference.

Personalized results

Using the options toggled above, we got these recommendations for tonight.

The Good:

• Yelpboxd v0.0’s best match suggestion was a workable dinner and movie plan that was within the user’s time request (7pm is the default for “tonight”), cuisine preference (Mediterranean), and a legitimate showtime of a film the user had not seen (Toy Story 5).

The Bad:

• The second suggestion was not going to work. Pharaohs is permanently closed. Yelpboxd was suggesting a plan worse than our motivating Benihana and AMC Kabuki Pillion combination: the dinner did not exist.

  • This is because Yelpboxd uses OpenStreetMap for free restaurant data, which is crowdsourced and thus prone to errors and outdated information. Yelp integration would have avoided this problem, but we do not currently have Yelp API access as explained above. Bug #1: Remove permanently closed restaurants from results.

The Ugly:

• The “Plan For” dropdown mapped “Right now” to the user’s current time, “Tonight” to 7pm today, and allowed the user to select a different time today, but did not provide any way to plan for a future date. Improvement #1: Add a date picker for future plans.

• We pulled restaurant logos from their websites when available because OSM restaurant images were even sparser, but this led to strange formatting e.g. SAJJ logo cropped arbitrarily. Improvement #2: Find other sources of restaurant images.

• Yelpboxd exclusively recommending Toy Story 5 was suspicious. To be fair, we were testing a “nearby” case with location set to SFMOMA, so expected to be shown showtimes at the AMC Metreon. Still, AMC is a major theater and is not limited to Toy Story 5; Toy Story 5 is not in the user’s watchlist and should not be boosted up in recommendations; the user’s preferred genres as detected by Yelpboxd were “crime/thriller” which is not Toy Story 5’s classification. Improvement #3: Diversify the recommender to ensure it doesn’t fall into a Toy Story 5 trap without reasonable explanation.

• Yelpboxd was randomly picking restaurants based on proximity and cuisine selection without considering reviews or similarity to the user’s dietary preferences. This was expected because we could not find a way to integrate with a user’s Yelp or dining preferences for free. But we could at least introduce a basic user input and feedback loop for dining. Improvement #4: Add an option to not show a user certain restaurants.

Improvements made in Yelpboxd v0.1

You can try Yelpboxd out for yourself for free (limited beta). We are currently limited by SerpApi’s free tier to 250 searches per month, so Yelpboxd will not return useful results after we hit that. For reference, developer testing has used ~60 of those 250 searches so far this month.

• (Fixed) Bug #1: Remove permanently closed restaurants from results. We knew that permanently closed restaurants like Pharaohs could still be listed in OSM, and without access to Yelp’s more accurate and updated API, we had to find a different way to detect restaurants that no longer exist. Claude suggested OSM lifecycle checks to detect stale data and running a DNS check on restaurants' domains, which flagged Pharaohs as nonexistent. We shipped this and Pharaohs no longer appears in results.

• (Fixed) Improvement #1: Add a date picker for future plans. We addressed the inability to plan for a future date seen in v0.0 by adding a calendar date and time picker in v0.1.

• (Semi-fixed) Improvement #2: Find other sources of restaurant images. Claude proposed Wikidata and Wikimedia Commons geosearch near the restaurant (with name checking to prevent neighboring images from surfacing) as fallback sources to the official restaurant logo when available, which was a reasonable tradeoff to increase restaurant image coverage. We still display a default emoji when no image is not available, but overall restaurant image coverage is higher.

• (Semi-fixed) Improvement #3: Diversify the recommender to ensure it doesn’t fall into a Toy Story 5 trap without reasonable explanation. Upon further investigation, the “Toy Story 5 trap” (where Yelpboxd exclusively recommended Toy Story 5 if the user requested plans close to the AMC Metreon at sufficiently late times) was actually linked to multiple recommender inefficiencies:

  • The Yelpboxd planner caps itself to reduce SerpApi calls. Claude initially tried to self-limit calls to SerpApi by capping searched restaurants at 22 per search radius, 3 cinemas at a time, 2 plans explored per showtime, showing a maximum of 6 plans. We raised these caps to expand the possible plans explored and displayed, but the tradeoff is higher SerpApi calls, and there is still a chance of falling into the Toy Story 5 trap if the search and time radius is not high enough.

    

  • Yelpboxd makes assumptions about users’ dining speed. One reason why Toy Story 5 kept dominating Yelpboxd recommendations was that it has the highest number of possible showtimes at the AMC Metreon and late showtime availability in general, so was most likely to satisfy the constraint that users needed to finish their dinner before heading to the theater. Indeed, longer than expected dining time at the Japantown Benihana before an AMC Kabuki showtime was the motivating problem behind Yelpboxd in the first place, and Yelpboxd was already factoring in time variation in dining style (more time for fine dining vs. fast casual), but we needed to give the user some option to adjust this. We added a Dinner Pace (Quick bite, Normal, Leisurely) to modify the default assumptions and open up additional planning options, but [Future Improvement #1] users may want to explicitly limit options to e.g. “dinners under an hour”.

• (Fixed) Improvement #4: Add an option to not show a user certain restaurants. We added a “DO NOT EAT (SKIP THESE PLACES)” option to list restaurants users know they don’t want to eat at beforehand:

Clicking the x next to a restaurant removes it from results and adds it to the user session’s DO NOT EAT list, but [Future Improvement #2] does not persist because we don’t have a user login to store our knowledge of user preferences yet. (Yelp integration would also help with understanding a user’s dining preferences, but we understand that’s not a viable option right now.)

What’s next for Yelpboxd

While building out Yelpboxd v0.1 we identified additional future improvements including allowing users to set time goals for their own dining experiences (additional filters on price point and casual-to-fine-dining scale would also be useful) and persistent user preferences that would require direct Yelp and/or Letterboxd integration, or maintaining separate Yelpboxd user login.

More existential and central to the future development of the Yelpboxd platform is the behavior of the Yelpboxd recommendation model. Currently, we maintain a philosophical parallel to Letterboxd’s view of its platform as an “organic recommender”, which is to preserve and elevate the craft of restaurant-film pairing by using human-defined heuristics and user-specific feedback loops without storing or building a model based on user similarity. But Letterboxd has the benefit of being a social media platform: even if it doesn’t recommend films to users in the same way Netflix does, it surfaces the recommendations of your friends (and filmmakers, directors, cast members, public figures) on the platform as a feed of logged and rated films. Yelpboxd does not have this organic social discovery in place, and the next big leap for Yelpboxd will be building a balanced user graph with sufficient data and activity to establish its utility as a platform without falling into the “Toy Story 5 trap”: Yelpboxd will need to not just be consistent and functional, but also highly personal, exciting, and responsive enough to inspire even the most seasoned dinner-and-movie connoisseurs of our generation.

Detecting everything, everywhere

Hot on the heels of successfully vibe-coding (in an hour) an on-screen facial recognition software suite that detected Daniel Molloy in Interview with the Vampire and played SexyBack in real time, I set out to solve a more complex problem.

As we started Company Retreat, Jenny (the inspiration behind Molloy Detector) pointed out a number of items on screen that she was interested in purchasing as party favors, including these chili pepper necklaces:

She was forced to manually search for them on Amazon, yet another glaring platform problem begging to be automated.

aside: Prime Video’s X-Ray

Prime Video’s X-Ray feature already overlays, in real-time, a panel of actor pages and trivia provided by IMDb:

This is an impressive feat of engineering but not actually in the ways you might suspect. As we demonstrated with Molloy Detector, facial recognition is a solved problem and Amazon can afford a high quality standard via its ownership of IMDb and AWS’s AI suite. The harder problem is delivering this much temporal information consistently and dynamically to many concurrent viewers. Perhaps Netflix, recently measured at twice Prime Video’s viewership, hasn’t implemented a version of X-Ray for more reasons than not owning a comprehensive catalog like IMDb.

But if the hard part isn’t facial recognition, how hard is it to extend this to brand and product detection on screen?

No-context window shopping

The closest solutions I could find deployed at scale were Shop Your TV and WornOnTV, which seem to asynchronously pull and publish screenshots labeled with product details and, at least on WornOnTV, direct links to buy them.

However, neither Shop Your TV nor WornOnTV had updated their catalog to include Company Retreat, indicating a data lag and popularity bias. Shop Your TV and WornOnTV could be using completely end-to-end human workflows to identify and label clothing- and jewelry-containing screenshots from popular shows to put on their websites, but I hope for their collective sanities that they are at least using some kind of automation to routinely collect and filter through screenshots to annotate and serve to their customers. Based on specific annotations it looks like they may also be relying on social media sleuths to help label this data.

Overall, I found the on-screen clothing detection market to be smaller and more handmade than I initially assumed.

Another hour with Claude: prototyping ShopMyScreen

• Seed prompt: “build a clothing detection software for tv and film that allows me to click links to buy clothing i see on screen” / “name it shopmyscreen”

• v0.0: Claude one-shots a UI similar to Molloy Detector that will make calls to Claude Vision, which will cost money. This is reasonable because I never added “don’t use Claude Vision it costs money” to CLAUDE.md. Still, I have to think of the end user (“ok is there a free method”), to which Claude proposes:

  • BLIP (Bootstrapped Language-Image Pretraining) requiring a 1GB local download, which will caption images that can be used as input to retailers

  • FashionCLIP, built on OpenAI’s CLIP (Contrastive Language-Image Pretraining) “adapted for the fashion industry”, requiring a smaller 600MB local download. The FashionCLIP paper correctly calls out “the huge operational costs of training and developing models” leading to “the value of ML innovations… mostly captured by a few players” motivating their decision to open source their model.

• v0.1: Since we’re running this locally, BLIP and FashionCLIP both require some boot-up time which Claude doesn’t initially incorporate into UI. I guide Claude to add progress bars and additional instructions into the UI so users don’t confusion-quit out of the application before local models have a chance to load.

• v0.2: Open-source BLIP and FashionCLIP can only identify at the granularity of “blue shirt”. Functionally, BLIP and FashionCLIP are both downloadable locally and ready to run on screenshots and Claude picks a few retailers to link off to (Google Shopping, Amazon, ASOS, Shopstyle) that I wouldn’t have picked myself, but it’s a start. Claude did warn me (paraphrased) that I would get what I paid for; these open-source alternatives to Claude Vision were lacking in their ability to recognize specific clothing brands.

• v0.3: I ask Claude to add an option to image search. Claude proposes Google Lens on the cropped part of the screenshot containing the clothing item, but Google Lens does not have a public API. There is a workaround: send the image in POST request to Google Lens reverse image search endpoint. On macOS this still requires some user action to click into Google Lens results, but brand identification accuracy is much higher than BLIP or FashionCLIP:

• v0.4: ShopMyScreen can identify categories of clothing on screen and direct link to Google Lens. From there, the user can adjust the image to search on and find exact match with high success rate guaranteed by Google Lens.

Show-stopper or show-flopper? Prime Video’s Shop the Show

Amazon recently launched Shop the Show on Prime Video for customers to “discover toys, apparel, and other products related to what they’re watching”, only available on the mobile app, which is definitely capitalizing on watching-TV-and-scrolling-phone behavior.

For example, Company Retreat’s Shop the Show features a huge number of t-shirts and merchandise that I will probably be buying at some point, but that doesn’t solve the ShopMyScreen problem: I could not find the chili pepper necklaces anywhere on this Shop the Show page.

ShopMyScreen final boss: chili pepper necklaces

Using ShopMyScreen, I was able to identify the chili pepper necklaces via Google Lens. Admittedly there was not much use for FashionCLIP or BLIP for this necklace detection task because I had to crop the screenshot myself, but at least it’s more specific than Prime Video’s Shop the Show.

Tailoring ShopMyScreen is hard to do all at once

At this point, we have a functional prototype (the software can detect clothing on screen and I can click off to determine where exactly to buy it), but there are a few major scalability and usability concerns:

1. Thin UI layer / minimal improvement over manual Google Lens search. Without Claude Vision, our options for accurate brand detection are limited to Google Lens, which (without unauthorized calls via some proxy API) still requires users to click off, crop the screenshot to the clothing or item they want to image search, and find the retailer themselves. This isn’t too different from the completely manual flow of taking a screenshot and pasting it into Google Lens yourself. In fact, it’s basically a “Google Lens wrapper” with some categorical detection from FashionCLIP / BLIP.

2. FashionCLIP / BLIP without Google Lens is basically useless, unless you really just want to buy a blue shirt on Amazon.

3. Training open-source clothing brand detection models without logos is a hard problem. There are open-source solutions to many subproblems that could collectively feed into a reasonable ShopMyScreen solution (logo recognition DeepLogo, brand/logo Brand Eye), but clothing poses an often logo-less problem. These open-source models should have no issue identifying Coca-Colas and BMWs on screen, but unlabeled clothing is far less generalizable and will also intuitively have less training data, putting open-source clothing brand detection models at a disadvantage. Google Lens can do it, Claude Vision can probably do it, and this is where the FashionCLIP paper’s observation that ML innovation is captured by a few big players in the market (because of prohibitively high training and operational cost) becomes particularly salient.

4. Integrating ShopMyScreen into video streaming platforms would be exponentially more complex. This prototype relies on an initial download of up to 1GB local models (FashionCLIP / BLIP) and runs locally by taking screenshots and updating links to Google Lens and retailers. The advantage is that any video streaming platform that can be screenshotted will work with this local software. But if we tried to build a browser extension or feature for any video streaming platform to do this, we would need to compute model outputs and serve them to the viewer alongside the actual video content, the same way X-Ray does for Prime Video. In ShopMyScreen’s case, the number of detectable products is at least an order of magnitude higher than IMDb’s actor database, so to scale this we would probably have to significantly limit the types of products that would trigger a detection and link off to buy.

5. ShopMyScreen is currently tuned to clothing recognition (e.g. FashionCLIP), but could be more effective as a simple logo detector. As referenced earlier, logo detection is an easier problem than facial recognition, so if product logos (brands, cars) are visible on screen, there are sufficiently good open-source models that we wouldn’t need to jump through all the hoops of using Google Lens to get around Claude Vision.

Next up for ShopMyScreen

v0.4 is up on GitHub, but it’s not a viable product beyond single user at this point. I’d like to push the limits of real-time brand recognition and explore serving model results via video streaming overlay next. I could also explore models tuned to identify logo-less clothing brands, but the most pertinent engineering bottlenecks in developing both ShopMyScreen and Molloy Detector are in linking images to model results and building a cost-effective way to stream these to viewers in real-time. I’d like to get to a point beyond locally run software and downloaded models processing periodic screenshots before exploring accuracy improvements to the actual models.

The motivating user story

On Monday, May 4, 2026, Jenny sent me this use case (referring to Daniel Molloy played by Eric Bogosian in AMC TV series Interview with the Vampire):

That same night I watched Jenny go into an episode with SexyBack queued up on her phone ready to play, but she quickly became so absorbed in the dialogue and action that she forgot to press play at least 50% of the time Daniel came on screen. Even more concerning from an end user experience perspective: she had to manually stop and rewind the track each time to prepare for the next instance, losing precious time and attention she could have used to continue monitoring on-screen appearances.

I set out to solve this problem once and for all.

20 minutes with Claude Code

On Tuesday, May 5, 2026, ~5:00-5:20pm PT:

1. Seed prompt. I prompted Claude Code: “build a functioning UI that you can set to listen to audio or monitor a screen. when an actor (in this case, Daniel Molloy from Interview with a Vampire) comes on screen or starts speaking in detected audio, play a certain sound bite (in this case, sexyback by justin timberlake)”

2. Claude is a better UI designer than I am. Claude generated a prettier UI than anything I ever compiled and ran locally in my at least two years of full-time computer science coursework and at least five years of full-time professional software engineering. Stylistically, I did not modify this UI at all and pushed directly to production.

3. Like many humans, Claude can struggle to parse an unstructured instruction stream. I asked Claude to “listen to audio… when… in this case, Daniel Molloy… starts speaking in detected audio”, which Claude interpreted as listening to literal audio triggers “daniel”, “molloy”, “mr. molloy”, “daniel molloy” and also added “the journalist” for good measure (evidence of Claude’s wider training, since I did not mention Molloy’s role in the show). Since Molloy is neither Siri nor Alexa and has no default wake word in the show, this was as good a detection as any. Reasoning that audio detection is lighter weight than facial recognition and that power users could go into the source code to modify these triggers themselves, I did not further refine the audio detection module and pushed it to production.

4. Claude is, by default, not cost-conscious. Functionally, Claude needed some direction. Its first solution for screen monitoring facial recognition was to periodically take a screenshot (fine) and send screenshots to Claude Vision for face matching with the request “Is Daniel Molloy (Eric Bogosian) visible?”. With a default 3-second screenshot interval over average 60 minutes per episode, this meant 1K+ screenshots per episode. Even before estimating any additional token spend beyond just a baseline one image per request sent to Claude Vision, this was already approaching $5 per episode (Sonnet 4.6 rates). That simply would not stand, as known potential users of this software complained about recent Netflix price increases from $7.99 to $8.99 per month. I prompted Claude to add a cost estimate to the UI for Claude Vision and to find cheaper methods.

5. Claude makes a reasonable cost-functionality tradeoff. Claude proposed less frequent API calls to reduce cost, which was sensible but would sacrifice a quality standard (low-latency facial detection) I wanted to uphold for this product. Claude also proposed a local solution to completely eliminate the need to make API calls, but this required adding a UI component to have the user to upload a reference photo that could then be used as input to the face_recognition Python library.

6. v0.0: User testing and refinement necessary. Now I had a working Python program to open an UI, requiring an exactly-named assets/sexyback.mp3 in the source code folder and a reference photo to be uploaded in UI if the user wanted to use facial recognition. I ran the Python program locally from command line to open MOLLOY DETECTOR (impressively, Claude was able to one-shot match the ethos I had hoped to convey in UI design), provided necessary local screen monitoring permissions, uploaded a reference photo (a screenshot from the show), and started an episode of Interview with the Vampire. MOLLOY DETECTOR was unable to identify the back of Daniel Molloy’s head, which was reasonable. When his face came on screen, MOLLOY DETECTOR printed a delightful log message and started playing SexyBack. I was elated. However, after a few seconds, SexyBack started to drown out on-screen dialogue. After 10 seconds, I realized there was no functionality to stop or fade-out the track. Since I had uploaded the entire 3+ minutes of SexyBack to assets/sexyback.mp3, MOLLOY DETECTOR was dutifully playing the entire 3+ minutes of SexyBack over critical, dramatic dialogue.

7. v0.1: Add audio auto-stop and UI options. I gave Claude the precise, technical feedback “is sound cooldown working” to which Claude responded “Yes, it’s working correctly.” However, the rationale provided referenced code for detection cooldown, so I added additional context “the sound isn't stopping”, which Claude understood as instruction to add configurable auto-stop functionality (defaulting to 10 seconds) and a STOP button to the UI for the user to hard override. This was enough for me. I re-ran MOLLOY DETECTOR, SexyBack started playing when Molloy’s full face came on screen, and stopped after a reasonable time.

8. v0.2: Addressing lags in on-screen face recognition. Not being able to detect backs of heads was reasonable, especially since I only provided one reference photo of Daniel Molloy mid-speech. Given the 3-second detection interval I expected a small amount of lag. But there were instances where Molloy was on screen for more than a few seconds and I did not hear SexyBack. I also noticed that the program did not detect any faces on screen when there were multiple faces on screen. I told Claude “can't tell when any faces are on screen” and Claude added a color-coded real-time UI diagnostic “no faces on screen” / “faces found / no match” / “match - molloy detected” in addition to existing logs. Claude also suggested experimenting with tolerance used as an input to face_distance, which had defaulted to 0.55. Per Claude I set this to 0.60, which theoretically would not actually solve any of the detection issues I raised but upon next run worked reasonably well. I tabled this to re-visit later.

40 minutes with myself and my text editor

At the 20-minute development mark, I pushed the source code to GitHub and started to write documentation.

I do not use AI tools to write since I prefer to write from the heart (even, and usually, at my own expense), so it took me the rest of the hour to write the README. Notably, I reinforced while writing a quickstart user guide that MOLLOY DETECTOR is unusable for a non-technical user and is far from ready to be packaged into a consumer product as-is. It requires command line proficiency, uploading likely copyrighted material to specific filepaths, and granting screen recording permissions that would and should face hurdles in any app store worth its salt in security practices. I pushed copyrighted audio (SexyBack) to the GitHub repository because I put my customers first and I am not trying to profit from this, but that is a huge no-go in the rest of the industry.

I filed the most pressing functionality improvements as issues (UI accessibility, runnable executable, uploading audio to playback and not requiring users to upload to assets/sexyback.mp3 exact filematch, and testing Claude Vision and audio detection).

What this means for all of us

MOLLOY DETECTOR is not ready to be a Chrome extension. It is not ready to go on any app store. And while Claude Code, even out of box without significant MCP configuration, gets something reasonable up and running in a short amount of time, it can also lead an inexperienced developer astray.

For most software engineers, product managers, and industry practitioners in San Francisco and New York reading this blog, this isn’t news. You have little trouble guiding Claude, Codex, Cursor to accelerate your development; you naturally integrate them into your daily work in an informed, controlled way. You understood MCPs immediately as a conceptual API. That’s because you’re the experts. You’ve built software from scratch, you’ve made cost and design trade-offs with definitionally irrational human beings before there were standardized recommendations. You are able to appreciate the situations where Claude can one-shot an implementation and you recognize the limitations it has in building extensible, maintainable software without your explicit enforcement.

You are also the exception in the massive population starting to rely on tools like Claude Code to build products that humans use.

One reason why I continue to use Uber instead of Waymo is that I learn a lot from drivers. My driver on Monday gave me a fairly well-informed take on OpenAI based on his conversations with riders in San Francisco and the tech podcasts he listens to while driving. He was also the first adopter of AI tools in his day job for a catering company and taught his coworkers how to use Copilot. His biggest criticism of Copilot? “You can still tell when writing is AI-generated.”

Personally, I don’t plan to take MOLLOY DETECTOR too much further. As with many things, I feel I’ve gotten at least 80% of the learning from the first 20% of my time with this problem. We’ll stress test this software at an upcoming watch party, which may reveal additional urgent issues.

The larger unknowns and opportunities for longer-term economic and social progress are in studying the ways coding and AI assistants influence the ways we learn and work. To me, this is a permanent work in progress. We already know the risks of the data and methods we use to train our machine counterparts. We know how badly it can go wrong. We have a sense that we need to establish and enforce safety guardrails and checks. The balance I am constantly seeking and revising my own calculus around is the trade-off between what we learn from these tools we’ve trained to be more competent than any single human in many regards, and how much we’re willing to give up in developing our knowledge and reasoning from first principles in the name of productivity.

spoilers: Stranger Things, The Pitt

On Thanksgiving Day last year, after a lifetime of actively avoiding anything described as “supernatural horror” and nine years out of the Stranger Things loop, I finally succumbed to overwhelming social pressure and gave the hit series a chance. I promised a friend I’d join them for the New Year’s special theatrical release of the Season 5 finale, so I had about a month to catch up on 40 hours of material. This was doable, I figured: even if I didn’t personally enjoy the show, it was a good opportunity to study a cultural phenomenon that captured a billion-dollar global market and launched the careers of multiple big-name stars in Hollywood and music.

I watched four episodes in that first night, stopping only because my brain was shutting down, begging me for more than a few hours of sleep. Not even halfway through December, I had barreled through every released episode and was starting a re-watch to study the cinematography and look for additional clues to the mystery that resolved in a $25 million finale at the box office. On the surface, I had little in common with the main characters — children, mothers, a troubled sheriff in ‘80s small town Indiana — but the core themes of fear and hiding, loss and grief, searching for meaning and finding the strength to persist were so universal and compelling that I could connect on the basis of the emotional weight alone.

Cinematic and thematic tribute to Steven Spielberg (E.T., 1982), Rob Reiner (Stand By Me, 1986), and Stephen King (Cujo, 1981; It, 1986) was immediately apparent from the beginning. So was product placement. There were recurring close-up shots of cans of Coca-Cola, boxes of frozen Eggo waffles, and enough scenes involving Jif peanut butter to constitute full-length advertisements all on their own, but the framing was beautiful and relevant enough to building a rich world rooted in reality that I didn’t mind, even appreciated reminders of the brands I, too, consumed religiously as a child.

It also got very dark, very quickly.

Many read the Upside Down (a parallel dimension overrun by violent monsters and controlled by a demonic force) as an allegory for depression, a desolate version of reality that can isolate and overtake you, something your loved ones will crawl and claw through to save you from, if you can hold on long enough. Some have read it as a symbol of closeted homosexuality and urge viewers to understand the reality of what being openly gay meant in the ‘80s, thirty years before same-sex marriage was legalized in the US, and during the AIDS crisis which mainstream media largely avoided, heightening misinformation and prejudice. There are on-screen depictions of post-traumatic stress and panic attacks both in the Upside Down and in the real-world dimension that hit me so hard and unexpectedly I found myself physically unable to stop crying. For a series that has been compared to Harry Potter both in content and cultural impact (“a show that helped shape a generation”), its willingness to paint such a wide range of dark reality, at times raw and intimate, at times angry and violent, and so often painfully, heart-wrenchingly beautiful, struck me as particularly bold.

Stranger Things cast such an unprecedented spell on the world because it had cross-generational, multi-genre appeal. Teenagers saw themselves reflected on screen, perhaps pulled into a world of teenage heroes and alternate-dimension monsters the same way I once saw myself in Hermione Granger and held onto hope of receiving a late acceptance to Hogwarts (bureaucracy and paperwork, am I right?); the true-to-the-era aesthetics and costuming reminded other viewers of their youth; film and music buffs were treated to an abundance of references that made for a visually and sonically rich walk down memory lane; sci-fi and horror fans got the world-building, tension, and adrenaline they were looking for.

Like Harry Potter and Twilight, Stranger Things became a franchise and chose to invest in characters and themes that were popular with viewers. As with any media franchise, viewers had strong opinions, and there are many things I personally would have liked to see done differently on screen. I don’t say this to detract in any way from the strong storytelling, production, and talent behind these series. I grew up on Harry Potter and made some of my closest friends over shared love of a magical, otherworldly universe. Twilight launched the careers of some of my favorite actors and directors. The Stranger Things cast and team deserve recognition for raising a new generation of film and television aficionados and choosing to tell a controversial story in a complicated sociopolitical environment.

But to let cultural media fade away into nostalgia without studying what made them resonate with so many, without analyzing the ways they may have lost and misrepresented large subsets of viewers, and without pushing for more, is to give up on a battle we have barely started fighting.

One core tension defines most content in the modern entertainment industry: whether to write for mainstream appeal or to optimize for a cult following. David Harbour may have claimed the Stranger Things creators thought they “were going to make a beautiful show that like ten geeks [were] going to watch”, but nothing starring Winona Ryder could have realistically been pitched as an underground show. It was always meant to be big. And, just like any product, the bigger something gets, the harder it is to preserve the quality of being at the fringe, of the vulnerability and specific color of ego behind crafting a labor of love. When I think about what this tangibly means for creators and consumers, it comes back to the same growing concern I see in music, literature, and art: taste degradation accelerated by exponentially cheaper, faster, easier-to-find and access art at the world’s fingertips. The proliferation of streaming services and media exclusively funded by and released on an increasingly consolidated set of platforms forecasts a cloudier-than-ever future for less provably monetizable, “riskier” perspectives.

It’s shortsighted and self-limiting at this stage of global socioeconomic development to insist on the removal of artificial intelligence from the creative process. I commend Sundance Now’s commitment to “algorithm-free” streaming, presumably guaranteeing human curation of independent titles, but it’s a marketing tactic aimed at film connoisseurs and will not scale to match Netflix or Disney viewership. Any science fiction fan can attest to the thousands of different ways we’ve imagined artificial intelligence going wrong, becoming an enemy to society and human existence as we know it. Any horror fan knows that willful ignorance is often the first and most fatal flaw. Opting out of learning how to use these tools to our advantage is, in similar vein to opting out of consuming mainstream content because it fails to tell our stories accurately and respectfully, akin to splitting off from the group early into the horror movie to stumble into certain death.

I am not dismissing what is a very real, paralyzing fear of the unknown in the AI-assisted future of entertainment. We see it in every industry. The Pitt summarized it well in a recent episode. A new chief attending physician brings in a fresh suite of protocols including an AI medical transcription tool, which incorrectly records a diagnosis. A resident rolls her eyes: “AI”. The chief attending reminds the team that this AI tool is “98% accurate, but needs to be carefully checked”. Experts disagree on how and when to use AI in every field, which is why we need to keep using it as practitioners in every field, a sort of “guerrilla AI safety activism”. Entertainment may not be fundamentally urgent to human life in the same way medicine is, but it’s life-affirming and culture-defining in far less quantifiable ways.

A common information safety principle is to focus on open-sourcing and making information available and accessible to as many people as possible. The idea is that democratization, putting information into the hands of well-intentioned members of society, leads to best possible outcomes. Of course, there’s undeniable risk of bad actors in the system taking advantage of these distribution channels, propagating malicious information, harmful software, but that’s part of what conscientious activism should counteract, in a net-good world at least. When Anna’s Archive scraped the entirety of Spotify’s music library last year and declared intention to distribute it via torrents, I read the manifesto. I was struck by the thoughtful engineering, the technical prowess required to pull off such a feat, and I was surprised by the depth of the data analysis, too. It read like a music data nerd’s blog, analyses of listening habits, artist patterns across genres. Yes, by distributing Spotify’s catalog for free, Anna’s Archive was funneling away a source of income to creators. But the realist in me reasoned that this was already going to happen — we’ve been pirating movies and music long before Spotify — so why not centralize it in a secure, protected place, safer for the consumer?

Spotify sued Anna’s Archive and won, but the fight continues; if anything, this adds fuel to the fire.

As for representation in media, the Oscar nominations released today notably left out Eva Victor’s outstanding directorial debut Sorry, Baby (2025) and Korean auteur Park Chan-wook’s No Other Choice (2025). We have a long, uphill road ahead, but we also need to recognize that we live in an exceedingly violent, terrifying, and unpredictable world; it is a privilege we should not take for granted to be able to keep fighting to see our stories played out on screen, our voices heard around the world.

I love movies for many of the same reasons I love music. A child might watch The Matrix with her father and completely not understand it, but she might note that it was an important film to him, return to it years later to discover something about herself and rediscover something about him at the same time. An adult cleaning out her mother’s attic might find a dusty record collection and have a chance to connect with her in ways she could not have imagined previously. Film and music, much like literature and visual art, are ways for us to understand ourselves, connect with others, communicate with our loved ones. They are ways for us to express complex emotions and share experiences that cannot be captured any other way. They are ways for us to be human together, to chase the infinite possibilities of the universe together.

Giving up on making and consuming these art forms critically, humbly, and honestly, refusing to learn how to play the cards we are dealt, would only be a disservice to ourselves.

“My perfect day? Sunday, my favorite day of the week. Raining, overcast. I go to my local coffee shop and order a black coffee and a pastry. I sit down with some research papers (at least one from Anthropic) and go to town.”

Paper review

I’m starting this post by reading and annotating publications by authors I respect as writers. If you’re looking for original writing, please skip ahead to Tech’s bitter lesson.

“The Bitter Lesson” (2019) by Rich Sutton (Rich Sutton & Andrew Barto received the 2024 Turing Award and wrote my favorite textbook):

Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation. These two need not run counter to each other, but in practice they tend to.

Computer v. Kasparov, 1997: When a simpler, search-based approach with special hardware and software proved vastly more effective, these human-knowledge-based chess researchers were not good losers. They said that “brute force” search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. These researchers wanted methods based on human input to win and were disappointed when they did not.

AlphaGo, 2015: Search and learning are the two most important classes of techniques for utilizing massive amounts of computation in AI research. In computer Go, as in computer chess, researchers’ initial effort was directed towards utilizing human understanding (so that less search was needed) and only much later was much greater success had by embracing search and learning.

DARPA speech recognition competition, 1970s: the statistical methods won out over the human-knowledge-based methods. This led to a major change in all of natural language processing, gradually over decades, where statistics and computation came to dominate the field. The recent rise of deep learning in speech recognition is the most recent step in this consistent direction. Deep learning methods rely even less on human knowledge, and use even more computation, together with learning on huge training sets, to produce dramatically better speech recognition systems. As in the games, researchers always tried to make systems that worked the way the researchers thought their own minds worked---they tried to put that knowledge in their systems---but it proved ultimately counterproductive, and a colossal waste of researcher’s time

As a field, we still have not thoroughly learned it, as we are continuing to make the same kind of mistakes. To see this, and to effectively resist it, we have to understand the appeal of these mistakes. We have to learn the bitter lesson that building in how we think we think does not work in the long run… The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.

One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.

The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries… We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.

“Burning out: An essay on overworking in AI” by Nathan Lambert, published yesterday:

Where OpenAI’s culture is often described as culty, there are often many signs that the core team members there absolutely love it, even if they’re working 996, 997, or 002. When you love it, it doesn’t feel like work. This is the same as why training 20 hours a week [as an athlete] while a full time student can feel easy.

Crafting the team culture in both of these environments is incredibly difficult. It’s the quality of the team culture that determines the outcome more than the individual components. Yes, with LLMs you can take brief shortcuts by hiring talent with years of experience from another frontier lab, but that doesn’t change the long-term dynamic. Yes, you obviously need as much compute as you can get. At the same time, culture is incredibly fickle. It’s easier to lose than it is to build.

I used to think that the AI bubble would pop financially, as seen through a series of economic mergers, acquisitions, and similar deals. I’m shifting to see more limitations on the human capital than the financial capital thrown at today’s AI companies.

Culture is a fine line between performance and toxicity, and it’s often hard to know which you are until you get to a major deliverable to check in versus competitors.

For me, the only reason to keep going is to try and make AI a wonderful technology for the world. Some feel the same. Others are going because they’re locked in on a path to generational wealth. Plenty don’t have either of these alignments, and the wall of effort comes sooner.

AWS postmortem on the outage that took down the Internet:

DynamoDB DNS automation miss: Automation is crucial to ensuring that these DNS records are updated frequently to add additional capacity as it becomes available, to correctly handle hardware failures, and to efficiently distribute traffic to optimize customers’ experience… In addition to providing a public regional endpoint, this automation maintains additional DNS endpoints for several dynamic DynamoDB variants including a FIPS compliant endpoint, an IPv6 endpoint, and account-specific endpoints. The root cause of this issue was a latent race condition in the DynamoDB DNS management system that resulted in an incorrect empty DNS record for the service’s regional endpoint (dynamodb.us-east-1.amazonaws.com) that the automation failed to repair.

EC2 delays post DynamoDB recovery: Customers experienced increased EC2 API error rates, latencies, and instance launch failures… After resolving the DynamoDB DNS issue at 2:25 AM PDT, customers continued to see increased errors for launches of new [EC2] instances. Recovery started at 12:01 PM PDT with full EC2 recovery occurring at 1:50 PM PDT.

NLB fallout from EC2: The delays in network state propagations for newly launched EC2 instances also caused impact to the Network Load Balancer (NLB) service and AWS services that use NLB. Between 5:30 AM and 2:09 PM PDT on October 20 some customers experienced increased connection errors on their NLBs in the N. Virginia (us-east-1) Region.

Other AWS services that rely on DynamoDB, new EC2 instance launches, Lambda invocations, and Fargate task launches such as Managed Workflows for Apache Airflow, Outposts lifecycle operations, and AWS Support Center were also impacted in the N. Virginia (us-east-1) Region.

“How Complex Systems Fail” by the late physician and software engineer Dr. Richard Cook, published 1998-2000:

All of the interesting systems (e.g. transportation, healthcare, power generation) are inherently and unavoidably hazardous… The frequency of hazard exposure can sometimes be changed but the processes involved in the system are themselves intrinsically and irreducibly hazardous.

Post-accident attribution to a ‘root cause’ is fundamentally wrong: There are multiple contributors to accidents. Each of these is necessarily insufficient in itself to create an accident. Only jointly are these causes sufficient to create an accident.

Hindsight bias remains the primary obstacle to accident investigation, especially when expert human performance is involved: The outcome knowledge poisons the ability of after-accident observers to recreate the view of practitioners before the accident of those same factors. It seems that practitioners “should have known” that the factors would “inevitably” lead to an accident.

In fact that likelihood of an identical accident is already extraordinarily low because the pattern of latent failures changes constantly. Instead of increasing safety, post-accident remedies usually increase the coupling and complexity of the system. This increases the potential number of latent failures and also makes the detection and blocking of accident trajectories more difficult.

Failure free operations require experience with failure: More robust system performance is likely to arise in systems where operators can discern the “edge of the envelope”. This is where system performance begins to deteriorate, becomes difficult to predict, or cannot be readily recovered. In intrinsically hazardous systems, operators are expected to encounter and appreciate hazards in ways that lead to overall performance that is desirable. Improved safety depends on providing operators with calibrated views of the hazards.

“Preparing for AI’s economic impact: exploring policy responses”, Anthropic’s follow up to their September Economic Index report I covered in a previous post:

Policies for nearly all scenarios:

1) Invest in upskilling through workforce training grants (e.g. subsidies to employers to create formal trainee positions with structured training programs; redirect higher ed subsidies to fund this; tax AI consumption)

2) Reform tax incentives for worker retention and retraining

3) Close corporate tax loopholes

4) Accelerate permits and approvals for AI infrastructure.

The suggestion to redirect higher education subsidies to fund workforce training programs is questionable given the current state of higher education funding. AI consumption tax is tractable, but face heavy opposition by certain states buoyed by AI investment. 2) and 3) don’t necessarily benefit Anthropic, its investors or corporate partners, giving these proposals some credibility, but also equally hits wider corporate AI competition (Google, Meta, X). 4) is clearly in Anthropic’s favor and would bolster AI development and funding generally.

Policy ideas for moderate scenarios:

5) Establish trade adjustment assistance for AI displacement: how the Trade Adjustment Assistance (TAA) model – in which affected workers are given opportunities to obtain new skills, or receive other support – might be adapted for labor market disruptions in an era of powerful AI. (“AI insurance” as a mechanism “to support those who lose jobs due to AI.”)

6) Implement taxes on compute or token generation… We believe taxes in this broader category deserve serious study, even though they would directly impact Anthropic’s revenue and profitability.

Policy ideas for fast-moving scenarios:

7) Create national sovereign wealth funds with stakes in AI… [for the UK] The AI Bond would aim to ensure adequate investment in “the AI stack” to capture AI’s benefits and then distribute its returns more evenly across Britain—even as AI research roles concentrate in a few cities, like London.

8) Adopt or modernize value-added taxes (VAT)… US is [currently] the exception… As AI transforms the economy, labor’s share of the production of value might decline significantly. A shift toward taxing consumption (as through a VAT) could become necessary to fund core government activities.

Briefly, sales tax in the US is collected by the retailer at final point of sale vs. VAT tax which is collected at each stage of supply chain.

9) Implement new revenue structures to account for AI’s growing share of the economy… [if labor’s share of economic output declines significantly] governments might require new revenue streams to complement income tax… [could explore] a “low-rate business wealth tax” as a complement to income taxes.

Many of these policy ideas are around government regulation, taxation, and wealth distribution methods (e.g. via subsidies and public investment instruments - the “AI bond”) which Anthropic explicitly recognizes would be downward pressure on its profits. Anthropic’s “$10 million commitment to scale up the Economic Futures Program” seems low compared to Google’s $75M AI Opportunity Fund and OpenAI’s $50M nonprofit fund, although Anthropic’s commitment is specific to funding economics research and symposia.

Tech’s bitter lesson

Recall Rich Sutton’s bitter lesson pulled from 70 years of AI research: in the strategy games of chess and Go, in complex modeling tasks such as speech recognition, method advancements made by human researchers were not as effective as scaling up compute power. He proposed a goal of building machines that learn and explore like us, not from us, that could be organized and leveraged to make greater breakthroughs, instead of being limited to the collective capacity of our own minds.

Abhishaike Mahajan came to a form of bitter lesson conclusion on developing cancer treatments, in “Cancer has a surprising amount of detail”:

In most areas of medicine, a diagnosis tends to unify patients under a single therapeutic approach: antibiotics for bacterial pneumonia, insulin for type 1 diabetes, thyroid hormone for hypothyroidism. The drug may differ in dose or formulation, but not in principle.

This is not the case for cancer.

I’d like to focus on one fault that all the papers mention: the inability for many novel biomarkers to improve on the current clinical standard.

In other words, the obvious next step is to stop asking for singular biomarkers to bear the entire burden of explanation, and instead ask how many small signals can be woven into a coherent, usable picture. But this creates a combinatorial explosion! If you have 20 binary biomarkers, that’s over a million possible patient subgroups. No trial, no matter how well-funded, can enumerate that space.

How can we escape this problem? It is increasingly my opinion that the only reasonable path forward is to delegate the problem of cancer biomarkers to machine intelligence.

Dwarkesh Patel interviewed Rich Sutton last month:

Sutton: I don’t think learning is really about training. I think learning is about learning, it’s about an active process. The child tries things and sees what happens.

Patel: Once we have AGI, we’ll have researchers which scale linearly with compute. We’ll have this avalanche of millions of AI researchers. Their stock will be growing as fast as compute… As a vision of what happens after AGI in terms of how AI research will evolve, I wonder if that’s still compatible with a bitter lesson.

Sutton: How did we get to this AGI? You want to presume that it’s been done… The bitter lesson, who cares about that? That’s an empirical observation about a particular period in history. 70 years in history, it doesn’t necessarily have to apply to the next 70 years… I do think succession to digital intelligence or augmented humans is inevitable… We should be concerned about our future, the future. We should try to make it good. We should also though recognize the limit, our limits.

Meta’s AI layoffs this week were the latest development in years of restructuring in the post-COVID, post-agentic tech industry. Most tech layoffs in 2025 cite flattening management (Google eliminated 35% of its managers leading small teams) and roles eliminated with increased AI adoption (Salesforce laid off 4,000 in customer support this year). It’s true that AI can do many things humans can do. It’s true that AI assistants, coders, and writers are more accurate and efficient than humans in many cases. It’s true that many managers are a net drag on their teams, introducing overhead via process, bureaucracy, morale.

But these are not new developments in the industry. Ineffective and inefficient management is a tale as old as commerce. Automation is definitely not new in customer support. The earliest “agentic AI” sales solutions were rule-based call routing, and honestly they were better than a lot of the solutions today.

Companies are incentivized to hire when business is good and talent is cheap. Often they are pressured by stock price and stakeholder commitment to prioritize short-term gains over proper long-term management. Big tech hired aggressively in the years following the 2020 COVID-19 pandemic, benefiting from temporary pandemic-era customer behavior shifts:

In 2023 many companies announced major layoffs, although CNN called out at time of publication that Apple grew its staff by only 20% from 2019 to 2022 and had not announced layoffs:

It’s tempting to conceptualize a “bitter lesson” for tech. Promo-driven development rewards engineers and managers hungry for greater responsibility, resulting in frivolous, buggy products and headaches and messes for others to weather. We’re seeing the risk of a messy software network start to become more publicly obvious in cases like AWS’s wide network of customers taking days to recover from a point of failure outage. Tying development outcomes to market behavior (and executive paychecks, funding and partnerships) encourages decision-making that ranks corporations above their constituents.

We have seen cases where widespread, normalized and instrumented financial and corporate risk capitalizing on short-term market-behavior led to societal financial collapse: in 2008, it was driven by inflated subprime loans that lost their value too quickly to recover from. We have been in a convoluted economic state for the last few years as sectors responded differently to the pandemic, recovery, and international trade developments.

The difference between the tech industry and the research picture that Rich Sutton drew his bitter lesson from is that the tech problem doesn’t have an objective function. We are not uniformly trying to build a chessmaster or the most accurate speech recognition system. If anything, corporate tech’s reward function is, like that of any corporate sector, profit — which is not actually quantifiable, when you consider things like optimizing for long-term profit being a valid corporate strategy that relies on trust and human capital to implement successfully while avoiding the trap of short-term capital needs. Tech’s bitter lesson is actually that, empirically in our 30 years of widespread Internet-based technology since dotcom, good social and research outcomes are byproducts of corporate development and an international markets game. And what of players like Anthropic, now actively proposing regulation and tax overhauls that may or may not benefit them? No different than oil and finance’s long-term lobby, just much more public. We may appreciate Anthropic’s human-first, socially-conscious marketing, but we don’t need to see it as anything more than a creatively-marketed software company.

So if the global economy and technological development is controlled by players and incentives we don’t understand, do we have any clear path to building a technological future we are empowered to influence to align to our values? Absolutely. While tech companies have capitalized on human attention and unsustainable workforce management to drive profit, great research advances and applications are enabled by competition in developing faster, cheaper, better versions of the underlying technologies. As Dr. Sutton highlighted to Dwarkesh, what defined the bitter lesson of the last 70 years in research need not continue into the next 70. The corporate tech patterns of the last 30 years need not define our next 30; in our current economic and resource systems, we can’t sidestep certain capital requirements to run successful organizations, but we can reward, both as consumers and producers, players with the power and potential to build a balanced world of humans and machines in harmonious symbiosis. We can shift tech’s bitter lesson by shifting its overall reward function.

Find me on X @valstechblog.

The title of this post is a reference to Machines of Loving Grace by Anthropic CEO Dario Amodei this month last year, itself a reference to Richard Brautigan’s 1967 poem. In it, he described “what a good world with powerful AI could look like”, highlighting the potential for transformative advances in physical and mental health technology, economic growth, inequality mitigation, climate and food supply, and governance systems.

I don’t know if this world is realistic, and even if it is, it will not be achieved without a huge amount of effort and struggle by many brave and dedicated people. Everyone (including AI companies!) will need to do their part both to prevent risks and to fully realize the benefits.

…It is a world worth fighting for. If all of this really does happen over 5 to 10 years—the defeat of most diseases, the growth in biological and cognitive freedom, the lifting of billions of people out of poverty to share in the new technologies, a renaissance of liberal democracy and human rights—I suspect everyone watching it will be surprised by the effect it has on them.

Basic human intuitions of fairness, cooperation, curiosity, and autonomy are hard to argue with, and are cumulative in a way that our more destructive impulses often aren’t… We have the opportunity to play some small role in making it real.

I’ll admit I attended the Death to Spotify Forum with an explicit intention to get out of the San Francisco tech bubble and to challenge myself to rethink and rework the ways I consume and interact with music and media. I wanted to emerge from Oakland five weeks later a boombox-sporting, vinyl-spinning, radio-supporting connoisseur of underground music. At minimum, I hoped to come away with a stronger sense of identity and morals as a musician and consumer.

Perhaps unsurprisingly in retrospect, I did the opposite. I’d hear our guest speakers talk about the brief magic of chancing upon music that made them feel a certain way, and I’d be reminded of my own surprise and delight at the ~0.1% of my YouTube recommendations that stuck with me: what made YouTube a sparse but high-quality recommendation system? Record label owners and radio station hosts discussed the complexities of reaching and engaging with their audience in a streaming-first economy, and I’d start thinking about ways to tip economic incentives at scale, how to make streaming alternatives more compelling for customers. We talked about the existential struggle indie bands face these days if they’re not on social media, and I thought to myself, “We can abstract that away for them.”

I thought about scalable systems design, which might sound far-fetched at first, but actually became a single unifying principle for me over the weeks. I realized that my friends and I might be able to intentionally rewire our brains for longer-term gratification, delete our social media, and move off streaming services. There is value to living a more unplugged life. But being totally offline isn’t accessible or realistic for much of the world. We are privileged to have so much stimulus in our lives that we can decide which ones to turn off. I was torn between voting with my actions by boycotting attention-based technology — turning back to offline, old-school methods — and becoming a majority consumer myself so I could identify the core incentives motivating our media consumption patterns. I wanted to piece apart what was systemically broken and work to fix it. How do we build a better future for ourselves without first understanding how we got here? How do we design for disparately-distributed populations of which we are miniscule segments?

To stream or not to stream: the artist’s conundrum

In the first week of reading group, we discussed Liz Pelly’s pieces on Spotify (Big Mood Machine and The Ghosts in the Machine) describing advertisement packages based on users’ moods and lifestyles inferred from music listening behavior and an internal “Perfect Fit Content” program that pushed playlist curation and the music itself in lower-cost directions for the company. We read about a wave of artists including King Gizzard & The Lizard Wizard boycotting Spotify over demonetization of already low payouts to grassroots artists, a growing and unsolved AI-generated music problem, and Daniel Ek’s investments. We talked about how we discovered some of our favorite artists on Spotify in the past, how we now found ourselves passively letting the algorithm decide what we should listen to, how we were increasingly dissatisfied with our recommendations but also didn’t necessarily know what else would be better.

Knowing that I was writing this post, a friend sent me their reasons to cancel Spotify to use as a visual aid:

Spotify, a blessing and a curse. Spotify, something that has started to accompany every part of life, something you might struggle to stop using because you use it when you’re cooking, commuting, working out, to boost your mood, shut off your brain, find and share music with your loved ones, even as a “healthier” alternative to other coping mechanisms. Spotify, the one platform that has become synonymous with a musician or band “getting serious” and “going pro”, which they might have to use to reach and stay relevant to their fans, but can’t count on to sustain them financially. Over the weeks the issue only became more complex.

As Bobby Martinez (Dandy Boy Records) and Ebben Saunders (Psychic Radio) expressed, the value platforms like these bring to humanity aren’t easy to quantify. The world’s music at your fingertips, what a beautiful concept, Ebben said. How intimidating to have this access be controlled and limited by a concentrated corporate landscape in which economic incentives don’t necessarily align with social ones. We heard from Haneen Sidahmed (Sudan Tapes Archive), tresX X Arriaga Cuellar, and DJ JUANNY Jorge Courtade (Rebane Récords): it’s really a question of access. There are many worlds of music not online; finding them is an effort in and of itself. Many cannot afford to keep physical music collections, let alone double check that everything they consume is properly and fairly attributed. Some artists make music to survive. Others put the money to questionable use. Where do we draw boundaries as consumers and creators?

I began to realize in these conversations that having the agency to decide what forms of media and which artists we support is itself a luxury, a continuous learning process of our own values and belief, a true labor of love fraught with immense political and economic meaning.

Curation as a medium of connection and community

Our second week focused on radio. Gabriel Lopez (KEXP) spoke about an early memory of hearing a classic song on the radio in the car his parents had just bought, driving back home. Mohit Kohli (Fault Radio) talked about an adolescent ritual of discovering and putting on music, reading the liner notes, feeling a connection with the lyrics and artist, and later, finding confidence as a DJ, independence in working for himself. There was a process to discovering and identifying with music which took time, trial and error, and made finding your favorites more gratifying and memorable.

I was struck by something Neroli Devany (Humboldt Hot Air) said (paraphrased, pieced together from my barely legible notes):

“Society likes comfort. We always have to be occupied. We want to be able to skip the track if we don’t like it. Sometimes I’m driving and I find myself moving to skip before I realize it’s the radio… But by accepting we’re not in control, that this is important to someone, and there is a human behind that… The act of going and staying out of our comfort zones is what it takes to be part of something bigger. That’s what it means to be a part of a community.”

Neroli mentioned hearing radio DJs making mistakes live on air, playing tracks over each other, and appreciating that only a human could be behind such things. I feel a similar appreciation for musicians like Mac DeMarco, who spent years chain-smoking cigarettes and making low-budget music from a lofted bedroom studio in Bushwick (a fan commented in 2016, “he wouldn’t be Mac without his cigs”) before quitting smoking and drinking and releasing a nine-hour, 199-track, five-year, largely-instrumental album in 2023 I call his “sobriety album”:

“I look at the world and the music industry and maybe I just want to be more of an artist, or pretend to be. Because I mean, there are plenty out there, but I think there’s another side of the music industry where it is like sport. But I like seeing all the crap. The human part is the most important part to me in any kind of art. I guess I’m just trying to remain human.” — Mac DeMarco for NME

Tshego Letsoalo (KALW) told us about the time she stood still in her kitchen, not expecting to resonate so strongly with something that had suddenly come on the radio. It was bittersweet to register a fleeting moment of delight even as it passed. That, we agreed, was central to being human.

Music as a cultural discovery and preservation tool

We heard from sound archivists. Haneen talked about starting Sudan Tapes Archive during the pandemic — “I wanted to listen to childhood tapes. I’m very nostalgic” — a quest to preserve the original music she remembered being given by family in Sudan, but could not find anywhere. But, as many raised in multiple cultures can understand, asking questions and attempting to document familial and cultural history can be met with confusion, resistance, anger, and pain. Jorge told us about the challenge of seeking connection to Honduran roots when family members did not like to talk about the past, finding understanding and community in music that would “never have been recommended to me on Spotify, and if it wasn’t recommended to me, there’s no way it would have ever been recommended to you.”

At the same time, preservers of local and regional music have mixed emotions about widespread discovery, fearing conglomerates coming in to modify and market it to sell to a global audience without proper credit or payout to the original artists, popularizing a narrow perception of the complex culture and history behind the music. At best, this globalization can result in new sound, representation, and the fusion of communities; at worst, the virality is a fleeting, insulting bastardization of sound created by centuries of unity and strength in the face of hardship.

The soundtrack of my childhood was Whitney Houston, Michael Jackson, AC/DC, and The Eagles on my dad’s radio and unstructured, rambling piano lessons from my family’s next-door neighbor. She was a retired concert pianist who took it upon herself to design a custom music theory and piano performance training program for me, including observing my mistakes and writing out corrective exercises for me on blank sheet music. I have a distinct memory of watching her hands shake as she drew out a series of sixteenth notes for me to practice consistent technique until I perfected them, telling me to do my hand-stretching exercises every day to avoid developing the arthritis so many pianists suffer from later in life. My first out-of-body experience happened accidentally during a recital when I slipped out of consciousness and let my practice take over, felt my heart and time slow (see: this scene in Soul) as if I had become a medium between the composer and the audience, a vessel via which they could communicate across hundreds of years.

I’ve spent much of my life chasing this feeling, seeking to understand it, reproduce it, enhance it; finding new and unexpected iterations of it is one of my life’s most indescribable joys. Some might call it being “in the zone”, “locked in”, “flow state”; to me it’s more than that. It’s a sense of peace, order, and purpose that comes from finding and following missions I’m deeply, spiritually moved to keep pursuing.

Taste is learned but not taught

Called to action by this discussion and lecture series, I reflected on my own consumption habits. I’d spent more than ten years listening to music and watching television every chance I got; at work, at the gym, at home, on planes. My problem wasn’t with the media itself. It was that the same stream of mediocre media I was putting on to have something to distract me started becoming less effective over time. Because I wanted to delay the feeling of being bored as much as I could, and the same methods weren’t working as well, I started looking for ways to enhance my experience. I didn’t like being bored. Actually, I was scared of being bored. I felt it was important to be always on, always productive, and if I wasn’t doing something, I didn’t know what to do with myself.

I started playing albums from start to finish, no skips, minimal interruptions. I started paying to see movies in theaters, establishing an hours-long process of getting to the theater, watching the film critically, sitting through boring and uncomfortable scenes (and often being rewarded by something beautiful after), studying the details. I developed a taste for classic horror and foreign black-and-white film I truly never thought I’d have the stomach or intellectualism to appreciate. I discovered 3ee, re-discovered Teyana Taylor, and found parts of myself in films I would never even have watched without taking a leap of faith on the Roxie’s staff picks.

Taste, to me, is a constant learning process. Some things are immediate hits but diminish in gratification as you return to them; others unfold and multiply in complexity as you approach them in different contexts. Taste is subjective in the sense that you are the one most consistently present variable, also the hardest to quantify and replicate. This hard-to-model individuality is precisely what separates human from machine. Perhaps the one true humanistic thing we can do in a world of complex interactions and incentives is to continue revising and refining our individual tastes, to double down on independent, critical thought.

When I was a fairly small child, I began to eat the whole apple. Not just the flesh, but the core with all the pips in it, even the stem. Not because it tasted good, I don’t think, nor because of any idea I might have had that I shouldn’t be wasteful, but because eating the core and the stem presented an obstacle to pleasure. It was work of a kind, even if in reverse order: first the reward, then the effort. It is still unthinkable for me to throw away an apple core, and when I see my children doing it — sometimes they even throw away half-eaten apples — I am filled with indignation, but I don’t say anything, because I want them to relish life and to have a sense of its abundance. I want them to feel that living is easy. And this is why I’ve changed my attitude towards apples, not through an act of will, but as a result of having seen and understood more, I think, and now I know that it is never really about the world in itself, merely about our way of relating to it. Against secrecy stands openness, against work stands freedom.

— Karl Ove Knausgård, “Apples”, Autumn

In three words, yesterday was wild.

In three launch announcements:

• Anthropic dropped Claude Sonnet 4.5, featuring a 20% performance improvement in computer use and “telecom agent tool use” (see tau2-bench tasks: things like debugging cell service issues) over Opus 4.1. They also highlighted model improvements in financial analysis, law, medicine, and engineering, with feedback from human experts. More notable to me personally was the explicit announcement of “our most aligned frontier model yet” (“reducing concerning behaviors like sycophancy, deception, power-seeking, and the tendency to encourage delusional thinking”), the Claude Agent SDK to build agents, and a research preview of a hybrid infrastructure-agent UI experiment called Imagine with Claude, “software that generates itself in response to what you need”.

• Microsoft also released Agent Mode (using OpenAI GPT-5, although Office Agent in Copilot uses Anthropic models, and a Microsoft Office SVP is quoted saying Microsoft is exploring other model providers outside of OpenAI), “bringing vibe working” to Excel, Word, and soon PowerPoint, achieving 57.2% accuracy in the SpreadsheetBench Excel benchmark (human score: 71.3%):

  “It’s work, quite frankly, that a first-year consultant would do, delivered in minutes.” — Sumit Chauhan

• OpenAI announced Instant Checkout in ChatGPT allowing ChatGPT users to buy directly from Etsy and “over a million Shopify merchants” (Glossier! Spanx! Vuori!) and open sourced the Agentic Commerce Protocol developed with Stripe. Nathan Lambert tweeted: “This is a way bigger deal than the Claude release.”

Eyes bigger than stomach. AIs bigger than wallet

Three things Bain first-years are definitely texting each other right now:

• bruhhh my associate just asked me about microsoft agent mode. i’m so cooked

• pls god don’t let my bf find out about instant checkout in chatgpt. i caught him whispering to chatgpt the other day and i don’t want to wake up to a hundred tubs of protein powder and a thousand dollar credit card bill 💀

• babe wake up new claude just dropped

Okay, maybe just me to myself for the last one (me n who fr 😩), but as I also tweeted, one of the biggest milestones of Amazon’s early e-commerce strategy was one-click checkout. A single click to buy. No re-entering billing, shipping, payment. Amazon held the patent for almost 20 years, and only then in 2017 did Shopify launch its one-click Shop Pay solution.

Instant Checkout in ChatGPT smells really similar to OG one-click checkout. It appears OpenAI is trying to capture more of the public consumer market while Anthropic is going the enterprise route. Yesterday I wrote about the Anthropic Economic Index which analyzed Claude usage patterns globally; OpenAI published a ChatGPT usage study on the same day as Anthropic published their update, highlighting faster growth in non-work-related messages. Although alignment and safety have been less publicly central to OpenAI’s research agenda relative to Anthropic’s, I take small solace in Chain of Thought Monitorability (2025), with contributors from DeepMind (including one of my favorite bloggers Neel Nanda), OpenAI, Anthropic, METR, endorsed by the “Godfather of AI” Geoffrey Hinton. I discovered this paper via a tweet by Miles Brundage:

chain of thought reasoning… Improves almost everything (sycophancy, hallucination, obvious misuse, etc.) [in language models] if you don’t screw up RL, and many things even if you do… not to even mention the benefits in terms of monitoring the chain of thought to detect shenanigans - I was just talking about improving the outputs themselves

The paper asks researchers to “study what makes models monitorable”, e.g. measuring readability, reasoning and task capability on “single forward pass”, and “track [chain of thought] monitorability of their models and treat it as a contributor to model safety”. They also call out, just as economists are begging researchers and policymakers to get ahead of measuring and guard-railing the socioeconomic impact of AI, that frontier researchers need to proceed with caution when weighting chain of thought monitoring as factors to assessing safety:

more advanced, situationally aware models… may become able to avoid detection by overriding their propensity to think out loud, and, when reasoning is required, to deliberately obfuscate it

We find it plausible, though far from clear, that some forms of process supervision could substantially improve model alignment by effectively intervening on model reasoning. Research strategies that aim to unconditionally preserve [chain of thought] monitorability in its current forms may miss productive safety opportunities of this kind

safety measures for future AI agents will likely need to employ multiple monitoring layers that hopefully have uncorrelated failure modes

What Would Claude Do? (WWCD)

My biggest fear in all of this is that AI safety and alignment will simply fall to the wayside as influential labs start openly driving their research to serve more consumer-altering, consumption-enabling behavior.

Of course, there are time-saving benefits to automating away the decisions involved in making mundane, routine purchases. A hundred years ago, supermarkets where you could find fresh produce and dry goods in the same place didn’t exist; in the past seven years Amazon Go has tried and largely failed to remove even the cashiers from supermarkets, while Instacart, Thrive Market, Walmart, and Amazon Fresh survive in the “groceries delivered directly to your door” space.

But this type of shopping isn’t what I’m worried about with ChatGPT Instant Checkout and the Agentic Commerce Protocol. We should be concerned that we’re rolling out agents that can encourage and enable people to make purchases while simultaneously setting model development goals to become more competent at entry-level professional tasks — while there are exciting avenues to bridging economic gaps here, perhaps diversifying and driving more business across Etsy and Shopify merchants, there is also staggering, immeasurable risk to exposing this infrastructure to scammers and fraudsters undoubtedly hard at work figuring out how to capture the eyes and wallets of unsuspecting ChatGPT shoppers.

I’m curious to see if and how Anthropic responds to OpenAI’s shopping integrations. I’m betting we won’t see such changes to Claude, at least not too soon. The romantic in me wants to believe it’s because Anthropic is so committed to its safety-first mission that they’d feel the need to invest a lot more into solidifying consumer-protecting model behavior before releasing something like this, but the realist in me understands that Anthropic has also placed a lot of weight on alignment and safety for research credibility, and that Claude’s piece of the AI pie is looking more and more like enterprise.

Regarding external influences to our behaviors and personalities, near puts the tension well in a blog post “Personality Basins” I found via maja:

One interesting way to frame personality capture is by combining it with the concept of attention economics. All of the apps on your phone want to turn you into the type of person that uses them all day because that is beneficial for their revenue models. In many cases this is mutually beneficial, but it’s nonetheless clear that the cat and mouse game is starting to favor the felines more and more over the last two decades as they have learned to perfect their craft of user acquisition, retention, and ARPU maximization.

near also quotes Dario Amodei on Dwarkesh Podcast:

“I’ve just seen cases with a number of people I’ve worked with, where attaching your incentives very strongly to the approval or cheering of a crowd can destroy your mind, and in some cases, it can destroy your soul.”

On forming independent views on AI safety, Neel writes:

Don’t be a monk - you form an inside view by going out in the world and doing things - not just by hiding away and thinking really hard

Inside vs outside views is a spectrum - there’s no clear division between thinking for yourself and deferring. Forming inside views starts out by deferring, and then slowly forming more and more detailed models of where I’m deferring and why over time… Getting here looks like downloading other people’s gears level models into your head, and slowly combining them, deleting parts you disagree with, adding ideas of your own, etc

The AI of the tiger, the thrill of the fight

Storm chasers cite curiosity and adrenaline as primary motivators to why they continue to race towards the danger. As we learned from the 1996 classic Twister, there’s a complex human bond formed in the shared, high-stakes endeavor to get as close as possible to something much bigger and more powerful than any of us, to hold it close enough to appreciate the stillness at the center, while constantly trying to predict and sidestep its violent edges.

That’s how the AI race feels to me right now. The hurricane is huge and looming. It’s moving and growing so quickly we won’t be able to outrun it entirely, and we can’t guarantee that our shelters will protect us. Storm chasers are driving and flying at it from all directions. There’s an eye in there, a central place of calmness and peace, which some of us might be lucky enough to experience, but most of us will have to avoid entirely.

There is already, and will continue to be, destruction in the path of the storm. The difference here is that it’s of our own making. We increase our chances of survival as a society by building and enforcing socioeconomic and cultural shelters in the form of technological model guardrails, social incentives to develop and retain independent thought, forging more personal connections and stronger communities.

And, if we can accelerate the passage of the unpredictable, unforgiving elements of the storm, we’ll be rewarded with fresh rain, nutrients in the soil, and a foundation on which to construct something stronger and healthier.

I used Open Paper to help me parse through the papers and technical reports I referenced in writing this blog post. Thanks to Saba, all my readers can use code VALSTECHBLOG for 50% off the Researcher Plan. Thanks to Ashi (check out her blog) for connecting us!

“The tragedy of the AI commons”

In 2021, computer scientists and philosophers Travis LaCroix and Aydin Mohseni published The Tragedy of the AI Commons. They described how the World Economic Forum had “identified almost three hundred separate efforts to develop ethical principles for AI”, but in large part because there was no way to enforce these guidelines, studies “suggested that such declarations have little to no practical effect”. Referencing ecologist Garrett Hardin’s 1968 formulation of the tragedy of the commons, they identified a core social dilemma (“no one has an individual incentive to cooperate, although mutual cooperation would lead to the best outcome for all those involved”), developed a stochastic evolutionary game theoretic model to simulate models of cooperation at research population scale, and concluded with a set of morals, including “Lowering the cost of cooperation increases the likelihood of cooperative success” and “Small, decentralised groups may benefit sustained cooperation for responsible AI research”. Importantly,

It would also be a mistake to interpret our results and arguments as a dismissal of current proposals or guidelines for the safe and ethical development of AI. To the contrary, we believe that such efforts are laudable and necessary; but to give ourselves reasonable odds for success, we must appraise ourselves of an understanding of the basic dynamics of such coordination problems

Earlier this year, Sahil Bloom published “AI and the Tragedy of the Commons” to his newsletter, providing an industry and life optimization angle to what LaCroix and Mohseni were approaching as a policy and collaboration problem. He outlined “possible (and scary) implications” of the messy socioeconomic incentives currently defining AI development and integration into work and life, most notably a potential future where “People are more efficient and productive, but lose a sense of texture, meaning, or purpose in the work… [having] unwittingly optimized the life out of their life.”

While LaCroix and Mohseni took a theoretical, statistical approach to modeling populations and concluded that costs to collaboration needed to be lowered to offset opposing incentives, Bloom took almost exactly an opposite approach referencing real survey data and anecdotal evidence and landed at the same conclusion:

It’s about how we make it rational for each actor to act with the longer-term future of our shared commons in mind… how we make it rational to work towards a future of human flourishing… In my view, it’s the only question that matters for our future: How can we lower the cost of long-term thinking?

A second-wave creative class

In 2022, Wessie du Toit wrote “The Rise and Fall of the Creative Class”:

By and large… the post-crash [post-2008] world starved and scattered the creative professions, squeezing budgets and forcing artists into the grim ‘independence’ of self-promotion on digital platforms.

One result of this was a general revulsion at capitalism, which partly explains why artisan ideals and environmentalism became so popular in creative circles. But despite this skepticism, and even as career prospects withered, the creative lifestyle maintained its appeal. In fact, the 2010s saw it taking off like never before…

Intellectual property really is the basis for growth and high incomes in today’s economy… the valuable intellectual property is increasingly concentrated in the tech sector. It is largely because IT and software are included that people can still claim the creative industries are an exciting area of job creation…

The tech world is, of course, a very creative place, but it represents a different paradigm of creativity to the arts and media vocations we inherited from the late-20th century. We are living in a time when this new creativity is rapidly eclipsing the old, as reflected by the drop in arts and humanities students, especially in the US and UK, at the expense of STEM subjects.

du Toit’s Rise and Fall came 20 years after Richard Florida’s Rise of the Creative Class. Florida wrote: “The key to economic growth lies not just in the ability to attract the creative class, but to translate that underlying advantage into creative economic outcomes in the form of new ideas, new high-tech businesses and regional growth”. In the years following, many critiqued Florida’s Creativity Index for economic growth (which factored in the creative class share of the workforce, innovation measured as patents per capita, the Milken Tech Pole Index as a High-Tech Index, and a “Gay Index, a reasonable proxy for an area’s openness to different kinds of people and ideas”) as exacerbating existing urban inequality and accelerating gentrification. Frank Bures, in the The Fall of the Creative Class:

[Richard] Florida’s idea was a nice one: Young, innovative people move to places that are open and hip and tolerant. They, in turn, generate economic innovation. I loved this idea because, as a freelance writer, it made me important. I was poor, but somehow I made everyone else rich! It seemed to make perfect sense… Soon after we arrived [in Madison], however, I was sitting at my desk wondering where all these creative, self-employed bohemians might be…

In a four-year, $6 million study of thirteen cities across Europe called “Accommodating Creative Knowledge,” that was published in 2011, researchers found one of Florida’s central ideas — the migration of creative workers to places that are tolerant, open and diverse — was simply not happening.

“They move to places where they can find jobs,” wrote author Sako Musterd…

“Even as an arts advocate,” said Mel Gray, “I want to do it for the right reasons.” The right reason, we can now say, is that these things are good in themselves. They have intrinsic value. They make the place we live more interesting, livelier, healthier and more humane… They do not make it more profitable.

Bures managed to capture in 2012 a sentiment so timeless it could just as easily have been written today in 2025:

…this was Florida’s true genius: He took our anxiety about place and turned it into a product. He found a way to capitalize on our nagging sense that there is always somewhere out there more creative, more fun, more diverse, more gay, and just plain better than the one where we happen to be.

But I’ve been down that road, and I know where it goes. I know that it leads both everywhere and nowhere. I know you could go down it forever and never quite arrive. And I know now that it may be wiser to try to create the place you want to live, rather than to keep trying to find it.

If the economic definition of the creative class was bolstered in the 2010s by tech work, in the 2020s we’re seeing the start of a massive shift in both what is viewed as creative work and how much time we’re spending on “manual / automated” vs. “creative” tasks. Sahil Bloom fears we’ll automate away so much of our work and our lives that we’ll lose the parts of ourselves that make us appreciate being human; Frank Bures warned us more than ten years ago as a creative in the economic transition following the Great Recession that we’d need to grapple with the dichotomy between “intrinsic value” and realized economic value, and actively mold the world we desire from scraps.

The reality is that the sociologists and economists defining and studying the “creative class” and economic ramifications of it are themselves members of a highly privileged, small socioeconomic subset of the world. Just as LaCroix and Mohseni qualified their study as not a critique of ethical AI guidelines, rather a call to action in areas where implementation was lacking, I note that socioeconomic studies of how AI is changing the workforce are not fruitless, but need to be coordinated at a global level and acted upon to shape real outcomes.

The latest Anthropic Economic Index report raises far more questions than it answers

Anthropic recently published updates to its Economic Index (N.B. Still no Anthropic Environmental Index which I was hoping for back in March; guess we’ll have to survive on Microsoft Environmental Sustainability Reports until then), observing that AI (Claude) usage “strongly correlates with income across countries”, “regional usage patterns reflect distinctive features of the local economy, [e.g.] elevated use for IT in California, for financial services in Florida, and for document editing and career assistance in DC”.

If the productivity gains are larger for high-adoption economies, current usage patterns suggest that the benefits of AI may concentrate in already-rich regions—possibly increasing global economic inequality and reversing growth convergence seen in recent decades…

This concentration in advanced economies with limited population sizes reflects their established patterns as technology pioneers. For example, both Israel and Singapore rank highly in the Global Innovation Index—a measure of how innovative different economies across the globe are—suggesting that general investments in information technology position economies well for rapid adoption of frontier AI. Overall, these economies can leverage their educated workforces, robust digital infrastructure, and innovation-friendly policies to create fertile conditions for AI…

The geographic patterns of AI adoption—where it is used, for which tasks, and how—suggest that in order to realize the potential of AI to benefit people across the globe, policymakers need to pay attention to local concentration of AI use and adoption, and address the risk of deepening digital divides.

The report doesn’t propose any meaningful policy suggestion or real-world advice beyond highlighting that regionally different AI usage patterns may become a global economic problem and that policymakers should pay attention.

Researchers affiliated with the National Bureau of Economic Research have just published a working research agenda for “the economics of transformative AI”:

AI capabilities have improved radically in recent years. Our institutions, organizations, skills, and economic models are struggling to keep pace. In this growing gap lie the greatest risks of the coming decade, as well as the greatest opportunities. We urgently need to improve our understanding of AI’s economic implications

They start tackling this problem by proposing precise questions to be answered: “we must define the fundamental questions, laying the groundwork for a productive research agenda… A question well-posed is half answered”:

As we potentially transition to a post-work society, economics will play a crucial role in shaping new institutions to support human flourishing. What frameworks can help us analyze what provides meaning when TAI (transformative AI) can handle most economic tasks? How might the distribution of meaning-generating activities be optimized for social welfare? Can TAI itself help create or facilitate new sources of meaning and fulfillment? How can we equitably distribute the benefits of TAI, not just in terms of material wealth, but also in terms of access to fulfilling activities and meaning?

They also mention the research use of LLM-based agents to simulate large-scale economies, first to drive “new insights into the economy and how AI might reshape labor, consumer behavior, industrial growth, and unforeseen economic bottlenecks”, and then more widely integrated scenario planning “[to bridge] technical, economic research, and policy communities, as well as the broader public”.

This research agenda is a promising and more tangible attempt to focus our understanding and actions around building a future in which humans and machines coexist safely and productively. I appreciate that it formulates many of the questions I had reading the Anthropic Economic Index report into real research areas. I have faith that many of our best and brightest minds are hard at work implementing these ideas, but as the authors repeatedly call out, there is an incredibly urgent unmet need:

While significant resources are being allocated to the technical development of TAI, there has been comparatively little investment in understanding its economic implications and in research to inform policy on how to steer TAI in desired directions and prepare for the impacts.

Economies of taste over scale

I previously wrote about recommendations and curation in a rapidly compressing, increasingly online-first attention economy. Another marker of cultural and regional shifts in consumption we are already observing is the proliferation of recreational “AI slop” (professionally, HBS terms this “workslop”). We’ll start to see a real divergence in the coming decade between consumers (and corporations and governments, other socioeconomic actors) that value and prioritize high-quality products and media and those that begin to accept lower-cost, lower-quality, more automated and mass-produced forms of entertainment and software. We’ve seen this before: the invention and optimization of assembly lines and manufacturing plants to mass produce things like cars, furniture, and mobile homes, turning years, months of human labor into single-digit days; the development of factory farms, pesticides, preservatives, freezer and microwave technology to flip food preparation from a full-time life-sustaining endeavor to a seconds-long afterthought. We doubled, tripled our free time with these inventions. Collectively, we were able to converse more, read and write more, make more art, music, film and television.

But, having automated away the hard work of clothing ourselves, building and keeping our homes lit, warm and safe, and putting food on the table, having enjoyed the fruits of our ingenuity via new forms of media, we’re now trying to automate away the labor that goes into creating that media (Spotify and Netflix both have complicated relationships with AI-generated material, where in the eyes of critics, Netflix already had a quality problem producing originals too quickly; Suno recently turned a raw recording of a jam session into a shockingly good, professional-level mixed and mastered track for my band, but also completely changed our lead singer’s voice).

A master woodworker who had trained at The Krenov School of Fine Woodworking once told me that the craft of carpentry has seen many forms, early on as a royal luxury, later as a common farm and homestead necessity, and then mostly as a nameless, thankless task in modern homes. But because of how the craft has changed, each individual woodworker now holds incredible responsibility to preserve and advance their craft, adapting to the world’s changing tastes while acknowledging the complex history and industrial processes already involved in their work.

If we put down our tools and turn back because the forest we’re in is too wild and unknown, we miss the chance to carve out a better future for ourselves.

OpenAI just signed a five-year, $300 billion contract with Oracle, driving a 30%+ Oracle stock price spike earlier this week even with an earnings miss, making Larry Ellison the world’s richest man for one day and the phones in my office light up with texts from newly-minted millionaires.

Prediction market platforms Polymarket and Kalshi also lit up with activity in contracts trading outcomes of Elon Musk remaining the richest in the world through 2025 (Polymarket: “resolve to Yes if Elon Musk is any rank other than #1 on the Bloomberg Billionaires Index at any point between June 30, and December 31, 2025, 11:59 PM ET”) and a bet (currently ranking Musk, Ellison, and Bezos) at the world’s wealthiest this year (Kalshi: “If [bet] is the wealthiest person in the world on Dec 31, 2025, then the market resolves to Yes. Outcome verified from Forbes.”).

Preliminary round: Musk vs. Ellison

“Ellison cemented his position as the biggest wealth gainer in Bloomberg's ranking this year with a $191 billion increase as of Wednesday's close. Musk led the list of greatest wealth losers with a $48 billion decline this year. Those shifts in fortunes reflect Oracle stock rising 97% and Tesla stock sliding 14% since the start of January.” — Business Insider, September 11, 2025

I was going to start with a deep dive on financial derivatives, at least futures contracts, because understanding these specific financial instruments makes understanding prediction markets much easier. Then I decided this isn’t a finance blog so I’m just going to let Investopedia (every first-year analyst’s best friend) and a Wharton blog do most of the explaining here:

• Event Contracts, Investopedia

  • “An event contract pays its face value if the event happens and $0 if it does not. The pricing of event contracts is based on other speculators’ judgments on the likelihood of an event.”

  • “Unlike traditional futures contracts, which are staked to the expected prices of commodities, financial instruments, or indices, event futures are priced according to the perceived likelihood of particular events winning or losing. The Chicago Mercantile Exchange (CME) has been offering event futures since 2022, allowing clients to choose yes or no on the predicted closing price of Bitcoin, commodities, and indexes. Kalshi is another major purveyor of event contracts; the online platform has been the most out front defending event contracts’ legitimacy while offering a far wider variety of them than the CME.”

• A Primer on Prediction Markets - Wharton Initiative on Financial Policy and Regulation (Liu & Bierwirth, 2025)

  • “Prediction markets are a little bit more nuanced than derivatives markets since they can cover solely event-based outcomes instead of price-based action in traditional derivatives exchanges. This opens up a realm of opportunity for traders both individual and institutional depending on the investments they pursue. This encourages a new market of potential hedges.”

  • Interesting prediction markets history in Prediction Markets (Wolfers & Zitzewitz, 2004): DARPA proposed establishing a Policy Analysis Market in 2003 “based on indices of economic health, civil stability, military disposition, conflict indicators and potentially even specific events… concept was to discover whether trading in such contracts could help to predict future events and how connections between events were perceived… Critics savaged DARPA for proposing ‘terrorism futures’… best-known prediction market among economists [at time of writing] is the Iowa Electronic Market” (election markets). Featured TradeSports (allowed trading in “current events” and “political futures” on top of sports markets; sports betting market now filled by DraftKings and PrizePicks) and the Hollywood Stock Exchange (Downton Abbey is currently performing well).

Regarding the “world’s richest person this year” trade, see Kalshi’s contract terms:

• Underlying is defined as “persons on the Forbes Real-Time Billionaire

  List”

• “Settlement Value for this Contract is $1.00”

e.g. Traders currently may buy a Yes on Larry Ellison for 27 cents; if Forbes ranks Ellison first in the world on December 31, 2025, holders of this contract are paid out $1 per unit less the transaction fees they paid Kalshi.

On the other side of each trade is another trader taking the opposite side, a similar pricing model to a traditional stock market, although trader demographics and behaviors might look very different in these markets compared to those of international stock exchanges.

N.B. Clicking into these contracts shows us just how fast and loose the terms can be, and how messy calculations like taxing gains on these bets could get. Protos correctly questions the validity of the Polymarket market remaining open given its resolution criteria on Musk’s Bloomberg standing bet — “Yes if Elon Musk is any rank other than #1 on the Bloomberg Billionaires Index at any point between June 30, and December 31, 2025, 11:59 PM ET” — “if someone was refreshing their browser consistently, it’s possible that the webpage displayed Ellison in the #1 spot at some point yesterday”.

It’s easy to imagine how traders in this market might be incorporating Oracle and Tesla stock price projections, Tesla’s board of directors proposing a trillion-dollar compensation package for Musk (contingent on Tesla peformance), and any number of other data points and opinions into their personal assessment of the worth of these contracts, informing their buy/sell decisions and setting marketplace contract prices.

It’s also easy to imagine how reporting capital gains from prediction market winnings, currently considered by the IRS as taxable income, could get complicated if tracked on a supposedly “decentralized” system. (More on this later.)

The playoffs: Polymarket vs. Kalshi

“Prediction market firms Kalshi and Polymarket are both eyeing multibillion-dollar boosts to their current valuations, according to The Information, citing anonymous sources. Per the report, Kalshi is ‘close’ to raising funds at a $5 billion valuation, while Polymarket has received an offer that could push its valuation as high as $9 billion” — LinkedIn News, September 13, 2025

Polymarket and Kalshi markets both out-predicted Nate Silver in the 2024 US presidential election. “The argument for prediction markets' accuracy is that people are more decisive with their money, picking outcomes they think will happen, not just ones they want to happen… They’re also ways to distill the daily news cycle…” but “didn't forecast the popular vote as precisely”. “They have their own imperfections, just like all of the other metrics do… [supporters of certain political candidates may find ways to place wagers to use their influence]… I suspect as these markets grow more prominent, it's possible their usefulness will decline” (Broughel).

Almost a year after the election prediction upset, Polymarket user “romanticpaul” bet on the Swift-Kelce engagement “less than 22 hours before the viral announcement dropped”, turning a more than $3,000 profit.

So what sets Polymarket and Kalshi apart? What justifies multi-billion dollar valuations for both? What regulatory challenges have they faced? Is one better-prepared than the other for long-term success?

Polymarket: “hybrid-decentralized order book”, emerging from years of US regulatory friction

In 2022, just two years out of stealth, Polymarket blocked US-based users after a settlement with the CFTC to pay $1.4 million and “facilitate the resolution (i.e. wind down) of all markets displayed on Polymarket.com that do not comply with the Commodity Exchange Act (CEA) and applicable CFTC regulations”.

This month, the CFTC approved Polymarket to re-launch in the US, which comes after Polymarket acquired CFTC-licensed QCEX for $112 million in July and announced an official partnership with xAI in June.

Polymarket traders think the odds of a 2025 US launch look pretty good:

Polymarket has always been “crypto-enabled”; it’s sometimes described as “decentralized”, although official documentation is careful to list a “hybrid-decentralized” Central Limit Order Book.

Annotated excerpts from a technical summary of how Polymarket works:

• CLOB - Central (orders collected in one place) Limit (supports limit orders) Order Book: “the order book is purpose-built for trading binary options. When a user places a limit order, the system also displays a mirrored order for the opposite outcome… This design is based on Polymarket’s core assumption: one YES share and one NO share always equal one dollar… Because of this, any order can be interpreted as its opposite”

• “Off-chain, there’s a backend operator responsible for tracking and matching orders, as well as preparing the data needed for market orders… When a user creates an order, they generate a data structure signed with their private key using the EIP-712 standard. Orders stay off-chain until execution, which lets users update or cancel them without needing on-chain transactions.”

• “On-chain, a smart contract receives that data and handles the calculations and actual order execution… Once the final outcome is confirmed by the oracle [“decentralized optimistic oracle” aka third-party data source used as resolver], a reportPayouts call is sent to the CTF contract [handles order processing, fees, etc.]. This stores the result on-chain. After that, token holders can call the redeemPositions method. It burns their shares [remember order mirroring; burn shares by removing them from circulation and paying out users] and returns their portion of the collateral.”

The sooner Polymarket is able to legally re-launch in the US, the sooner it will be able to capture additional market share.

Kalshi: “first fully CFTC-regulated event contracts market in the US”, making crypto and AI investments

“A federal judge has cleared the way for people to bet on U.S. election results for the first time… struck down a decision from the [CFTC] that prohibited gambling companies and users from betting on the outcome of Congressional races… U.S. District Court Judge Jia Cobb ruled in favor of the New York-based company Kalshi” — The Hill, September 12, 2024

“In a Sept. 2023 decision, the [CFTC] told Kalshi it could not offer yes-no prediction bets on which party would control the House of Representatives and the Senate, ruling that it constituted illegal gambling activity that is contrary to the public interest… Such bets “could potentially be used in ways that would have an adverse effect on the integrity of elections, or the perception of integrity of elections — for example, by creating monetary incentives to vote for particular candidates,” the agency wrote.” — AP, September 11, 2024

Kalshi has not operated without its fair share of regulatory hurdles, but its main differentiating factor from Polymarket is that it was able to stay within regulation in the US in the three years and counting Polymarket has been officially locked out.

Still, regulated access to US customers is just a slice of the global betting market:

“Kalshi operates as a [CFTC]-regulated platform within the United States, requiring dollar deposits and traditional KYC procedures, which limits its user base but provides regulatory certainty for American participants… Polymarket, conversely, operates on Polygon using USDC settlements, offering pseudonymous trading and broader global accessibility but facing restrictions in the U.S. market due to regulatory uncertainties” — The Block, September 8, 2025

For example, Kalshi processed $875 million in volume in August, while Polymarket processed $1 billion. The future success of Kalshi will depend on collective user appetite for betting markets and relative incentives of trading on these and similar betting platforms especially if its main competitor Polymarket’s US re-launch goes smoothly. It’ll be a question of how many winners can co-exist in this space, much like the conversation around how many players can co-exist in the AI space.

Just as many investors are questioning when the AI bubble will burst, just as the crypto wave continues to ebb and flow, these two major players entering their own unprecedentedly global regulatory race will drive similar volatility in the prediction market industry.

Play-by-play: Sports betting

Sports betting is a $100 billion market. For a practice that started thousands of years ago with chariot and horse racing in ancient Greece and has long been associated with gambling and crime, betting markets are rather poetically plagued by often-paralyzing regulatory hurdles. The US Supreme Court federally legalized sports betting in 2018, formerly banned by the Professional and Amateur Sports Protection Act, “allowing each state to set its own sports gambling regulations without federal restrictions”.

Technically, sports betting is currently illegal in California, but there are ways around it; PrizePicks is the “Official Daily Fantasy Partner of the San Francisco Giants”, legally classified a “fantasy sports operator” and not considered “online gambling”. DraftKings Sportsbook does restrict users from placing real money bets in states where sports wagering is not permitted. DraftKings is working hard to get California on board.

Because event contracts are regulated as financial instruments not structured and regulated in the same way as sportsbooks, prediction market platforms “could reshape the sports betting landscape by introducing lower-cost alternatives to traditional state-regulated sportsbooks”; the caveat, of course, is this only becomes a sustained market if “[these platforms] can withstand CFTC regulation”.

With college football, the NFL, US Open, and MLB in full swing, just to name a few, trading volume on Kalshi recently hit $300 million in one weekend, almost $200 million on football Sunday, with more than 95% of all trading involving sports contracts (Kalshi launched sports contracts at the start of 2025).

Kalshi specifically differentiates trading event contracts from sports betting and stock trading in the mechanics of how their platform works. Kalshi matches counterparts such that someone needs to take the other side of what you are betting (charging transaction fees per trade), unlike traditional sportsbooks that accept bets on both sides of outcomes markets and charge commission on losing bets.

There are open challenges from state regulators:

“[Massachusetts] Attorney General Andrea Joy Campbell alleged in a filing that sports event contracts, which Kalshi introduced in January 2025, violate the state's sports wagering laws, which require operators to be licensed. Campbell is asking for a court to block Kalshi from offering sports prediction markets in the state without a license, as well as seeking monetary and other relief.” — CoinDesk, September 12, 2025

But these are not regulatory challenges specific to prediction market platforms. As much as sports betting keeps prediction market platforms afloat via pure trading volume in the seasons between other high-volume events like elections, more at risk in this regulatory space are platforms like DraftKings and Fanatics built on traditional sports books. The question now is how long dedicated sports betting platforms will survive in a landscape increasingly overtaken and defined by new players creating new markets with new rules.

As they say in the rodeo, “Cowboys may come and go, but the rodeo is forever”. That couldn’t be more true of the cowboys riding these modern bull markets.

Friends like this remind me of the immeasurable impact of openly developing and sharing personal taste. In the past week alone, sparse peer-to-peer recommendations (read: friends that know my taste, and whose tastes I trust, telling me I’d really like something) have led me to a clothing purchase I’m happy with, an unexpectedly hilarious TV show that broke me out of a TV-watching slump, and a ticket counter purchase 15 minutes before showtime of a film that lived up to all the indie hype. However, and importantly, often a recommendation that by any other metric should hit just doesn’t do it for me; whether that’s a product of an expectations mismatch (don’t buy into the same hype), preferences mismatch (different personal weighting of dimensions), or inherent randomness (off day), these are the factors that are hard to model when we try to scale personal taste.

What large-scale recommendation systems get right

Using the Paine framework, I’ll begin with a starting thesis.

Thesis: More data feeds better recommendations.

YouTube has consistently surfaced high-quality content made by lesser-known videographers to my home page of recommendations. YouTube’s sell to the user is that we’ll get increasingly personalized recommendations in exchange for allowing YouTube to collect our “user history and context” (watch history and interactions), which is fed into its recommendation system:

The model objective is to “predict expected watch time” using weighted logistic regression where clicked impressions (positive training examples) are weighted by video watch time and unclicked (negative) impressions are unit weighted:

Optimizing for video watch time across users understandably helps YouTube’s business goals as an entertainment and advertising platform incentivized to maintain a marketplace for users to watch videos and creators to create content.

However, a recommendation system trained on a population-scale dataset only works well if it knows enough about “user history and context” to draw meaningful preference connections between users. Extrapolating what one user might like to see next (or would watch the longest) from a set of inferred similar users and their watch behaviors yields to outcomes such as my YouTube homepage being filled with a hundred viral, sensationalist videos for each high-quality recommendation.

So while YouTube has the advantage of being the only platform where you can find a detailed, well-researched tutorial by anyone on anything, it lacks the inherent curation of platforms like Netflix and MUBI. For selective consumers, using YouTube to find quality content feels like a game of cornhole.

Less noise, more signal

“TV was in a language I didn’t understand, there was nothing to buy, no advertisements anywhere, so all I’d be doing was walk around, think and write. My brain felt like it was at rest, free from the consuming frenzy… Maybe it could have seemed like boredom at first, but it quickly became very soulful.” — Celine, Before Sunset (2004)

Counter-argument: Recommendations systems are the capitalistic product of social media. Eschew data in favor of specifically curated experiences.

In 2021, I spent at least half of my free time on TikTok and Twitter until I finally tired of my TikTok For You page and deleted my account. In parallel, I realized my Twitter feed was becoming an infinite scroll of distraction with decreasing reward.

I experience this pattern with most platforms I try. There’s an introductory period, assuming I get past the signup flow, where my delight factor is high enough to keep me using the platform. (The x-axis here is time elapsed since signing up, not the time / effort it literally takes to engage with the content per interaction, which is a long-form vs. short-form media conversation that requires a separate essay.) After some time, my delight factor starts to oscillate around a point where it’s increasingly unclear to me what value I’m deriving from the platform. Enough time spent under the quit threshold and I’ll stop coming back for more. This might be months or years. Regardless, without a periodic boost (new socials, new content, new interactions) the baseline delight of novelty generally bottoms out below the quit threshold.

But my delight factor matrix across all my services is entirely personal and variable over time, a mixed product of confounding variables including my mood, what’s going on in my life when I’m consuming the content, who I’m sharing it with.

The dumbphone movement and the sheer fact that we as a population have established and documented social media detox methods show us that there is a strong and growing market here. Whereas traditional advertising may have prioritized capturing attention in a competitive entertainment landscape, now we’re seeing products that are explicitly vending distraction as a service. Scroll Instagram during your commute to make time go faster. Stare at Twitter to avoid making eye contact with anyone else when you’re waiting at the doctor’s office.

The problem with ignoring data when curating taste is that a certain degree of “online” is now a prerequisite to modern forms of discovery. Critically, media is now both fueled and judged by its online reception. Organic (word-of-mouth) and sponsored social media marketing drives box office outcomes, concert turnouts, and series renewals. You could see it as an internet-age tragedy of the commons where everyone is allowed to create, everyone is allowed to consume, and fringe content either develops enough of a cult following to become mainstream and economically sustainable or (so much more commonly) ends because the creator is not incentivized to continue developing content in a data-driven online economy.

Curation in the age of hypnotic media

“Good writers tend to be good and broad readers, and their sources tend to be a better curation tool than any other.” — Henrik Karlsson

Rebuttal: There’s a way to use social media platforms to inform ourselves and others, but effective and productive personalization requires effort and time to curate that may not be worth it for everyone.

My [fig 3] above is my own delight factor charted over time for an arbitrary service. This curve changes from individual to individual and that’s the retention curve no recommendation system is going to get precisely correct for anyone, because I can’t even properly quantify it for myself. My current formula is no TikTok, reduced Twitter consumption, and no saved YouTube watch history, just a collection of YouTube videos I know I like and following a limited path from YouTube’s recommendations formed from just a few data points. I use Netflix to watch specific peer-sourced recommendations and then personally find Netflix recommendations utterly useless. I’ve found a little more success with MUBI because it’s a highly curated collection of films, but enough flops to keep me wary of my subscription.

The social element of Spotify and Letterboxd encourage music and film discovery and are a way for friends to provide loose curation to each other. The problem of mapping peer-to-peer recommendations to individual preferences remains to be solved. Crucially, there’s no consistent way to collect quantifiable feedback on these recommendations. So much is kept in our brains but not communicated. So many things we may also not realize we would like, or need to be in a certain headspace or time of life to appreciate.

Even the acts of researching and writing this post have made me reflect on why I’m doing this:

“When your goal is not only to present ideas but to make them palatable and digestible for a distracted audience, the boundaries of intellectualism start to blur. Ideas are no longer for discussion, they’re for consumption.” — The death of the public intellectual

The main point I’m trying to get across is that it’s a constant but rewarding battle to go from passively consuming recommended content to proactively curating your methods of sourcing and discovering new media. Personally, it’s not even that my recommendation signal-to-noise ratio overall necessarily increases in doing so, but that the sparklier outcomes (unexpected, surprisingly good, random recommendations, often social) weight higher in my personal ranking.

Parting words. I’ll simply leave you with this:

Don’t forget to call me beep me on X

“Life-changing”

I describe many things as “life-changing”. Kaveh Akbar’s Martyr! was life-changing, as was Tommy Orange’s Wandering Stars, as was learning that Akbar and Orange “[traded pages] every Friday [while writing these books], and… called it band practice”. This Rolling Stone Australia feature of a relative newcomer to the indie scene (Borderline) was life-changing. So was Khruangbin’s Fender Signature Session (yes! I’m so basic and I love it!) that reminded me of the meditative power of improvisation and letting an instrument of choice carry you forward (guitar here, but can also be freeform writing, walking meditation, etc.). I once repeated the phrase to describe a salad so many times over the course of one dinner that it was presented at the next one as Life-Changing Salad (the life-changing ingredients are insiders-only but I’ll tell you if you DM me).

I say that things are life-changing, that I’m obsessed, that I’ve fallen in love with something, about once a week. It’s true every time, which is why I’ll continue to say these things whenever I want. Naturally, after two hours listening to Professor Sarah Paine and Dwarkesh Patel last night the first thing out of my mouth was “I’m literally in love”. “With what?” our OG Dwarkesh lecture series attendee asked, to which I said “all of it, everything, everyone”. As much as I stand by these statements, it also means I struggle to communicate the gravity and intensity of this life-changing moment, which is why I’m writing about it now.

Yesterday’s session (“Why Russia Lost the Cold War”) was the last of six in-person lectures and interviews organized by Dwarkesh Patel featuring Professor Paine. At the time of this post’s publication the recording isn’t yet published on Dwarkesh’s YouTube. I typed up my lecture notes last night for anyone looking for a high-level overview of Professor Paine’s talking points and an outline of what to research next.

I considered adding a personal section to my takeaways, but to be completely honest the LaTeX was not giving ‘personal’, so here I am using Substack as God intended (to share my personal reflections with my 1.5 dedicated readers).

The Paine Meta-Framework

Professor Paine is very good at frameworks. Specifically in the context of analyzing history, such a framework might look like:

• Thesis/Starting Argument

• Counter-Argument (paper requirement from Naval War College)

• Rebuttal (different perspective, not your starting argument)

I noticed early in last night’s lecture that this framework to outline and trace through parts of history, to approach a problem from different angles and offer supporting evidence for each part, then synthesize a consolidated picture of events and interactions, was remarkably logical and thorough.

Why I’m calling this a “meta-framework”: at risk of sounding like Tim Ferriss, Professor Paine’s lecture was my “aha” moment for realizing the common thread that pulled together a few previously disparate fabrics of productivity and life optimization I’ve been exploring for approximately the last year. I’ll break down this “meta-framework” into three pieces: 1) Questions you want to answer, 2) Research methods to answer them, and 3) Sustaining lessons.

1) Questions you want to answer

In Professor Paine’s lectures, these questions are things like “Why did Russia lose the Cold War?”; “Does America Have a Crisis of Strategic Thinking?”; which we can then begin to answer by going through the combinations of actors and interactions in the system and synthesizing into a thesis, counter-arguments, and some form of conclusion. In historical analysis, the correctness of a thesis is inherently subjective; historians analyze evidence and form arguments that define the discourse.

Since designing resilient, scalable software systems has been on my mind for years, especially last night, I couldn’t help but extend this analysis framework to conversations engineers have on how to build systems to a set of business requirements, where the questions might be “How do we build this system in a maintainable, traceable way?”; “Are we happy with how our system is built? Is it serving our current business needs in an operationally, developmentally-streamlined way?” (almost always no); “How the actual fuck does this system work?” etc. etc.

If we insist that engineers chart out their systems design paths in the same way Naval War College students are required to structure their papers, this might look like (hypothetical engineering problem that is totally not inspired by a recent end-user experience I had with a ticket booking system):

Question: How do we design a system to persist a customer’s hold on a ticket in our backend for longer than a few seconds without requiring them to manually re-enter all their credit card information if the page decides to randomly re-load or stall at payment?

• Thesis: We should implement a local state store and webhook listener that’s always receiving page events so we can write local state to a persistent database if we get any flow-disrupting event notification.

• Counter-argument: That’s a lot of work to maintain and process for every single user session. No webhooks, let’s just write to a persistent database as soon as the user confirms their information and delete when we know we won’t need the data anymore. [See: why careful data management matters]

• Rebuttal: Let’s approach this from a different angle. The frustrating part of the user experience is that the page is reloading unpredictably. Let’s solve that before we over-optimize to solve a symptom rather than address the core problem.

This is of course not a comprehensive set of system design concerns, just a starting example of how this thesis and argument-based framework can be used to structure thought both in historical analysis and forward-looking decision-making scenarios.

Predictably, I started applying this framework to my own life. Here I want to note that Professor Paine’s starting thesis for “Why did Russia lose the Cold War?” was “Reagan won it all!” and the “combined” counter-argument incorporated points from “internal” and “external” perspectives, so if some of my statements seem reductive and silly, it’s because they’re just tools to drive actual, legitimate research and arguments. Put another way, these sub-components need not be strictly delineated.

Question: Why am I stressed?

• Thesis: Too much coffee, too little sleep, too much to do

• Counter-argument: More coffee, sleep less, so you can finish everything you need to do

• Rebuttal: Different perspective. Why is your to-do list stressing you out?

  • Busywork that you feel like you have to do but don’t have the motivation to do? Do the mental exercise - what happens if you don’t do them?

  • Chores that you don’t like to do? How do you make them enjoyable or at least more tolerable?

2) Research methods to answer them

I touch on this in 1). In the mapping exercise stemming from any initial question, the first things that follow are usually more questions. For example, in the ticket booking example, the rebuttal is actually a question of “are we solving for the right thing”, and stem out into research topics of what’s affecting the page load volatility and how to test page rendering with different end users. In my personal example, these second-order questions of “Is the grind stressing you out?” “Are chores stressing you out?” lead to research topics as varied as “What happens if you switch back to matcha?” “What happens if you sleep more now for a fresher brain in the morning instead of losing sleep over a problem that won’t change overnight?”

The results of these research topics, personally, are data points collected over time, as I experiment with modifying individual variables in my lifestyle. I’ve collected consistent enough data on myself over the last year to know that I personally respond well to matcha (not coffee), cutting out alcohol, trying to sleep ~1-8 a.m. instead of 3-9 a.m. (+1h of sleep every night), preferring music, film and long-form content over short-form or low-intellectual stimulation media, learning to cook a few go-to meals I’m proud of without overwhelming myself by trying to cook everything from scratch all the time, and physically and intellectually traveling (immersing myself in a new environment, learning something entirely new, and most importantly, reflecting and applying learning back to my life) as frequently as possible.

3) Sustaining lessons

By no means am I presenting this as a silver bullet meta-framework to solve any problem. The last paragraph is me trying to squeeze a year of concentrated learning and personal trial-and-error into a few sentences; each component, especially since I’m naturally verbose, deserves an essay.

Critically, I haven’t been able to consistently apply my own learnings over the last year. I still watch Veep re-runs when I am objectively learning nothing from it, because I love Veep and it makes me laugh. I still let vegetables rot in my fridge. Even as I write this, I have a pressing work deadline to do before tomorrow, I may have to sleep after 1 a.m., I probably won’t need to turn to coffee but I may need a Coke.

The thing that Professor Paine and Dwarkesh showed us last night was that anyone with the interest, attention and time can begin to understand complex history, and that a consistent framework of structuring arguments is an effective way to research and present something. Returning to my own personal meta-framework, I’ve established reasonable frameworks to study components - supporting evidence, if you will - sleep, caffeine, hobbies, routines - but have not achieved an optimal consolidated state of all components and interactions, which is a constant and dynamic process.

Closing notes

I am the last person to buy into a bio-hacking, life optimization, productivity boosting lifestyle. The Paine meta-framework is life-changing to me precisely because it’s oriented around an ever-changing set of arguments and conclusions that are not strict directives. The sheer fact that we SF tech workers filled an auditorium last night to listen to a Naval War College professor talk about the Cold War and sat in rapt attention is enough proof that there is incredible demand for structure and organized learning, that we need to double down on active education for the common person after high school and college, and that frameworks can be an effective way to communicate and encourage that learning to people that think like me. For me, the meta-framework is optimized for a life satisfaction metric which is inherently dynamic, unpredictable, and thus benefits from an often uncomfortable but necessary introspection and reflection framework I’m constantly assessing and formulating.

I have to leave it at that tonight so I can get back to the day job I’ve been procrastinating. Thank you for reading and let’s keep the conversation going.

Big players continue to compete at scale

This week Anthropic dropped fourth-generation Claude models and Google I/O almost exclusively focused on AI launches including a swing at Perplexity with AI Mode, which (from a brief side-by-by test I just conducted by asking both Google AI Mode and Perplexity to find me dinner in the Castro tonight) can provide more accurate and personalized results due to integration with things like Google Maps and your personal Google data. Shop with AI Mode, providing an “agentic checkout” experience and virtual try-on via Search Labs, has the potential to become a direct competitor to Amazon, although it will take time and a significant investment in quality to convince customers at scale.

These are crowning developments in the last year of the incorporation of AI and AI-adjacent technology into every and any vertical. The AI space is becoming exponentially more saturated, competitive, and bubble-y (see: OpenAI’s $3 billion acquisition of Windsurf) and yet multiple international conglomerates continue to funnel investment into expanding feature sets in a race to establish monopolistic control over the most valuable commodity globally: human attention. Similar milestones in the past where we saw this form of large-scale technological shift influencing human behavior and driving global manufacturing and software development trends were the development of televisions, personal computers, smartphones, and social media. The difference now is how much easier it is to spin up hardware and develop software to serve these customer applications. Iteration loops and talent are now orders of magnitude faster and more available with the proliferation of hardware and software infrastructure and developer tools, accelerated applications-oriented training and remote work.

Software engineering as a discipline and art

While we can all agree coding assistants and generative app developers aren’t currently capable of replacing experienced human developers entirely, the way we are studying and viewing software engineering as a career path is still changing significantly. Corporate developers are now far more focused on learning to apply AI tools to their work. Earlier stages of software development feature heavy use of assistants that can generate in a matter of days what previously might have taken weeks or months. Educators now have real concerns around student use of AI and assistants hindering their ability to actually understand concepts from first principles, which in my opinion is a valid concern, but also just another flavor of how technologies like Google Search and Wikipedia changed and democratized the ways we learn and process information from last-century era phone calls to information lines. The power still lies with the people to use this information in ways they wish.

To those that argue that the software engineering class will shrink significantly over the next century, that’s absolutely true in the traditional, 2000s sense of what being a “software engineer” means. Modern mobile and web app development, site reliability, devops, infrastructure; these can and will be increasingly abstracted away in the same way AWS and cloud service providers abstracted away actual data center and hardware maintenance serving software applications. At the same time, this creates a unique economic and social incentive for high-quality ideas and implementation to become noticeably better than competition. This opens up opportunity for engineers to become an increasingly creative class, where the purpose and skill required to build reliable, beautiful, human-serving software is amplified and even demanded by a continued and universal human desire for connection and simplicity.

The flip side

I’ve written before on the need to ramp up investment in sustainable hardware, environmental protection, and safety and security measures. These continue to be my personal areas to watch. Naturally and understandably, successful venture and investment funds tend to lean toward seeking out and therefore encouraging customer-facing product development to create and capture greater market share of the human attention economy. A pessimistic and capitalistic view of health and longevity investment might be that longer lifespans lead to more literal economic time. It is widely known that the food and pharmaceutical industries has twisted research incentives and public health outcomes, and critics draw parallels to social media and internet companies funding machine learning research.

My view is that it is impossible to avoid some level of “incentive impurity”, as any action taken will have some downstream effect that isn’t ideal. And, in fact, given the way our global economy has been built for centuries, perhaps we shouldn’t be trying to minimize our downstream negative impact but rather maximizing the positive influence of our actions and technology.

Here I want to break from the tech talk for a bit and reference two pieces of art that influenced me recently. (Some spoilers ahead.)

The first is John Wilson’s How to Cook the Perfect Risotto, which in classic John Wilson fashion documented his quest to surprise his landlord with her favorite dish and became an honest reflection on addiction, the inherent randomness and unfairness of much of human life, and thus the immeasurable importance of human connection, even the simplest of actions we take to show our love for each other. In one scene, John goes on a long drive to take himself out of an environment that reminds him of smoking, but is stuck behind a truck with a broken exhaust; upon questioning, the driver reveals that he is “addicted to diesel” and refuses to consider the environmental impact of his actions, citing (paraphrased) that bad chemicals and processes are impossible to avoid at this point, including the plastic that packages our potato chips. John comments to the viewer that it’s interesting how little regard some have for the impact of their actions on others and then continues his risotto quest. My takeaway from this in the context of this tech blog is that there will always be makers and competitors in any space with less regard for human and social outcomes (aware or unaware) that will continue to exist and even be rewarded for their actions, and our best course of action may be the peaceful stance of counteracting in ways we can. We can still learn to make risotto for our landlords in a world where diesel truckheads drive freely.

The second is an excerpt from Garth Greenwell’s Small Rain, in which a man teetering at the edge of death during the COVID pandemic reflects on the potato chips his sister (not knowing about his effort to reduce his sodium intake) has brought him at the hospital. I’ve termed this method of acknowledging and even appreciating the beautiful outcomes of a complex set of economic incentives the Potato Chip Model:

…Chips were exactly what I shouldn’t eat. But that would mean telling G she had brought the wrong things, I imagined her feeling hurt, and so I reached into the bag, taking just one chip. It was thin and almost weightless, an amazing object, really, if you think about it, a miracle of engineering… nothing in nature was like this, so perfectly tuned to our pleasure… I knew it was pernicious, one of the manipulations of capitalism… But maybe a Snickers bar [referencing another delicious product of capitalism] is a wonderful thing… as a product of science and experiment, and as the end point of a whole system of production and distribution, the ingredients sourced I’m sure from all over the world, which can only be abstract to me… whatever machines had been designed and built… systems of distribution to carry [and package these impossibly fragile things] all over the country, the world, so that you can walk into a store in the middle of America, in a college town in Iowa, and for a few dollars fire up those points in your brain that mean pleasure… it was at once amazing, proof of ingenuity and genius, and also the product of unimaginable suffering, of exploitation and violence and labor… and there was truth in that, too, the intrication of wonder and depravity, pleasure and violence.

I’ll leave it at that today. Thank you.

Very brief history of recommender systems

Modern recommender systems are “designed to provide personalized recommendations to users of online products and services, enhancing the user’s online experience” (Li et al., 2021). But recommender systems have a rich (pun intended) history that began decades before widespread internet use. One of the lesser-sung heroes of computer science is Dr. Elaine Rich, who pioneered one of the earliest recommender systems, Grundy, at CMU in 1979.

Dr. Rich writes:

Grundy exploited relatively deep models, both of the books it knew and of its users. This meant that the models of the individual books had to be created by hand. And what about the models of each user? The goal, in designing Grundy, was to minimize the amount of interaction a user would have to have with Grundy in order to get going. To achieve that goal, Grundy used stereotypes, which could be triggered for a particular person just from a small set of words that the person provided as a simple self description. Grundy generated book recommendations by comparing its model of the current user to the models of the books it knew about. It chose the best match and generated a short description of the book, which emphasized the reasons it thought this user would like the book. Then it asked the user whether it liked the recommendation, and, if not, why not. Using that feedback, it updated both its stereotypes and its model of the current user.

Dr. Rich compares this approach, which still required “relatively deep [hand-created] models, both of the books it knew and of its users”, to collaborative filtering, which predicts a user’s preferences based on the behavior of users similar to them.

Referencing the 2023 IEEE Recent Developments in Recommender Systems survey paper, a good example of a hybrid recommender system (at a high level, combining elements of collaborative filtering and e.g. content-based, which is conceptually similar to Dr. Rich’s original approach oriented around learning from user behaviors to form a user model that can be used to personalize recommendations) is Netflix’s recommendation system. The following is just a snapshot of the input data Netflix considers, just the tip of the iceberg, but already we see how complex these systems can become and what bottlenecks we may face in designing these to be scalable, memory- and time-efficient:

We estimate the likelihood that you will enjoy a particular title in our catalog based on a number of factors including:

• your interactions with our service (such as your viewing history and how you rated other titles),

• other members with similar tastes and preferences on our service, and

• information about the titles, such as their genre, categories, actors, release year, etc.

In addition to knowing what you have watched on Netflix, to best personalize the recommendations we also consider factors including:

• the time of day you're enjoying Netflix,

• the languages you prefer

• the devices you are enjoying Netflix on, and

• how long you enjoyed a Netflix title.

Yanir Seroussi highlights two primary motivations to work on recommender systems:

1. Money

2. True data science problem: intersection of software engineering, machine learning, and statistics.

This blog post is not going to be a comprehensive overview of recommender systems. Instead I want to propose and focus this post on a third primary motivator to working on recommender systems, a topic that has finally been getting the airtime it so desperately needs in the general artificial intelligence space, but still remains underrepresented and understudied in the heavily commercialized field of recommender systems:

3. alignment.

Traditional alignment models don’t economically serve large-scale social recommender systems

Alignment is rapidly growing as a global coordinated research effort to ensure humans are able to retain control over AI systems and “align” them with human values (survey paper, 2023). In 2021 researchers in SF and Berkeley published this short paper “What are you optimizing for? Aligning Recommender Systems with Human Values” proposing that “recommender system design can draw from the study of the value alignment problem: the problem of ensuring that an AI system’s behavior aligns with the values of the principal on whose behalf it acts”. Applications of explicitly values-aligned recommenders included identifying and down-boosting or removing clickbait, harmful content, and misinformation, increasing fairness and reducing bias in algorithmic discovery to boost e.g. smaller artists on streaming platforms, and more proactively encouraging positive social media outcomes over toxic ones.

These are in theory big improvements over first-generation recommender systems, but practically money still talks much bigger game; social media remains one of the roughest, hardest landscapes to “align” because a few huge players are actively shaping the way humans process each others’ online content. These networks and feedback loops specifically engineered to increase profit via human engagement have the power to bring people together, democratize education and information (including this very post), and change lives for the better in ways previously unheard of, but at the same time introduce pressures and dangers completely unique to these platforms.

Alfie Lamerton of Formation Research is one of the first I’ve seen to start actually formalizing alignment risk in recommender systems, something Lamerton is generally calling “lock-in risk”. (Somebody get Lamerton into a PhD program ASAP.) Lamerton highlights from literature that recommender systems are “believed to contribute to polarization by promoting extreme content”, but that “the evidence for this is inconclusive and it’s possible that this effect is not very pronounced… either because it was never the problem in the first place, or because companies are effectively mitigating the effect in their in-house recommender system”. However, “The way recommender systems curate information can lead to the isolation of information into a filter bubble”, and Lamerton argues that these are bad outcomes of these phenomena that measurements of lock-in risk could help counteract.

How to flip the problem

Lamerton is off to a rolling start with funding for Formation Research and a huge existing book of work on fairness and generalized bad actor detection in recommender systems (including the many in-house systems mentioned with controls and protocols built in and in enough cases open sourced), and it’s disappointing but not shocking that there isn’t the same level of attention on aligning systems that can influence and ultimately change human behavior as there is on aligning the machines and models we are building. After all, in a universal human attention-based media industry where corporations have so much of a say in what societies define as beautiful, desirable, creative, and aspirational, it simply doesn’t make economic sense to focus on driving forward some nebulous definition of “what makes us good as people” while we’re also basically building potentially too-smart killer robots we need to protect ourselves against by continuously checking and re-calibrating them (traditional alignment).

Perhaps it’s more convenient to dismiss these claims as fearmongering sensationalism; perhaps this post itself is an example of a polarized outcome of a filter bubble where I’ve been fed post after post and paper after paper covering alignment risk and how to build responsible and safe AI. But I’m hopeful that consumer trends are beginning to support a larger movement here. There is a higher appetite for personalized mental health solutions and “healthier social media”. One huge benefit of existing media platforms are that diverse lifestyles and differing reward systems can be highlighted; a good example was the “Danish way-of-life-turned-phenomenon” hygge a few years ago. There are a number of reasonably high-profile influencers that are proactively and publicly disconnecting to reconnect with themselves and their loved ones, including one of my personal favorites, plant-based chef Derek Sarno.

The easiest way to contribute to fixing the alignment problem with recommender systems is to provide human feedback that rewards the online behaviors you do want to see more of, even aspirationally. Personally, if I discover organically or via algorithm some small musician or band I like, I’ll put all of their tracks on repeat even if some of them are a little too experimental for my taste, I’ll subscribe and like their videos and watch all the way through even if the camera work looks like my own (amateurish). These are small steps but at population scale they are the core input signals to the vast majority of recommender systems.

We can get there without sacrificing progress on AI alignment in robotics and defense and fighting deepfakes; in fact, in the same way governments and organizations have the power to encourage healthier nutrition and fitness habits, it’s time to take a wider and stronger stance on healthier internet consumption habits. To me, approaching this movement as an alignment in recommender systems problem democratizes it because we are the humans these systems cater to; we actually have utmost power to mold them to better serve our purposes.

And you can now take one second to contribute to recommender systems alignment yourself by subscribing to me here and to Alfie Lamerton’s Substack as well and boosting my blog engagement metrics and increasing healthy, informed debate on the internet by leaving your unfiltered comments below.

“What’s this software?” I asked him, noticing how Pearl visually highlighted sections of the x-ray and made them clickable. “Oh, it points out possible problem spots, like this tiny one,” he said, pointing at what would have looked like a smudge on the screen without the software’s overlay. “It helps me remember to monitor these over time. This one looks like it’s not bigger than it was last year, so I’m not too worried about it for you.”

“Nice,” I nodded. “Yeah,” he said, “AI is everywhere now.”

I’ve seen a noticeable uptick in AI bubble speculation in conversations with friends and coworkers recently. Dave Friedman just published some useful notes on the current state of the AI industry as compared to dotcom and telecom:

• If today’s AI boom follows historical tech bubbles, many AI startups will go bankrupt, valuations will crash, and only a few dominant players will remain. The hyperscalers will use their balance sheets and distribution channels to consolidate much of the AI market.

• However, AI differs from dotcom/telecom in that it already has commercial utility, is bottlenecked by supply constraints, and is primarily backed by tech giants rather than startups and venture capitalists.

Last month Anthropic launched the Anthropic Economic Index (not to be confused with the OpenAI Economic Blueprint) and an initial report presenting “a novel framework for measuring AI usage patterns across the economy”. Their analysis mapped 4+ million Claude conversations to tasks by US Department of Labor defined occupations and generated a few interesting plots I want to highlight.

[Anthropic open sourced their data and Jupyter notebook supporting this initial report and invited public feedback, so I went in and ran the analyses myself on my local machine to see how easy it would to be to do this. It’s incredibly easy and I like the locally generated color scheme better so I will be using plots from my own local run.]

Disproportionately high usage of Claude in computer and mathematical sciences, entertainment and media, life and social science, and education relative to US worker population in those areas; disproportionately low usage in transportation, sales, administrative, and food service. Caveat: “the occupational classification of a conversation does not necessarily mean the user was a professional in that field.” Even so, this is consistent with my own recent empirical evidence that software engineers, teachers, media marketers, and students are increasingly turning to AI assistants to answer questions and help them write, while specialized and physical skills like transportation, construction, and food service either have less need for these types of outputs and/or do not find Claude as effective to assist tasks.

Software ranks high on the wage-Claude density spectrum, but OB/GYNs (by task classification) are the highest paid, least frequent Claude users. Of course my initial reaction to this was “What types of prompts would be classified as OB/GYN tasks?” but the data is produced by “privacy-preserving platform” Clio (Claude insights and observations) so I could only find a meta-level view of occupational tasks in the data, which is good for building customer trust (your personal interactions with Claude will not be released in an open-source dataset) but insufficient to answer questions like this, and other questions I have such as demographic (e.g. age, education level, geographical location) patterns in Claude usage.

I adjusted the plotting thresholds to provide a little more color (sacrificing cleanliness of the graph). Waitstaff and orthodontists join shampooers and OB/GYNs at the salary extremes for low frequency Claude usage by task; technical writers and more software-related occupations join the ranks of higher usage; notably, tutors emerge as a higher usage occupation on the lower end of the wage spectrum.

For the final point to highlight from this paper, I will use the original graphic from the paper. “AI use leans more toward augmentation (57%), where AI collaborates with and enhances human capabilities, compared to automation (43%), where AI directly performs tasks.” In other words, users still ask Claude more “help me do X” and “optimize Y” type tasks (augmentation) over “do Z” (automation).

Of course, the authors themselves are most aware of the limitations of their analysis (task classification vs. occupation, limited occupational dataset, limited context into user workflows) and future directions (task subsets within occupations, broader impact of the work). I find their final words consistent with my own thoughts around these general trends:

While these patterns are informative, they capture just the beginning of AI’s integration into work. As AI systems expand beyond text to handle video, speech, and physical actions through robotics, and as AI agents become more capable of carrying out extended tasks autonomously, the nature of human-AI collaboration is poised to transform dramatically. New tasks and even entirely new occupations may emerge around these capabilities. Empirical frameworks that track these changes dynamically will be crucial for anticipating and preparing for the evolving landscape of work. The challenge ahead lies not just in measuring these changes, but in using our understanding of them to help shape a better future.

While I actually found the dataset and analysis a little underwhelming, I realize that isn’t the point of this paper. The point is to establish some methods of thinking through these patterns and to get researchers and economists talking, which Anthropic is well-suited to do.

Pull request: Anthropic Green / Anthropic Environmental Index

This is going to be a less corporate-friendly take on something that would be valuable to study and measure over time: an environmental impact index to save on the environmental cost of prompts to Claude and other technologies.

An openly AI skeptic friend sent me this CJR study ripping apart generative search tools in news citation capabilities which was doubly hilarious to me because I have no shortage of Bad Experiences with Gemini.

She also raised a very valid point: even setting aside the inaccuracies we see often in generative search, what of the environmental impact?

We already know AI research is energy-intensive but also that AI has the potential (if used responsibly) to streamline and recharacterize much of the global economy. This is the core dichotomy / trade-off I’d like to see much more investment in understanding, especially if it can be deployed to frontline users to guide their behaviors and interactions with the assistive techology itself. I am imagining a combined AI/ESG Index that goes far beyond the current measures which might be basically market cap weighted averages of AI companies. I’d like to see something akin to Anthropic’s recently released token-saving updates (optimizing throughput and reducing token usage with Claude 3.7 Sonnet) e.g. explicitly balancing environmental impact of queries in console, IDE, or at the search application layer.

Just as Uber allows a Green option to select a certain vehicle class or wait / share a ride for marginally lower environment impact at the price of time and cost, I could imagine Anthropic Green (or Cursor, VSCode, or any other tool serving the highest-use occupations from this study) providing a similar service. Understandably, the best of intentions will struggle to scale in the face of economic and investor pressures — we do need to make sure Anthropic and other big players continue to make good money to see the biggest possible lift to influence customer behavior here — and I’m under no pretense that ESG investing and corporate environmental commitments are of completely pure and good intentions (naturally, it may also be driven by regulation, marketing, and involve questionable actual environmental impact analysis) but I also can believe that something like this at global scale could still encourage and enforce better, more informed user patterns.

I can’t believe I’m writing this into a blog post, but here goes — I am just a member of the software occupational class, standing in front of the behemoths of the AI industry, asking them to consider popularizing environmental impact analysis frameworks on top of the economic analyses they’ve already published.

This blog post is meant to provide a high-level overview of what agents are, how to assess them, how they are being monetized and priced, and open questions I have.

Last logistical note before we get started: I’m shifting to a monthly blog post cadence (and a private, unpublished weekly reading and writing habit) moving forward. I’m hoping this is a better balance for me, but will reassess periodically.

What are agents?

Talent agents, especially if you believe what you see on Entourage and BoJack Horseman, make calls, book auditions, and might do some form of brand management for their clients. Real estate agents (Selling Sunset, anyone?) can identify potential homes, schedule tours, and negotiate on behalf of their clients.

In the same vein, AI agents are autonomous actors set to achieve some objective in an initially unfamiliar world that they can be trained to understand and act more intelligently in over time [1]. If this sounds familiar, it’s probably because these are the same core principles that inform robotics and reinforcement learning (RL). (As much as things would be so much easier if I could walk around and say “it’s just RL!” to everything, there is in fact much more to agentic frameworks and their societal implications than just RL.)

As a friend put aptly over dinner recently, “my agent will be talking to your agent” is no longer the stuff of science fiction or A-list Hollywood parties. And offloading the mundane or otherwise uninteresting is not a new human desire. Before Anthropic’s computer use and OpenAI’s Operator could take over repetitive tasks in our browsers on our behalf (“even creating memes”) we had Google Assistant calling restaurants and making reservations for us (although not without growing pains). Just think of how Silicon Valley street style might have evolved differently if Stitch Fix had managed to deploy AI agents acting as personal stylists at scale instead of leaving an unsatisfied if niche fast-fashion demographic defaulting to the proverbial capsule wardrobes built around black turtlenecks and clear frames (no judgement, I’m one of them).

AI agents today are deployed models that can be trained on and accomplish tasks in specific domains, such as AiSDR for automated sales reps, Decagon for customer service agents, and highly programmable actors such as Cloudflare Agents that can browse, persist state, send emails and texts, and even call APIs.

Agentic quality and performance

What separates a good agent from a bad one? How can we universally improve the performance of agents when there is such a wide range of tolerated outcomes? If you’ve ever watched a parent or member of the pre-Internet age try to interact with a chatbot (or worse, an automated scammer), you might understand that what appears to you as an objectively terrible and low-effort simulation of human behavior may be completely convincing to someone else. Can we even call these things agents?

Here’s where the classification and application becomes more of an art than a science. Erik Schluntz and Barry Zhang at Anthropic make a few relevant points in “Building effective agents” [2]. First, they define both workflows (predefined aka rule-based orchestration of LLMs and tools) and agents (self-aware and -controlled aka autonomous LLMs leveraging tools) as agentic systems. And:

When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.

When more complexity is warranted, workflows offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale. For many applications, however, optimizing single LLM calls with retrieval and in-context examples is usually enough.

Intuitively, it makes sense that for most of our day-to-day tasks we are looking to get off our plates, we may not need a full-fledged agent. But for applications where we want to actively learn and compound our own base knowledge limited by literal time, AI agents are at a stage of development that could pave the way for a global socioeconomic revolution. This vision is so powerful that it’s dangerous. Daniel Kokotajlo predicts an “intelligence explosion” that will need to be actively monitored and controlled [3]:

when zillions of copies of the AI are autonomously conducting AI R&D across several datacenters due to having surpassed human abilities -- i.e. when the intelligence explosion is underway -- the AIs will plausibly scheme against their human creators so as to achieve their actual goals instead of the goals they were supposed to have… Whether they do this depends on the extent to which the Spec/initial-role emphasized stuff like honesty and not doing this sort of thing, and on the extent to which the agency training distorted and subverted it.

This knowledge compounding is already happening across fields. Researchers at Google Research and DeepMind just published “Towards an AI co-scientist”, demonstrating a “virtual scientific collaborator to help scientists generate novel hypotheses and research proposals, and to accelerate the clock speed of scientific and biomedical discoveries” [4]. Recall last year’s discussion on innovation at the intersection of traditionally orthogonal fields; the AI co-scientist was similarly “motivated by unmet needs in the modern scientific discovery process… including the ability to synthesize across complex subjects”.

Where’s the money?

It’s tempting to see the independent, time-saving parallel workstreams promised by well-trained agents and demonstrated compounding effect of collaborative agents and draw a conclusion of exponential growth. Perhaps much of Wall Street would even agree. So should we go all in?

Recall Schluntz and Zhang: “Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense” [2]. We already know costs in this industry are notoriously opaque, and I’ve previously done back-of-the-envelope calculations placing SOTA LLM training costs at ~$10M (for strict budgets), but what about specifically agentic development which you could argue is more about the creative infrastructural application of already mostly-trained (with perhaps an active learning component) models in tandem with assistive tooling?

I decided to go directly to a trusted source with firsthand experience and asked Perplexity a few questions, which I should have known would give me “mid” answers as evidenced by the first image result under “how much did OpenAI operator cost to build”:

Before you start commenting things like “Perplexity is not an agent!” and “graphic design is definitely not your passion”, let me point your attention to the screenshot on bottom left: OpenAI’s Operator being offered as part of the $200/month ChatGPT Pro subscription plan. Naturally, one of the clearest ways to monetize agents is to make people want them so much they pay for them. Offering what is effectively a public beta (research preview that will “evolve based on user feedback”) as part of an all-inclusive Pro plan may even make Operator a consistent loss leader for OpenAI, just as predictably-priced hot dogs and rotisserie chickens bring loyal customers back to Costco time and time again. The promise of a customizable, self-correcting browser operator that can send emails and shop online for you, if properly implemented and kept as a free add-on, can be a big ticket to ChatGPT Pro becoming OpenAI’s business-solidifying Amazon Prime equivalent, which offers a set of services such as 2-day shipping, Audible, Prime Video, and grocery delivery bundled into one package.

A strategy like this also builds consumer trust, as long as OpenAI holds prices constant (at the very least under, let’s say, the US CPI index for eggs). Admittedly Anthropic has cultivated a stronger direct trust ethos in this space with earlier and more central investment in AI safety, and broadening the research charter of computer use for “general computer skills” beyond the browser (e.g. spreadsheets, data entry between views). Claude scores 22% on OSWorld, a multimodal agent benchmark that evaluates humans at 70+%, but the previous SOTA models were at 7-8%, so at least we’re seeing marked improvement.

What would a world of fully-integrated AI agents look like? We’d be able to deploy personal assistant agents to find and book appointments with dentists, therapists, barbers, gyms and trainers. Restaurant agents well-trained to our tastes could monitor soft openings and grab highly-coveted tables for us. Concert agents could stay up late for us to grab pre-sale tickets to our favorite bands’ upcoming local shows. Travel agents could track flight and hotel prices for us and propose a tailored itinerary based on our inputs and preferences. The make or break for these solutions will be how reliable and trustworthy they are relative to their human counterparts.

Corporate valuations

Last year Microsoft mostly acqui-hired Inflection for $650M [5], Amazon mostly acqui-hired Adept paying around $25M, hiring in staff at the startup with formerly $1B valuation, and paying back $414M to investors including Greylock and General Catalyst [6], and HP just acqui-hired most of Humane (shutting down the AI Pin); The Verge deputy editor Alex Heath argues that Big Tech is one of the only reliable sources of cash to build expensive “leading AI models”, largely stemmed only by antitrust law [7].

OpenAI was valued at $157B last October and is now in talks to secure funding that would put it at $300B+, driven by investments from Microsoft and SoftBank. Anthropic is closing in on a $60B valuation, with $8B in direct investment from Amazon (now estimated at $14B). xAI is at around a $44B valuation (the same amount Elon bought Twitter for in 2022).

Traditional methods to assess company valuations start with subtracting liabilities from assets, but “this simple method doesn’t always provide the full picture of a company’s value”. Clearly investors aren’t going this route when assessing the value of companies like OpenAI, projecting $5B losses on $3.7B revenue last year. Fundamentally, valuations for companies like this are collective bets placed on a company’s potential to grow and exceed at a rate that will make up for the current state, but they are, after all, just bets. Foundation Capital GP Ashu Garg puts this very well [8]:

the economics of AI don’t work like traditional software. OpenAI is currently valued at 13.5x forward revenue—similar to what Facebook commanded at its IPO. But while Facebook’s costs decreased as it scaled, OpenAI’s costs are growing in lockstep with its revenue, and sometimes faster…

At the height of the dot-com bubble, the entire internet economy generated $1.5T in revenue (adjusted to 2024 dollars). Today, generative AI companies produce less than $10B in revenue while planning infrastructure investments that could exceed $1T.

In tech, the winners aren’t always those with the most advanced technology—they’re often those who build the most compelling ecosystems… This creates brutal economics for OpenAI. While they reportedly plan to raise ChatGPT’s subscription price to $44/month over the next five years, Meta can give away their AI for free.

New technologies, no matter how revolutionary, don’t automatically translate into sustainable businesses. OpenAI’s $157B valuation suggests we might be forgetting this lesson.

This isn’t to diminish what OpenAI has achieved. They’ve shown us that AI can do things many thought impossible just a few years ago. They’ve forced enterprises to rethink how they operate and changed how humans interact with computers. But starting a revolution isn’t the same as profiting from it. Today’s headline-grabbing AI companies are creating tremendous value, but that doesn’t guarantee they’ll be the ones to capture it in the long run.

But is there a possible truth in these valuations? In a hypothetical future where OpenAI and Anthropic become the next Microsoft and Amazon, both currently multi-trillion dollar companies, it’s hard to imagine any reality other than both players carving out multiple applications serving the global population, probably applications that can co-exist and drive business to each other. OpenAI and Anthropic are currently head-to-head in an exponentially higher-stakes race to “AGI” (artificial general intelligence, or Dario Amodei’s preferred term “powerful AI”). It will legitimately take shooting for the stars to land at an outcome where the hundreds of billions of dollars to be invested in research and infrastructure yield trillions of dollars in generated value, and would likely require significant advances in robotics (androids, defense, aerospace), biomedical research (preventing and curing disease, longevity), and enterprise productivity (sales, automation, development).

Open questions

Immediate questions following above discussion of company fundamentals and product offerings:

• Which agent will win? Does there have to be one clear winner? For example, Google still dominates general search today, historically beating out Yahoo, Bing, and Ask Jeeves. Perplexity is gaining popularity and is something to watch if it can expand its use cases and provide value beyond traditional search. Of the general public consumer agents provided via OpenAI’s Operator, Anthropic’s computer use, and sector players developing similar technologies for enterprise (Adept, Codeium, Cursor), will we see similar market behaviors unfold?

• What about Ashu Garg’s “compelling ecosystem” observation? Meta and DeepSeek continue to reinforce the open source LLM standard (Meta with a significantly more established customer ecosystem), Apple has been “quietly buying AI startups”, and OpenAI and Anthropic are developing loyal customer bases. Will partnership be possible to consolidate customer bases with the products built for them, what regulatory hoops will companies need to jump through, and how will our lives and the way we interact with technology change as a result?

Longer-term questions:

• At what point will “general computer agents” like OpenAI’s and Anthropic’s shift labor away from employment opportunities such as call centers and data entry? Training costs are still a thing, but computer agents don’t need the physical material and manufacturing cost of, say, warehouse robotics. I was reminded, while hearing about a friend’s recent trek around the city to find a working printer for a shipping label to make a return, that mobile, online-first experiences have antiquated entire subsectors of manufacturing (RIP Sears). Restaurant QR codes for menus and ordering are now a norm instead of a COVID-necessitated precaution. Will “computer use” agents similarly shift us away from the repetitive tasks that currently define much of our interaction with our phones and PCs? What will we be able to do with all that extra time?

• Will certain employment sectors see a boost from these jobs being increasingly automated away? Will the barrier to entry and eventual ROI become more compelling for artists, musicians, chefs, and in-person experience curators? Are our economic systems prepared to handle this shift? What can we do governmentally and infrastructurally to get ahead of a decreased need for entry-level IT professionals, which is currently a common re-skill for many to enter the middle class? Personally, I am doubling my investment in education and funding accessible forms of therapy and connection such as music and art initiatives for the at-risk, but this will require large-scale societal awareness.

Highlighted reading:

1. Google’s “Agents” whitepaper

2. Anthropic’s “Building effective agents”

3. Daniel Kokotajlo on “agency training”

4. Google’s Gemini 2.0-based “multi-agent AI co-scientist”

5. Microsoft’s “unusual deal” with Inflection

6. US FTC pays extra attention to Amazon’s acquisition of Adept

7. “Big Tech’s playbook for swallowing the AI industry”

8. “Why OpenAI’s $157B valuation misreads AI’s future”

Acknowledgements: Many thanks to Steve for proofreading drafts of this post, providing thoughtful and useful feedback, and for writing inspiration. Go subscribe!

My deepest thanks to my readers new and returning. To my personal friends that continue to read and support my blog, thank you for everything.

I re-read Neel Nanda’s weekly review process and realized, tactically, I’ve spent about three hours each Sunday since I restarted this blog furiously collecting papers, news, and thoughts from that past week and throwing them all together into a hodgepodge post, usually racing against the clock before dinner plans. While time-effective and technically true to the Sunday ritual, this doesn’t exactly create the environment I need to produce work I can be proud of; thus my “right lesson to take from [last week’s blog post]” was to start drafting my blog posts earlier in the week (in smaller, more frequent chunks) so my Sundays are more relaxed and provide space for higher-quality writing. So I wrote the first part of this post on Saturday — progress, if I may say so myself ;)

In my last post I mentioned “heightened collective social desire and nostalgia for organic experiences and technologies in the way we were accustomed to before the current wave of online-first assistants, chatbots, large-scale automation” — there’s more to this I didn’t have the time to write out. Here it is!

hybrid consumption

Well-curated record shops and dive bars still exist as reliable methods to discover local bands and new music, but the next generation’s nostalgia will be for Spotify, YouTube, and (dare I say it) TikTok algorithmic discovery and online communities. This was happening long before COVID as the internet was one of the only ways for people to get and stay connected across the world. I’ve never been able (and don’t intend) to fully grasp how many realities are represented on the internet, and as a consumer take a hybrid approach myself to curating the online content I am served. I don’t use Instagram or TikTok; I do use Twitter, YouTube, and Substack. The users I engage with on Twitter are primarily people in tech in their 20s-40s in urban centers; I find my musical tastes underrepresented on my Twitter feed but shockingly well-understood by YouTube’s recommendation algorithm. I keep thinking about this:

Lazyrecords is in alpha (adding friends should work though) and already offers a vinyl-themed discovery platform, social media feed, and playlist (self-curation) functionality. Technology is fundamentally cyclical; I’ve written about this before; I know “this time feels different” is such a cliché but something genuinely feels different about these niche-to-mainstream shifts in taste and content consumption. Doji and Obsidian are also products I’ve discovered through Twitter, both marketed at people that want to up-level and outsource their personal style and productivity to a machine, and both developing cult followings. Apple has managed to stay in style because it is the style (although the critical reception of Apple Intelligence, Genmoji & Events might signal the start of a perception shift that will take Apple in the direction of Microsoft — still functional but less “in with the trends”).

I find the global shift towards remote work fascinating because it correlates, at least anecdotally, with increased demand for self-reflection (e.g. productivity, tracking) and in-person community-building in the extra hours per day no longer spent on commutes or sitting in offices. For example, Toronto-based internetVin is building Futureland (“software to track progress on any aspect of your life”) and New Systems (“an experimental project aimed at improving Toronto’s systems related to creativity, innovation, and culture”) including New Stadium, a “new home for art, technology, startups, and urban innovation… in the West End of Toronto”. Within my demographic of “people in tech in their 20s-40s”, I see a lot of “lifestyle optimization”; go for a quick try-on and fitting at a Warby Parker physical store and then have prescription glasses shipped directly to your doorstep; place an UberEats order while you’re still in your Uber such that it arrives when you do; pay for meal prep services that decide what you’re eating so you don’t have to.

If we formulate this as an optimization problem with a few variables and constraints, lifestyle optimization is usually not as simple as “just save time”; there are axes for how effective in-person vs. online experiences are and unquantifiable effects like how they can make you feel (e.g. Trader Joe’s is becoming a stressful experience in over-crowded city centers, bodegas and delis are an accessible third place for many people, and every clothes & furniture retailer is releasing software to help you “virtually try-on / see” without going into a store).

I visualize the consumer landscape today as relatively niche circles of products and services initially catering to certain demographics that can expand by appealing to universal shared human needs such as community, nutrition, and productivity. Apple, Microsoft and Google are beginning to compete more seriously in the “one provider for all services” space, Meta falls behind in hardware but has an advantage in social, Amazon continues to provide grocery, pharmacy and fashion. A single megaconglomerate buying and beating out the competition feels a little too dystopian for me and global regulations will not make this an easy path, but stranger things have happened in the last few years. I continue to protest this by supporting small cafes, corner stores, and record shops, and friends keep telling me that megacorps and the consumer behavior popularized by Walmart and Costco are uniquely American and capitalistic, but [edit: it’s Sunday and I haven’t finished this thought — will have to circle back next week ;) ]

generative media

Gemini tells me what to write in my emails before I’ve even started typing, designers have Figma AI and Photoshop AI, programmers have a number of well-established coding assistants to choose from, chatbots / agents are being developed for literally every vertical possible. These are all becoming normalized and incorporated into standard workflows in these industries.

Music and art continue to be relative outliers in automation & assistance. Music is definitely being generated and pulling in profit but the music-making industry has always had a higher barrier to entry, and there are simply far fewer professional musicians than professional designers and programmers. Music creation tools like GarageBand do offer AIs for music generation and noise reduction and DeepMind quietly released a suite of generative AI tools for music creation last year.

Tero Parviainen has a good presentation covering the history of generative music, including Brian Eno’s “self-generating musical systems” in the 1990s that were programmed to change with each listen.

There is mainstream music and art and then there are basement DJs and local artists in every city; even though making music and art is on average not nearly as lucrative as programming and design, I think there’s enough of a collected “cottage industry” across musicians and artists (not to mention the globalized online influencer and localized education and hobby industry) to warrant greater investment in tools for these creators. The challenge will not be whether there will be enough demand to use and pay for these tools, but actually in educating potential users on how to use the tools and what they can gain from them.

In fact, the value of these tools may lie explicitly outside of the direct creative process for the media industry. I see a lot of value in improving agents for marketing (e.g. automated suggestions to help smaller artists boost their online presence), logistics of making good art (e.g. sourcing and purchasing equipment), and components that many artists may see as a hindrance to their creative process or otherwise need to outsource (e.g. music mastering). Players that emerge in this space will likely need to find a way to cater to both professionals and hobbyists in the same way Ableton and Native Instruments do to stay business-viable; to tie out this point with the previous section, it will involve collaboration with influencers and taste-makers that hold power to shift societal perception and adoption of this next generation of instruments.

According to WIRED, DeepSeek started as a deep learning research arm of High-Flyer, a Chinese quant hedge fund established less than ten years ago:

focused on PhD students from China’s top universities… who were eager to prove themselves. […] The hiring strategy helped create a collaborative company culture where people were free to use ample computing resources to pursue unorthodox research projects […] For many Chinese AI companies, developing open source models is the only way to play catch-up with their Western counterparts, because it attracts more users and contributors, which in turn help the models grow.

I am quite literally rubbing my hands as I get ready to read this paper. Let’s get into it.

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, DeepSeek-AI (2025)

DeepSeek has open sourced “DeepSeek-R1-Zero, DeepSeek-R1, and six dense models (1.5B, 7B, 8B, 14B, 32B, 70B) distilled from DeepSeek-R1 based on Qwen and Llama”. (Qwen is developed by Alibaba Cloud. Llama may be more familiar to Americans — Meta’s open-source AI models.)

In their paper, DeepSeek publishes results benchmarked against OpenAI o1 models, which was first introduced as a “new LLM trained with RL” — “chain of thought” reasoning with the idea that having the model “think for longer” and learn mechanically how to structure and reason could generate better results — DeepSeek-R1 outperforms OpenAI o1 models on AIME (competitive math) and SWE-bench (GitHub issue resolution ability), performs on par in Codeforces (competitive coding), MMLU (covers 57 subjects from elementary to advanced professional level), GPQA (graduate level sciences) and MATH-500 (subset of OpenAI’s own math benchmark).

DeepSeek’s approach is unique because they set out explicitly to investigate the capability of LLMs to “develop reasoning capabilities without any supervised data, focusing on their self-evolution through a pure RL process”. DeepSeek-R1-Zero, the product of applying GRPO (Group Relative Policy Optimization, a DeepSeek RL algorithm originally developed to come after SFT) directly to DeepSeek-V3-Base model skipping supervised fine-tuning is able to match AIME performance of OpenAI-o1-0912 after “thousands of RL steps” but run into other problems like “poor readability and language mixing”, so incorporate “a small amount of cold-start data and a multi-stage training pipeline” for DeepSeek-R1.

— liner notes: DeepSeek-V3-671B beats Llama-3.1-405B model performance at a literal fraction of the training cost —

DeepSeek-V3-Base was released last month and estimated the total cost of training the 671B parameter model at $5.6 million (based on requiring “2.788M [NVIDIA] H800 GPU hours for its full training” — by the way, US regulation of AI chip exports is something to watch).

Model training cost is notoriously expensive and opaque in this industry, with Sam Altman cited as claiming training cost of OpenAI’s trillion-parameter GPT-4 at “more than $100 million” and Anthropic’s Dario Amodei projecting a $10 billion model by 2025.

Let’s do some back-of-the-envelope calculation for an apples to apples comparison to DeepSeek-V3’s (671B parameter) training costs; take Meta’s Llama 3.1 (405B parameters), which required 39.3M [NVIDIA H100] GPU hours. NVIDIA H800 GPUs are H100 GPUs formulated for Chinese export; “H800 mainly reduced the chip-to-chip data transfer rate to about half the rate of the flagship H100.” Rental costs seem similar between H100 and H800, at $2 per GPU hour DeepSeek assumes in their calculation so let’s just use that. That means Llama-3.1-405B cost $78.6M to train vs. DeepSeek-V3-671B at $5.6M, more than an order of magnitude difference, and DeepSeek-V3-671B was able to outperform Llama-3.1-405B in a slate of general and competitive reasoning benchmarks.

DeepSeek is able to achieve this unprecedently low training cost by designing a FP8 mixed precision training framework, an approach first published in 2023 out of Microsoft Azure and Research leveraging low-bit data formats to reduce LLM training cost.

— end liner notes —

Back to the DeepSeek-R1 paper. We’re still analyzing DeepSeek-R1-Zero, which is kept cheap by using GRPO for pure RL and forgoing supervised fine-tuning.

At this stage DeepSeek-R1-Zero is competitive with OpenAI o1 models across benchmarks and takes more time to respond as RL steps increase by the thousands, a marker of the model developing an “intelligence”.

I genuinely love it when deep appreciation for the craft comes through in any research paper, and want to highlight this section:

One of the most remarkable aspects of this self-evolution is the emergence of sophisticated behaviors as the test-time computation increases. Behaviors such as reflection—where the model revisits and reevaluates its previous steps—and the exploration of alternative approaches to problem-solving arise spontaneously. These behaviors are not explicitly programmed but instead emerge as a result of the model’s interaction with the reinforcement learning environment. This spontaneous development significantly enhances DeepSeek-R1-Zero’s reasoning capabilities, enabling it to tackle more challenging tasks with greater efficiency and accuracy.

Aha Moment of DeepSeek-R1-Zero A particularly intriguing phenomenon observed during the training of DeepSeek-R1-Zero is the occurrence of an “aha moment”. This moment… occurs in an intermediate version of the model. During this phase, DeepSeek-R1-Zero learns to allocate more thinking time to a problem by reevaluating its initial approach. This behavior is not only a testament to the model’s growing reasoning abilities but also a captivating example of how reinforcement learning can lead to unexpected and sophisticated outcomes.

This moment is not only an “aha moment” for the model but also for the researchers observing its behavior. It underscores the power and beauty of reinforcement learning: rather than explicitly teaching the model on how to solve a problem, we simply provide it with the right incentives, and it autonomously develops advanced problem-solving strategies. The “aha moment” serves as a powerful reminder of the potential of RL to unlock new levels of intelligence in artificial systems, paving the way for more autonomous and adaptive models in the future.

Of course, there’s a catch to DeepSeek-R1-Zero. You can’t expect a model effectively allowed to teach itself succeed in all tasks, even after thousands of iterations, and DeepSeek addresses the “poor readability and language mixing” alluded to earlier by “construct[ing] and collect[ing] a small amount of long CoT data to fine-tune the model as the initial RL actor” to produce DeepSeek-R1.

On the flip side, DeepSeek also produces distilled models that directly fine-tune Qwen and Llama (applying only SFT now without an RL stage) “using the 800k samples curated with DeepSeek-R1”, proposing it as a method to “equip more efficient smaller models with reasoning capabilities”. On the relative tradeoffs between SFT-forward (distillation) and RL-forward methods:

Distilling more powerful models into smaller ones yields excellent results, whereas smaller models relying on the large-scale RL mentioned in this paper require enormous computational power and may not even achieve the performance of distillation… While distillation strategies are both economical and effective, advancing beyond the boundaries of intelligence may still require more powerful base models and larger-scale reinforcement learning.

Finally, it’s nice to see an explicit callout on Unsuccessful Attempts in training, including a failure of an AlphaGo-inspired Monte Carlo Tree Search approach to scale in an exponentially growing token generation defined search space as opposed to a relatively well-defined search space for a board game.

This was my favorite paper to read in a while because the researchers’ genuine curiosity and excitement about the unflashy, conceptually simple reinforcement learning approach really came through. This paper was indeed “a powerful reminder of the potential of RL to unlock new levels of intelligence in artificial systems.”

Last night I bought a custom domain name for this blog (an investment, I promise), but it’ll take a few more days for the DNS records to fully propagate so this will still be published as a vanilla Substack post because I’m trying to honor my ritual and don’t want to miss a Sunday. Fingers crossed this blog is as web-optimized as possible by next Sunday!

Market Street; “The Path of Gold” after the 1906 earthquake and more recently and colloquially “The Street They Keep Power-Washing When There’s a Big Conference in Town”. Today, Market is probably the single most culturally-encapsulating street in SF. It starts at the Ferry Building and Embarcadero to the east — symbolically and literally the root of commerce in SF. It passes through a gorgeous historical architectural section — City Hall, the symphony, the ballet, the opera house — in stark and sobering contrast to the activities that happen there. Finally, it juts down through the Castro before looping off into the quieter edges of the city.

Welcome to the first edition of the Market Street Journal, established in SF to report on the Modern AI Gold Rush ;)

Energy shifts in SF and US policy

This week on Market we saw even more Patagonia jackets than usual coming from the JPM conference in town. Nightly rates for hotel rooms in the city broke records by a factor of ten. (Some last-minute travelers were forced to book hotels in Oakland, which might have sounded cool until they realized they had to shove themselves into overcrowded transit every morning to make it into SF on time for meetings.)

Daniel Lurie was recently inaugurated as SF’s new mayor and showed us what Chinatown could actually look and feel like by putting on a night market rivaling the foot traffic of Times Square and booking Bay Area native ZHU.

Perplexity perplexed many by bidding to merge with TikTok and TikTok… did TikTok. The TikTok ban drove interesting user behaviors across social media apps, including an influx of users to RedNote and wider recognition that ByteDance’s influence extends far beyond TikTok (CapCut).

President Biden signed an executive order to “accelerate large-scale AI infrastructure development at government sites”, the “AI energy bill”, in tandem with the Biden administration preparing to transition the $52 billion Chips and Science Act to the incoming administration. Unsurprisingly, TSMC beat Q4 earnings expectations citing “surging demand for advanced chips used in AI applications” as well as boosts to NVIDIA and Tesla. AI chips powering self-driving cars at scale is still a longer-horizon bet but futures traders should take note. Anecdotally, Waymo is picking up traction as a legitimate price/wait time competitor to Uber and Lyft in SF (beyond just being a tourist attraction).

Major moves in multimodal

Shanghai-based MiniMax (valued at $2.5B+) released three new models: MiniMax-Text-01, MiniMax-VL-01, and T2A-01-HD, for text, image/text, and audio. MiniMax claims 456B parameter MiniMax-Text-01 beats Gemini 2.0 Flash on MMLU / SimpleQA, that MiniMax-VL-01 can compete with Claude 3.5 Sonnet on multimodal e.g. ChartQA, and TechCrunch assesses T2A-01-HD to be comparable with Meta’s audio models.

London-based AI video generator Synthesia (valued at $2.1B) backed by NVIDIA raised $180M (doubling its valuation since 2023) in a round led by NEA. Runway is another player in the AI video space I’ve been following for a while. Generative multimodal technology, especially audio/visual, is similarly characterized as self-driving automotive technology as having reasonably functional iterations in production scaled to end customers, but with a long path ahead to achieve parity with human-made, human-driven experience. Arguably self-driving is a step ahead of AI audiovisual, but that may soon change with the lower cost and energy barrier to entry for multimodal model development as I mentioned in my last post.

The meme economy and crypto

$TRUMP (launched Friday on Solana) hit $13B market cap; Solana still trails by a factor of at least 5x behind Ethereum in TVL but as of today beats Ethereum by almost 10x in last 24h volume. GameStop, Taylor Swift, and now unironically Solana meme coin Fartcoin — money is as unserious now as it’s ever been, which should be the loudest signal for us to get real and start talking about stablecoin.

In addition to the natural use case of fiat-, precious metal, or other asset-backed cryptocurrency that theoretically solves the volatility problem making Bitcoin and Ethereum more amenable as trading instruments than actual reliably-valued currency to exchange for goods and services, Chainalysis now reports stablecoin has officially surpassed Bitcoin as the currency dominating the $50B+ cybercrime industry (malware, scams, stolen funds, darknet, and ~$10B in FTX creditor claims in 2022), taking the majority (63%) of all illicit transaction volume at 77% YoY growth.

Before we get too hot on stablecoin, let’s revisit its monetization mechanisms. When Stripe bought stablecoin infra startup Bridge for $1.1B last year, MIT Cryptoeconomics Lab founder Christian Catalini highlighted in Forbes:

Relying solely on reserve interest isn’t a sustainable way to monetize a stablecoin… Stablecoin issuers like Circle and Tether [N.B. the two largest stablecoins by market cap and both backed by USD, although Tether has faced skepticism] seem to overlook that today’s high-interest environment is an anomaly, and a sustainable business can’t be built on a foundation that’s likely to crumble when market conditions shift.

…what options do stablecoin issuers have? Unless they’re relying on temporary regulatory loopholes—which are unlikely to hold long-term (more on that in the next section)—they’ll need to start competing with their own customers.

…Stripe doesn’t face this dilemma. As one of the world’s most successful payments companies, they’ve mastered the art of deploying and monetizing a streamlined software layer on top of global money movement—a model that scales efficiently through network effects without being slowed down by the need for country-specific banking licenses. Stablecoins accelerate this approach by acting as a bridge between Stripe and domestic banking and payment rails. What was once a network constrained by legacy institutions—including card companies—can now overcome its last-mile problem, delivering significantly more value to merchants and consumers.

Editor’s desk

You didn’t really think you’d get away without a paper dump from me, did you?

Last week I put A Hitchhiker’s Guide to Scaling Law Estimation on my reading list. This 2024 paper from Leshem Choshen, Yang Zhang, and Jacob Andreas at MIT and IBM compiled and released “a large-scale dataset containing losses and downstream evaluations for 485 previously published pretrained models” including Llama and GPT, used these to “estimate more than 1000 scaling laws”, and “derive[d] a set of best practices for estimating scaling laws in new model families”. Recall model loss measures distance between a model’s predictions and the actual labels, hence we generally want to minimize loss when training models, but the core tradeoffs with computationally-intensive training methods such as deep learning are in cost and compute. This paper is important because it established the largest public dataset describing scaling behavior across model families and a generic mathematical framework to define a set of scaling laws estimating performances of new models f given model families F. Publishing the “secret sauce” here and opening it up to public contribution is an admirable and influential way to empower researchers and practitioners alike to move faster and cheaper in the otherwise exponentially more complex, compute- and cost-intensive space of deep model training.

Chirag Shah and Ryen White at UW and Microsoft submitted Agents Are Not Enough (Dec 2024) to ACM Proceedings, which felt like a breath of fresh air to read, especially in the current industry wave of every remotely e-commerce application releasing a AI customer service chatbot and marketing emails put together so haphazardly that they forget to replace “[Your Name]” with my actual name:

While generative AI is appealing, this technology alone is insufficient to make new generations of agents more successful. To make the current wave of agents effective and sustainable, we envision an ecosystem that includes not only agents but also Sims, which represent user preferences and behaviors, as well as Assistants, which directly interact with the user and coordinate the execution of user tasks with the help of the agents.

Shah and White describe a number of agent development waves starting in the 1950s to the 2000s, including multi-agent systems (still significant development here, including a recent benchmark for multimodal multi-agent systems in Minecraft and Altera AI also using Minecraft as a simulation space for large-scale agentic behavioral studies) and cognitive architectures like Soar from CMU. The primary concerns and obstacles to continued widespread societal AI agent adoption they raise are in scalability and robustness (predictable) and ethics and safety (predictable but still needs way more research attention as I’ve been saying for months). The notable difference this paper makes is in actually proposing an expanded paradigm for developing agentic AI, e.g. a world of Agents, Sims, and Assistants. No doubt, large-scale social media and customer applications that run on user preferences and monetize targeted ads formulated “Sims” first at such scale (e.g. using user profiles, preferences, settings in recommender systems) but what they haven’t yet popularized in the literature is this idea of “Assistants” that operate in that space between “Sims” and “Agents”.

An Assistant is a program that directly interacts with the user, has a deep understanding of that user, and has an ability to call Sims and Agents as needed to reactively or proactively accomplish tasks and sub-tasks for the user. In this regard, an Assistant is a private version of an agent that can have access to a user’s personal information and could be fine-tuned to that user, allowing it to act on the user’s behalf.

Just last night I was imagining a world where I could deploy AI versions of my online behaviors (modeled after my existing behaviors on Twitter and Substack, for example) that could perhaps not write for me (because I like writing so much and am not looking to automate that away) but interact with Tweets in a way true to my behavior, go out and find Substack posts and papers that I would find interesting, tailor my social media algorithms for me at scale and help me find inspiration in potentially 1/10 or 1/100th of my current time spent. Those would effectively be my Assistants in this Agent/Sim/Assistant model. Watch this space :)