Last week, EU regulators handed Uber a €825 million fine for using automated algorithms to deactivate driver accounts without meaningful human oversight. The ruling was about one key element: the regulators argue that consequential decisions about people, like firing someone or deactivating their driver accounts, cannot be made by software alone.

As we discuss how much better AI is getting at human tasks, it is probably also an important time to discuss what tasks AI should or should not be doing. A lot of companies are offloading tasks to AI and are rushing to deploy agents across customer support, HR, financial services, and more. The pitch is an easy one: tokens are a cheap alternative to full-time employees. It is easy to see why a company like Uber would hand over a deactivation decision pipeline to an agent plus a quick human thumbs up. The EU does seem to be on the right path when it argues that a symbolic, quick human glance doesn’t count as a genuine human approved decision.

You can think about Bill Gates what you want, but I think he is right in saying that the AI tech shift is different from previous ones because it substitutes for human cognition directly. Looking at the way the technology is shaping up, it is clear that the corner companies will want to cut is the one that is most expensive as well, humans. Generating code and having agents making decisions is cheap but delegating consequential decisions to AI seems like a loss for humanity.

While AI might be able to diagnose our cancer better than a human doctor soon or fire us quickly and objectively, we can all agree that we can’t lose our human touch on the things that impact human lives so deeply.

The Tidbit: An astrophysicist and science communicator was presenting his newest, AI vibe-coded work when his peers pointed out that it was totally wrong. This was quite embarrassing for him, so he wrote a long piece on AI in science titled “AI Made Me Look Foolish. Then I Learned”, another publication found that 56% of his text was written by AI.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

AI is helping companies and developers ship faster than ever, but it also seems like our software products haven’t really been improving much. According to Linear (Jira 2.0), pull requests are up 111% in the past two years. And AI agents author almost half of all the issue tickets, while the amount of time humans spend planning has stayed the same.

But, I think somewhere along the way we forgot to stop and ask ourselves if we really need so many AI pull requests and so many minor updates. Don’t get me wrong, I believe there is significant value in using AI to write better code, faster, but it does seem like many eng teams are drowning in AI generated code. The same Linear data also shows that actual deployments are down, review times have increased, and teams are spending more time triaging than doing helpful work.

We can count the number of pull requests and we can see it going up, but it is unclear if they are leading to product improvements and features that users are actually looking for. I think for a lot of savvy users, AI is able to help them ship good code at a reasonable pace, but speed isn’t everything and I fear that a lot of junk is being shipped.

At the core of it, AI adoption is clearly rapid and the next steps are to get an understanding of what is being shipped, why we’re still taking forever to plan, and figure out how to properly, and quickly, review AI code. It seems like the jury is still out on if the introduction of AI has improved our products, maybe it is just the humans that are in the way.

The Tidbit: It’s finally happening… A little late but social platforms are finally working on reducing the amount of AI slop polluting feeds. Spotify is bulk removing AI content, you can now report AI slop on LinkedIn, and YouTube is doing a detox. We can keep AI slop to our pull requests for now.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

Following last week’s discussion on how AI is changing the web, I wanted to zoom in on Cloudflare’s role in this as they’re quite focused on getting to this next stage of the AI agent dominated web. Cloudflare says that more than half of the internet’s traffic comes from AI agents, so it is clear we can’t be operating with a “business as usual” mentality.

It’s been interesting to see them roll out significant infrastructure updates to help AI agents navigate the internet. The standard way that agents interact online is token heavy and works by taking screenshots, parsing HTML, and guessing where buttons are. To fix this they’ve made changes in three areas.

First, every agent gets its own dedicated SQLite database so all these agents can do their thing, go to sleep, and then recall state at a much lower cost. Next, they’ve made improvements to identity frameworks so that each AI agent is linked back to its owner, so a site knows on whose behalf the agent is working. And then lastly, they’re adding a payment layer so that folks can charge agents per API request or data pull.

These three layers essentially create a second web optimized for Agentic AI that runs underneath the web that you and I interact with which is optimized for humans. It makes much more sense for AI agents to be working in a space optimized for them, and for us to browse an internet that is optimized for us.

The Tidbit: It has felt like Spotify wasn’t quite ready to deal with or admit the impact of AI generated music would have on their platform and their users. They’ve finally announced that they will prominently label AI artists and not recommend AI generated content.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

Last week a judge overturned Amazon’s injunction against Perplexity’s AI shopping bot. The key piece was that Perplexity’s browser accessed Amazon under direct user instruction so it was the end-user accessing Amazon, not Perplexity itself, giving Perplexity the legal win since Amazon could not prove that the company was accessing Amazon services without authorization.

While perhaps confusing, I think it brings up a good question around agents on the internet: how much is human intention and how much is it AI agents action on their own?

I think in the Perplexity vs Amazon case, Amazon is attempting to buy time to get consumers using their AI products and they do not want something like Perplexity, ChatGPT, or Gemini to be a consumer’s primary product shopping agent. But for sellers, what happens to e-commerce when traffic is no longer driven by traditional ad-based buying by human consumers?

Shopify is trying to make AI agents a complement to traditional search rather than a substitution. Agents are great because they can browse full product catalogs and pull out the right, niche products directly for the end user.

And then on the infrastructure side, we see Cloudflare launching traffic authentication mechanisms where reputable agents must cryptographically sign their requests. This allows site owners to distinguish between malicious scrapers and legitimate purchasing agents acting on behalf of a user.

I don’t think we’ve seen the end of the lawsuits but it is interesting to watch the industry adapt to a new form of search and consumption. Can’t wait to see how we build ads targeted at AI agents!

The Tidbit: In very related news, you can now shop Safeway directly from ChatGPT with their new plugin. Now you don’t have to go all the way to DoorDash to order your groceries… but it is interesting to see another alternative to traditional ads with suggested things to order directly in your AI chat app.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

Quarterly earnings have come out and to nobody’s surprise, the big tech players are still spending big on AI. Reactions were a bit different this time, some companies got a stock boost and some got a bump down, so let’s see why the different giants are getting different treatment.

I think at a high level, the market now is more AI knowledgeable and isn’t lumping all of AI into one category and it understands the unit economics behind AI a bit better. It is evident that AI demand is higher than ever before so the question is whether the massive spending is translating into visible near-term revenue and margin expansion.

The two big winners this quarter were Microsoft and Amazon who saw growth in revenue, so investors are thinking that their large investments into datacenters are paying off in clear revenue growth. And these two companies are saying they don’t have enough compute available to keep up with the demand they are seeing, so the spending on datacenters seems justified for them.

On the other hand you have Google, Meta, and Oracle who all got hit negatively. Google saw its free cash flow turn negative for the first time ever, Meta is spending massively on AI for its own products but isn’t yet seeing a return on that investment (earnings dropped 14%), and then Oracle can’t afford to fund its datacenter expansions with cash and is taking on so much debt.

So basically when big tech says they’re planning to spend $730B this year on AI, we’re now scrutinizing the return on these investments a bit more closely and expecting more than “AI is the future”.

The Tidbit: Google Earth released and within 24 hours removed a new feature allowing users to create AI-generated satellite images. The weak guardrails on the feature allowed folks to easily generate satellite images of things like a nuclear plant in Iran and the AI watermarks were easily circumvented. I think my main question is why would you create this feature, what is this for?

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

I’m sure most of us have come across a human influencer trying to sell us something online, promising it will cure all our ailments. Turns out we’ve probably also come across an AI generated influencer selling us something that is truly too good to be true.

The New York Times found how pervasive and industrialized these online scams have become with the AI tools now broadly available. Health scams have been around for forever, so why do these work so much more effectively? AI makes it easy.

The barrier to entry is virtually nonexistent. Anyone can spin up hyper-localized campaigns without needing actors, studios, video editors, or even fluent English speakers. And then on top of that iteration is key and also basically free. These organizations can easily spin up hundreds of variations until they find the exact combination that converts their audiences. They can test different ages, demographics, accents, scripts quickly and see what gets picked up.

The AI video generation tools that we now have access to are honestly so good that it can be hard to tell what is real and what is fake. If you’re looking for clues you might be able to tell, but the general audience doesn’t know they should be wary. It is so easy to effortlessly create fake doctors, traditional medicine gurus, and experts complete with realistic (but fake) wall diplomas and clinical offices that look convincing on camera.

And then to top it all off, unlike television ads, there are no laws regulating what can be posted and claimed on these social media platforms, and no regulation of synthetic human avatars in advertising.

The tools are here and folks are as easy to scam as ever, I won’t be surprised if a lot of our real ads also become AI generated and hyper-personalized.

The Tidbit: And now onto a real medical miracle… Researchers used AlphaFold to find and engineer a novel enzyme capable of reversing aging damage in human tissue. By using AI structure prediction, the team redesigned how the enzyme fits onto damaged proteins to undo molecular aging previously thought to be permanent.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

We keep waiting for an LLM to be small enough and competent enough to run locally on an iPhone or similar device.

Running a competent AI model with advanced reasoning capabilities that are good enough to get something out of requires a massive amount of compute that is simply not available on smartphones. You have to shrink the model down so much it becomes unusable both via intelligence but as via the strain it takes on your device to try to power it.

As we all know, running AI on an edge device comes with some benefits like covering privacy concerns, reducing dependency on cloud computing, and reducing the current latency that is a result of sending queries off device. There are fortunately many tasks that don’t need massive amounts of intelligence behind them, where speed is more important like email drafting or summarizing texts or documents. But none of this matters if the models aren’t competent enough to work.

Two pieces that could help with this puzzle were released last week: PrismML released Bonsai 27B, an open source highly compressed model that shrinks down in footprint to 3.9GB that still retains 90% of its intelligence. For those like me who didn’t know, that fits within recent iPhones’ memory limit. Combine this with an open-source training patch called Antidoom from Liquid AI, which helps fix a common tiny model problem of doomlooping (stuck in an infinite loop of reasoning).

Bonsai serves a compressed model that fits inside a phone’s memory budget and Antidoom helps drop the loop failure rates on small models to reduce processor burden, allowing us to actually successfully run on phones. As AI becomes more engrained in our everyday lives I think we’ll see more efforts towards on-device success.

Subscribe now

The Tidbit: Amazon’s Mechanical Turk is officially coming to an end. Mturk was AI before we had proper AI and it is interesting to see it come to an end while they never managed to grow the business into something like Scale.ai or Mercor.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

→ subscribe button here

Meta can’t seem to get their AI products right. Last week they launched a new image generator integrated into their various product surfaces, and instead of it being a big catch-up moment there was so much backlash.

Muse Image itself is an impressive image generation model and is unique in that it operates like an agent, utilizing tools to search or code for more accurate outputs and has a built-in “self-refinement” process to check and edit mistakes. They built an impressive image and accompanying video model, and it does signal that after a lot of time and money, Meta is trying to break into being a relevant AI model player.

But it doesn’t matter how good your image model is if you bungle the roll out. Three days after the release they rolled back a feature utilizing the new model. They had allowed users in Instagram to generate images based on other people’s public photos, which means a lot of people had their likenesses used in AI content without their consent. It turns out their invisible AI watermarking feature also didn’t really work…

Roll outs of new technologies are always a little rough but it seems like Meta didn’t learn from Google, OpenAI, or Grok’s image chatbot mistakes. I think we see it everyday but launching AI tools isn’t always as easy as just plugging it into existing software or consumer apps. Image generation continues to be where AI progress most clearly shows how fast it’s moving and also where the potential for misuse and harm is most evident. This pattern suggests the race to ship is still winning out over the discipline to learn, and users are the ones left dealing with the fallout.

The Tidbit: Google made what I think is an appropriately positive move with respect to their ads. Users can now see if the ad they come across has been generated using AI. It isn’t perfect but I think it is important to build the transparency muscle as more and more of the content we come across is AI generated.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

In the midst of the many layoffs blamed on AI, there is an interesting profile on SAP that claims that with the introduction of AI, employees can reinvent jobs rather than SAP laying them off. The SAP story is a little funny because two years ago they eliminated nearly 10,000 jobs due to AI and are now acknowledging that AI is doing much of a software engineer’s job, but apparently the company is now taking a different approach.

Germany, and Europe in general, do have stronger workers’ rights than in the US, but I think it is fair to say that Europe and the US have some fundamentally different approaches to AI adoption and its regulation. A report from March showed that the U.S. has established a substantial lead over Europe in both the adoption rate and intensity of AI use. SAP seems to be on the leading edge but while the report found that 43% of US workers have adopted AI, only an average of 32% across Europe have.

A recent economic report focused on the EU by OpenAI found that in Europe, 47% of jobs will not be immediately impacted by AI while 14% have relatively higher near-term automation potential. Companies need to start planning for what this looks like for them, like SAP, so they can adapt or pivot so we don’t have to read more headlines about AI layoffs. It is also interesting to think about the role of AI in nations that have a shrinking workforce, which is expected across Europe, where the workforce will need to both adapt to AI changes but also to the retirement of skilled workers.

The Tidbit: As part of Midjourney’s legal disputes with three Hollywood studios, Midjourney is asking the studios to disclose how they use AI themselves. While I’m sure they have a legal reason behind, if approved, I’m more interested to hear the answer as I think AI usage is likely much more pervasive in Hollywood studios than we realize.

Share

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

Recently a paper on benchmarks was released by researchers from UC Berkeley, arguing that while recent AI systems achieve strong results on standard metrics, these gains rarely translate into meaningful economic deployment. They suggest this gap is largely an evaluation problem, as widely used benchmarks fail to measure sustained performance on real, economically valuable workflows. They also propose a new benchmark, Agent’s Last Exam (ALE), which is built to assess agents on complex, extended workflows; however, today I’d rather focus on exploring why evaluating these types of agentic tasks is so difficult in the first place.

When benchmarking conversational LLMs, you grade against completion and the answer key, but with agents you need to grade that the agent had the right behavior. An agent needs to make the right output while going through a chain of steps.

The paper argues that there are three elements in tension when looking at a benchmark: realism, verifiability, and breadth, and when use cases get complicated (and more like the real world) it becomes hard to verify and reproduce the outcomes. It is clearly non-trivial to properly evaluate agents. On code it is easier because there is a correct output, but when we move to non-coding use cases it isn’t so simple. The more realistic agentic workflows become the harder it is to verify them and as AI solutions become more agentic, we’ll continue to need better evaluation methods.

The Tidbit: Last week I mentioned how the pre-GPU supply chain has been getting an AI boost. Now we’re seeing it play out as Apple increases their prices, blaming it on the increasing costs of storage and memory due to AI snatching up all the components.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

Nature recently published a poll of 1,900 scientists (researchers) about their thoughts and usage of AI tools. The results match what I’ve also been hearing in real life. At a high level, scientists are experiencing a strong sense of peer pressure (FOMO) to adopt AI tools in their work. Basically, either you adapt to the new age or you get left behind by those who are willing to utilize the AI tools. Nearly 60% of surveyed researchers feel they like they will be left behind if they do not incorporate AI into their research.

I am biased but I do think that AI tools have the potential to positively impact research, especially as models get smart and better at utilizing scientific data. Researchers should resist the temptation to deploy AI across every possible use case, all at once. Instead, focusing on adoption of mature applications that we know work, like accelerating complex data analysis or streamlining document creation, researchers can get immediate benefits.

I also believe that in academia there is a fear of AI usage because what have all those years of specialized training been for if an AI can just step in and do a portion of your work? And how quickly until it can do all your work? In my daily work I have seen how quickly AI has been able to do parts of my job, so just like all other professions, researchers need to use these tools to get ahead and then adapt so they can stay relevant in the age of AI.

The Tidbit: We talk a lot about Nvidia’s recent exponential growth due to the AI boom, but there is a whole boom happening earlier in the supply chain. Fascinatingly enough these chip companies are not based in China, leaving the big AI player out of the chip supply chain.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

It took a while but Microsoft has finally joined the LLM model party, releasing seven new models. This shouldn’t come as a surprise, it was only a matter of time before they used their own resources to get in the game so let’s take a look at what they shipped. The models all seem fine, are on the smaller side, and none are competing at the frontier.

For differentiation, they’ve launched “Frontier Tuning” focused on providing customization for enterprise customers. This is a bit of a different direction than traditional fine-tuning as we know it, where this is run inside a reinforcement learning (RL) environment. At a high level the models, inside this environment, learn from real workflows, tool usage, and evaluation signals from the customer. Then, in theory, these models should be tuned not only on general, static data but also on the downstream outcomes that are desired by the customer. If we treat workflow completion as a reward signal, then the model is able to learn the optimal sequence of API calls, reasoning steps, and tool calls needed to complete a specific task.

I think this is an interesting approach as a stop-gap solution while Microsoft tries to scale up building models. Though many times we do see that bigger models still outperform specialized models, maybe the frontier tuning approach borrows some of the best of both worlds.

The Tidbit: More than 34,000 Instagram accounts were attacked by hackers who found a bug in Meta’s AI tools. Basically, anyone could reset the password for an Instagram account by asking the company’s AI chatbot to reset the password. Customer service chatbots shouldn’t be quite that helpful.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

As I wrote last week, AI does democratize access to things that are currently outside of a person’s skill set or budget, and this includes legal advice. There is a lot of swirl at the moment about the number of AI assisted lawsuits being filed by individuals increasing but a piece that isn’t getting as much attention is the implications of using AI as a bonus legal advisor.

TL;DR: Side legal advice from an LLM isn’t quite as private as you think, and it isn’t because your data is being trained on.

We have all heard of attorney-client-privilege (ACP), where discussions between you and your lawyer cannot be used in a court. This extends to emails, and it’s fine if you use a word processor or add files to a system. But when you begin a legal discussion with a consumer chatbot, it becomes discoverable in court. The chatbot is considered a 3rd party, so you have broken ACP, and this is not because the data is being trained on but because consumer chatbots store your data for a short period of time after your conversation. If you used an enterprise version with zero data retention (ZDR), well then you’re back in safe ACP territory.

So basically if a person engaged in some legal dispute uses the consumer version of a chatbot to get additional advice or confirm something their lawyers are working on, that whole chat would be discoverable and could be used against that person. However, if they had used the enterprise version with ZDR, the court then thinks this is a confidential interaction that is not discoverable.

So next time you have a little legal trouble, double check what version of your chatbot you’re using before admitting to what you did to it. My potentially controversial opinion is that I think it is a little interesting that if you interact with someone with a law degree who passed the bar you have confidentiality privilege, but if you use a smart (consumer) AI to help write your brief or give you a second opinion, you are out of luck.

The Tidbit: Some farmers are at the forefront of AI adoption. While you might not think of a farmer as an ideal first adopter of AI, the agricultural industry is quick to pick up money and labor saving technology to stay afloat with things like weed lasers and automatic cow milkers.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

I would say one of the best things about the incredible capabilities of AI these days is its ability to open up the door to so much that previously had a barrier to entry. The most obvious so far has been the democratization of coding. With the tools available today, anyone, even those with no coding experience at all, can spin up a new website or app, all they need is an idea.

Now, two examples that are not in the same realm but represent where we are headed in my opinion: First, this New York Times author successfully sold his house without using a realtor, relying only on Gemini to get him through the process. Realtors have long held onto the story that their commission was worth it because their niche knowledge would be critical to a successful house sale. But as the author proves, AI is able to fill the knowledge gap he has and working together, he is able to succeed without needing to hire an expert and pay the 3% commission. You’ve essentially given everyone with access to the leading AI services the opportunity to be their own realtor, and looking forward, will force realtors to provide actual value to remain relevant. I think we call this disruption.

Secondly, in a very different sector, one that has already been hit hard by AI, you see the creative and film industry being transformed by this technology. Runway recently shared the progress their AI video generation has made (it is a lot of progress!). Their example clips were produced on insanely short timelines and a team size of one, and honestly do not look like they were produced by AI. We’re entering a societal “uncanny valley” where anyone with a story can produce an accompanying professionally looking video without any video producing skills, team, and budget. Something that was unimaginable just a couple of years ago.

There is a tricky balance here, where in some cases the democratization will help me save on buying a house in the future, but also these are industries where folks will have to pivot and figure out where the human touch remains invaluable.

The Tidbit: Every time I think crypto is dead… it is not. My younger brother brought up a new cryptocurrency mining approach where you use the same GPUs for AI inference and mining at the same time. Basically, you sneak in mining math into the computer’s “free time” while it is processing AI so you can run AI models and mine digital currency concurrently without slowing down.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

There is an interesting new trend of bringing AI edited photos of yourself to your plastic surgeon to show what work you’d like to have done and what you’d like the end result to look like. In theory, this makes a lot of sense, why wouldn’t you want to show your preferred outcome on a version of yourself. But in practice, the AI world doesn’t magically translate to the real, physical world and a lot of plastic surgeons are having challenges explaining why physically unattainable AI features are not a real option.

Together with AI, people are able to dream up looks that are simply not feasible in the real world. AI still needs to be calibrated by human beings to function well. I think this is also why a lot of AI-based robotics work remains challenging as AI often does not interact well with the physical world.

When you’re crossing into the biological world hand in hand with AI, these doctors become the symbol of guardrails and reality. Waymo just suspended service in six cities, after two of its cars stopped on flooded roads, clearly unsafe. But this is just another example of how AI’s understanding of the world isn’t quite ready and doesn’t quite understand the physics and implications of a flooded roadway. That said, I believe we’ll see rapid improvements as AI becomes smarter, driven by increasing demand for physical AI products.

The Tidbit: Amidst the chaos of many tech jobs being cut “due to AI”, California’s governor has signed an executive order that is supposed to look into labor policies to deal with job displacement from AI. While not wrong, it does seem a bit handwavey and light on actual action while we’re seeing the job cuts unfold.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

The data center scramble is real. Demand for GPUs continues to go up but the infrastructure to power those GPUs is not as easy as just plugging into the power grid. We’re continuing to see issues with these “AI hyperscale facilities” that require an immense amount of land, water, and electricity to run. Especially in the US, the aging utilities grid cannot reliably support the rapidly growing demand.

None of this is particularly new news, but I do expect it to become more and more talked about as AI demand and consumption increases. One area that I think plays into this is the growth of agentic AI. We’re seeing a shift away from chatbots (conversations) to longer running or always-on agentic AI, where agents are completing multi-step tasks. Instead of a real-time back-and-forth with a human, we’re able to set our agents to work on something for hours to days at a time.

Agents also require more than just GPU inference. They need data storage (RAM) as they’re generating and retrieving large amounts of data in real time. And when agents are running (CPU) over multiple hours they need orchestration, memory management, and tool usage, which are all contributing to the crunch.

Not only do we see the crunch at the data center level, we’re also seeing shortages and backorders of Mac Minis as folks have turned to them to build and run AI agents. I don’t think there is a short term supply solution, it’ll be interesting to see where we end up being most crunched as AI becomes more and more agentic. But one thing seems clear: CPUs are officially hot again.

The Tidbit: Cerebras launched onto the public markets last week with a monster IPO valuing the chip making company at $67 billion. This is the first big AI IPO this year so it will be interesting to see the market reaction and if the price of their shares continues to go up.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

We’re entering a weird, new paradigm where AI is at work, at school, and at home, but we don’t quite have the social constructs in place to guide usage or even how we actually feel about it being everywhere. A recent Gallup poll found that 50% of American workers use AI at least a few times a year at work, and 13% use it daily. At the same time, 18% of workers think it is “very” or “somewhat” likely that their role will be eliminated by AI in the next five years.

Just in the past week, Cloudflare announced it was reducing its workforce by 20% (1,100 people) due to AI and Meta is planning to lay off 10% (8,000 people) to invest more in AI. These seem to be pretty regular headlines these days, and I think we’ve moved past AI being an excuse and now is maybe actually the cause of the layoffs.

I think there is a lot of panic because we can’t predict the future and we’re seeing AI quickly become increasingly competent. It is an exceptionally confusing and brutal time to be a college student as well. This is an interesting essay from a professor whose students feel the pressure to use AI to improve the quality of their work, but without thinking about how they should instead focus on improving their own (human) writing skills to become better writers.

The end result is that no AI is allowed in the class. But, these are the same students that need to graduate into a workforce that is now relying on AI and if they’re behind on their use of and the ability to get the most out of the technology, they’ll be less competitive.

There is a tricky balance here that will be key to get right as the world evolves, and we need to set up more clear social guidelines on how AI should be used and how use of AI needs to be made clear. Do I need to have a more clear label that this newsletter continues to be thought about and typed without AI?

The Tidbit: If you thought ping pong was safe from AI… you’d be wrong about that. An AI-powered robot has beaten elite table tennis players, which is a challenging game for robots since it requires fast reaction time and solid perception of a quick moving object.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

A recent Harvard study has shown that LLMs can outperform doctors in triaging and diagnosing emergency room patients based on the information available upon arrival to the hospital. Initial ER admission is challenging because doctors must make split-second, life-altering decisions using sparse, “messy” data from electronic health records and the paramedics (if you have watched The Pitt you know what I am talking about). The study specifically isolated text-based clinical reasoning to see how well the AI could process raw medical data compared to a human’s interpretation of that same data.

The study tested OpenAI’s o1 model and human physicians’ abilities to identify correct diagnoses at three specific stages of care: triage, evaluation, and admission. It found that the model generally outperformed human clinicians and existing diagnostic tools, and the more information available to the AI (and humans) quickly pushed the accuracy up. Initial patient triaging is critical and decisions are often reliant on low-information and incomplete data situations, where this type of AI can quickly process the information to hand off to a doctor to manage IRL patient care.

Systems integrating AI can act as a sophisticated second opinion that evaluates a patient’s medical history alongside the doctor’s initial assessment. Implementing LLMs more strategically in triaging can potentially help support overstretched emergency rooms. This isn’t replacing doctors, but can be a key partner in improving patient outcomes.

The Tidbit: Earning reports came out last week and AI spending is higher than ever. Across Amazon, Microsoft, Google, and Meta the expected spending this year is $700 billion, an insane amount of money which is largely going into data centers. This spending was contributing 1.5 percentage points to the 2% GDP growth rate in Q1 2026.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

I’ve written extensively about sovereign AI previously, the idea that nations should be able to host and serve AI on their own terms to ensure that their data is secure from foreign companies or governments.

There was news last week of Cohere merging with Aleph Alpha as a sovereign AI alternative, which has further solidified my thoughts that we’ll see growth in this area in the coming months. There is a lot going on in the world of AI and the model space is increasingly tough to compete in if you are not one of the big players. Most companies need to find a niche where they can build a moat and I think we’ll see a lot of the smaller research labs move that direction.

It is not unreasonable to assume that governments would welcome this development, especially in the privacy focused EU. For example, if a school system wanted to use a model hosted in their own datacenter, the requirements for their model could be relatively simple. As long as the model outperforms humans at a specific task, organizations don’t need the most advanced options available; they can rely on smaller, niche, or open-source models. The same goes for a tax system, where it would be reasonable for a government to want to ensure that tax data does not leave the country.

A merger like Cohere and Aleph Alpha’s’ allows these companies to compete in this probably lucrative niche market, whereas they would struggle to compete against the larger model developers in the broader market. It’ll be interesting to watch this play out and see if they’re able to get governments to set up the infrastructure and processes needed to implement basic AI.

The Tidbit: Google, Meta, Microsoft and Amazon will all release their quarterly earnings this week. It will be interesting to see what the numbers show after news that Microsoft is offering voluntary buyouts for 7% of the company and Meta announced it would be cutting its workforce by 10%. It is crazy to see the persistent layoffs being blamed on AI when it is likely before AI has an actual impact on job functions.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now

AI has gotten so much better so quickly that sometimes I get surprised when I see headlines like “AI hallucinations cost lawyers $110,000 in Oregon vineyard lawsuit”. The proliferation of AI into everyone’s lives has happened fast but it doesn’t mean that all those people who have picked up the tools realize they aren’t perfect yet.

In the case of this vineyard lawsuit, a woman was suing her brothers over the control of their late mother’s winery and tasting room seeking $12 million dollars. The fine was imposed by the judge after they discovered that the woman’s lawyers had filings that were citing nonexistent cases. The court said it found 15 references to made-up cases and eight fabricated citations, clearly AI generated.

Two things can be true at the same time: 1) AI has gotten much better and is a much more reliable tool, and 2) it still makes mistakes and hallucinates regularly. I actually think that the increased quality of outputs makes us trust the AI more and makes us less diligent in checking for any mistakes. You can say that better AI makes hallucinations more dangerous, not less.

I find that using AI to double check itself is a fruitful choice, basically using a new chat or a different model to verify that the work outputs are true and make sense. AI these days is quite good, sometimes better than humans, at figuring out if something is fake and we should use that to our advantage. And of course you could just check the sources yourself.

We also have to recognize that in the case of the vineyard, at the end of the day the woman did use AI but also two different lawyers signed off on the work. The problem is that experts see output that matches their mental model and wave it through. Remember to always check the work your AI does!

The Tidbit: The pivot from Allbirds (shoe company) to NewBirdAI (AI company) sounds like something from the tv show Silicon Valley. After selling its assets recently, the company is going all in on high-performance compute infrastructure. Allbirds continues to be a peak SF tech bubble company.

Know someone who might enjoy this newsletter? Share it with them and help spread the word!

Thanks for reading! Have a great week! 😁

Subscribe now