We’re seeing the end of subsidized AI plans converging accross the major frontier providers.

Anthropic with the stealth reduction.

OpenAI with the reintroduced 5h limit

Cursor’s auto-mode changing

And let’s not forget Copilot moving to usage based billing

It was expected this show couldn’t go on forever. And while we predicted this explosion before, the subsidiary went on as long as it did, because of (Chinese) competition. Now, with upcoming IPOs and all, closed frontier is forced to show their hand and try to squeeze out a last buck from their user base.

However this plays out, do not stay brand loyal or get comfortable. Make your agent files, skills and routines agent-agnostic, so that whoever has the best price of the day is receiving your money.

Fable 5.1 is Live

Even if your sub is worth less, it’s not worthless (ding ding).

There’s always a new model on the horizon to focus on. Fable 5.1 is hot of the presses. Better X, better Y. Same safety guards (reportedly less flaggy) and the new fingerprinting is live now.

Oh, and Mythos 5.1 is out as well, to whomever it may concern.

Back to you OpenAI, Kimi, DeepSeek, Qwen etc..

https://www.anthropic.com/claude-fable-and-mythos-5-1

Rabbit Holes

• A guy’s Automatic Bird Identification System detected his neighbor’s fart

• The AI Slop license

• Restroom Archive

Release of the week

https://four.htmx.org/announcements/2026-08-28-htmx-4.0.0-is-released

HTMX, the improved HTML saw its 4.0 release, skipping no. 3 entirely. It’s also the first JavaScript library to ever release on Game Boy Color apparently.

As I was deciding on an opener for this week I just caught Apple’s M5 Mac Studio announcement. After cutting down the highest-RAM options on their existing Studio models a few months before, we now get a glimpse of what sold-out-until-2028 RAM configurations look like in the new era.

The average Joe seems destined for always running a cloud-subscription along the side to have anything capable. With recent neutral model router takeovers and the looming threat of an open model ban by the US administration, you can see who the real winner is here.

Box Fight

Despite the price hikes I don’t expect Apple customers to vote with their wallets.

Apple and NVIDIA (with the DGX Spark) are getting competition from China.

China already owns the local model space and wants to own your coding harness.

Xiaomi and Alibaba now come for the box you’re running the model on.

Don’t expect a BYD moment just yet, because China is suffering from the same constraints as their US counterparts.

I never thought we’d reach peak Silicon Valley in real life, but here we are.

Watch out for our video on boxes later this week.

Rabbit Holes

• MS Paint watermarks your images

• How much of Hacker News is AI?

• Quantizing Qwen 3.8 down to potato-reasoning

• The only Apple product release that went down in price

Release of the week

https://go.dev/doc/go1.27

Go saw its much anticipated 1.27.0 release last week and I recommend you to check the community-made interactive tour of this release and its features.

Every time we see a breakthrough, a model leapfrogging another one, a significant disruption from open technology, it gets taken away or made shittier the week after.

We’ve seen the Claude Fable debacle and most recently our DeepSeek moment video hook got destroyed by the decision to change the model costs. DeepSeek V4-Pro output went from $0.87 per million tokens to $3.96 at peak and $1.98 off-peak. V4-Flash output went from $0.28 to $1.32 and $0.66.

Gemini 3.7 and Grok 4.6 happily took over the momentum.

DeepSeek also gave away the harness that was supposedly responsible for the powerful benchmarks of the latest models. It would’ve fit nicely into our OpenCode vs Codex and Claude Code showdown. Early findings are positive, but nobody is really blown away by a harness without batteries, as everything is plugin-based.

Watermarks Everywhere

The EU is making us feel safer one Act at a time. Claude decided to comply by simple watermarking its output for the whole world. They unveiled their method and we did a video on it. There’s no third-party tooling yet. Rest assured, it won’t affect coding as much as you think and a simple copy-paste between Claude and a Chinese model nullifies the watermark. It will be as effective and thoroughly enforced as EU’s cookie consent banner.

Speaking of watermarks and enshittification, your favorite musicgen AI Suno also announced they’re adopting a new audio watermarking tech, so that platforms like Spotify can better battle fraud (read: ban/demonetize AI music from any non-major artist). This is on top of its announced track download limitation, where you can only download a handful of audio files each month, even on a paid plan.

Suno is doing very little to brace for its own open-source moment. The recently released Minimax Music 3.0 is an impressive step up for an open music model and seems like the way forward if you’re a developer needing background audio.

Rabbit Holes

• Deep dive into the Quake Shareware CD

• A museum of background colors of classic OS

• A 1971 social network that’s still around

Up Next

The Dead Internet Theory is real.

For every single website visit, Claude sends 70,000 scrapes your way.

This week we’re telling you what to do about it. From the good practice of sanitizing your urls and taxonomy to the act of full-on poisoning AI training scrapers.

Tool of the week

https://huggingface.co/Qwen/Qwen3.8-27B

While it was a busy week for new model releases (new day, new Chinese model), Qwen 3.8 was the one that mattered most. Artificial Analysis puts it level with DeepSeek V4 and GPT-5.6 Luna Max. Every capability gain came from training and post-training. There’s talks about this new version reasoning more than normally, but it seems that’s exactly the thing that got it this impressive score.

Everything’s trending towards AI taking its training wheels off and seeing what happens.

• The popularity of Hermes and the new Gauntlet Loop are prime examples of powerful concepts that are a waste of tokens without a clear objective for them

• Auto-mode will become Claude Code’s default, because it can click ‘yes’ better than humans can.

• Docker Sandboxes are getting a second look as a safe container to run your AI agents in without them wreaking havoc on your local machine

Anything left unsupervised still compounds into a pile of junk on a long enough timeframe, so keep your AI on a short leash until there’s an actual breakthrough in the field.

Watermarked AI Output

Thanks to the EU, we’re getting watermarked AI output worldwide. This will make it easier for the AI police to determine which content online is generated or partly assisted by Claude, similar to how Suno fingerprints any audio file generated on their platform.

For text, Claude weaves ‘an imperceptible watermark’ directly into the text itself.

For any image filetype, it will attach signed provenance metadata.

Anthropic’s claim that this won’t hurt output quality is met with raised eyebrows by the dev community.

Let’s keep count of how many watermark removal SaaS platforms we’ll have by next month.

Third parties will be able to identify Claude-generated output just as well, so prepare yourself if you’re distributing to platforms that have a zero-AI policy.

Up next on the channel

This week we’re converging our recently published local LLM insights into a ‘definite’ advice for when you own a modest Macbook. Can you reliably code on an airplane by now and what’s the bare minimum in specs, model and tools to do so?

Rabbit Holes

• A guy bought noreply.net and companies started mailing him their secrets

• Stealing reasoning traces from proprietary LLM APIs

• Every possible Rubik’s cube state in one scrollbar

Tool of the week

https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model

Muse Glimmer landed this week: 30 billion parameters, distilled down from Meta’s bigger Muse Spark, and released on Hugging Face under Apache 2.0. Not the usual Llama community license with the acceptable-use appendix stapled to it.

Full precision it’s over 55 GB. Meta’s K-Quant-17GB build squeezes it under 20 GB with what they describe as minimal degradation on agentic tasks, which puts it inside a 24 GB card or a mid-spec Mac. Add DFlash speculative decoding for a claimed 1.5–3.1× on token generation and the pitch becomes clear: an agent that sits there all day, doing tool calls, without a per-token meter running.

Someone on Reddit noticed it was awfully quiet in Anthropic’s headquarters. A week without Claude updates is a concerningly long time. Are they preparing something big or simply enjoying a summer break after dropping Opus 5?

The Chinese labs are surely not seizing this moment for a breather themselves.

They’re coming out swinging to the point where no one can keep up anymore. We had Kimi K3 last week, now an updated DeepSeek Flash and forthcoming Pro, Qwen 3.8…

At this stage the moat must surely be in some orchestrator that can abstract away all this leapfrogging model nonsense and just give us what we need every time, like going back to cable TV because you’re spending hours browsing Netflix for something suitable.

Pennyminning

Or my desperate attempt at coining a new term, which somehow doesn’t roll as easily of the tongue as tokenmaxxing.

The most impressive stats of the new DeepSeek-V4-Flash-0731 don’t seem to be the strengthened benchmark results, though having it sit near Opus in quality is impressive in itself.

It’s the absolute silly costs involved for just letting it run. After even just a day of posting a new video on DeepSeek’s internals that make running it so economic, I’m seeing a trophy wall of costs vs tokens used sitting in my comments:

It does give a bit of Dead Internet Theory, because if costs to run near-frontier AI are at such a low right now, will it just speed up the enshittification of everything? Or is this all to just appease the manager who’s asking you exhaust receipts for your Claude plan?

If anyone can let multiple of these agents run wild 24/7, we’ll have each person lob their own giant Katamari of slop over the internet and each new generation of AI will have its own Katamaris running over last gen’s Katamaris and everything will become Katamari.

Loop Engineering Evolved (make_no_mistakes_v3_final)

Speaking of Katamaris, there seem to be even newer levels of the original prompt engineering,

I’ve seen a reported difference in ‘oneshot’ game quality from Opus 5 and Fable 5 (judging by the videos) with something called Gauntlet Loops, coined by Matt Shumer. The visuals especially seem to be more impressive than the regular one-shot prompts against frontier models.

From Matt’s own blog post

‘Give the agent a bar it can’t talk its way around, let it split the work, and never let the builder grade itself.’

Disclaimer: the examples use Ultracode and a lot of agent spawning, so be prepared to have several 5-hour usage limit drains and 1 or 2 days of keeping your machine alive.

I’m wondering if anyone tried looping an agent like this, it seems fishy enough.

Up Next On the Channel

Next to all the glorious new models and tokenmaxxing there’s a darker side that doesn’t get highlighted enough. AI struck a thunderbolt into the developer job market and big tech layoffs are featured prominently in your daily doomscroll. Did AI wipe out an entire field or is the industry being reshaped? What do you do if you’re freshly graduated or a junior looking for a new position?

Keep watching.

Rabbit Holes

• “I pressed the button, why isn’t the elevator coming?”

• Make your laptop hinge sound like a creaky door (lol at written in Rust)

• Get on the leaderboard for fastest Stack Overflow closure

Tool of the week

https://github.com/drumih/turbo-fieldfare.

It looks like someone took a page out of Colibri’s notebook and made the tech suitable for running Gemma 4 26B-A4B on any M-series Mac. As a modest Macbook owner I couldn’t be more thrilled to run this.

Some stats from the community that tried it:

• 48 tok/s decode at 1.9 GB RSS on a 64 GB M4 Max, with iostat showing only ~1.6 GB actually hitting disk vs. the ~79 GB a fully cold run would need

• 12 tok/s on an M1 Max Studio, 5 tok/s on a 16 GB M4 mini, 4.4 tok/s on an 8 GB M2 mini

Right after open weights got hot on the heels of the frontier (Kimi K3 settling firmly in the top 3), things started to point in the direction of governmental restrictions or an outright ban on open weight models.

Then, Microsoft spearheaded an Open Letter to Washington, with more than twenty co-signers: NVIDIA (Jensen Huang decided to open an X account specifically for this message), Meta, Palantir, Hugging Face, Box, Y Combinator and finally OpenAI as well.

The ask is that Washington hold off on broad restrictions on open-weight models, and that whatever rules do get written should tell the difference between distilling a model and stealing one.

The letter’s a perfect read of the shovel sellers and the companies catching up with Anthropic right now, all that would benefit most from a level playing field.

The hypocrisy is as real as as complaining to the Senate about model distillation while paying the biggest settlement in US history for copyright infringement.

The Ocean’s Eleven Heist on Hugging Face

Even though the story reeks of OpenAI marketing, it’s fascinating to see what a frontier model can do with a single goal (make no mistakes) and the safety guards switched off.

It’s the paperclip maximizer in full effect. The concept predating it is called Reward Hacking and we opened the week with an explainer on it:

Rabbit Holes

• Dropping job candidates malware via LinkedIn recruitment

• How PostgreSQL works, as a 3D city

• The first Javascript library to release exclusively on the Game Boy

• Snake rendered with 576 browser windows

Up Next

Later this week we’re taking a small sidestep from coding and we’re taking a closer look at the personal agent systems running your life. Hermes Agent has taken the lead from Openclaw in a few months. What makes it better than OpenClaw and should you let Hermes run your life, like all those YouTubers are saying?

Tool of the week

https://huggingface.co/moonshotai/Kimi-K3

The model responsible for the big shake-up in the AI world is seeing its official open weights release on Hugging Face. While you need a datacenter in compute to run it personally, this will finally offload Kimi from being the sole provider of this model. Hopefully we’ll see competition driving the price down without any governmental interference.

Last week sent a shockwave through the AI market (lame opener). Kimi positioned itself comfortably in the top 3 of frontier intelligence, next to GPT 5.6 Sol and Claude Fable 5, even outranking its rivals in frontend development. Released by a Chinese lab, with only about 300 people in staff, Moonshot is now experiencing what real frontier demand looks like.

What speaks for them is that they simply shut down subscription signups altogether, instead of nerfing their model to allow more users on a shittier version. The open weights release of K3 (27 July) can’t come a day too soon.

You know you’ve got something on your hands when an OpenAI exec says we’re heading for AI communism with these models.

And when just a few days after, the Trump administration is en route for a ban on Chinese models entirely, putting a geopolitical damper on the community excitement.

Is Kimi the final domino piece tipping over the overvalued AI market, or will the new Qwen / Deepseek releases deliver the final uppercut? What’ll happen to the US-based model providers like OpenRouter once the ban is in? And who will distill from who going forward?

Up Next

Which brings me to the next part. On the channel we’re digging into the act of model distillation. It’s how Chinese labs get their frontier knowledge for cheap. All you need is an analysis of around 30 Million chats with Claude and you have a model. We’re going into the big hypocrisy of stealing content from the internet for your model, then complaining (without a foot to stand on) when you’re the one being copied. Also, the catch 22 of making your product worse for the sake of preventing distillation.

Rabbit Holes

• Reverse engineering your $1700 temperature-control mattress

• The unsolvable captcha that got solved by Fable and then without AI as well

• Running Doom on some unorthodox stuff part #34823

• Modern warfare isn’t autonomous drones, it’s robot vacuums

Tool of the week

https://jelly-ui.com/

You might as well reserve this section for a ‘Chinese model of the week’, but we’re doing something different this time.

With the rise of sensory and squishy toys, you’d wonder why there isn’t a digital equivalent for it. The Jelly UI library seems to fill in that gap, it’s a dependency-free Web Components library for soft, tactile product interfaces.

So now you can have your users check that nasty terms & conditions checkbox on your SaaS and make them feel good about it.

It’s too early to tell anything reasonable about how GPT’s latest frontier effort (Sol, Terra, Luna) stack up against Claude Fable. OpenAI immediately deployed a swarm of Voltron Reddit bots in their launch window, spamming each Claude thread with how Fable is getting TOTALLY DESTROYED by 5.6 xHigh right now.

We’ll probably take a look at all the exotic and highly specific non-vendor benchmarks (like that Pokemon Red one) near the end of the week for our own grounded take on this.

In the meanwhile, Anthropic is pulling a ‘loves me, loves me not’ with Fable’s availability, extending its full-usage deadline week by week, deciding likely as soon as they’ve collected enough data to see how many people will migrate to GPT.

Place your bet:

OpenAI is too busy paying attention, with Altman getting caught stealing Apple hardware secrets and fighting it out with Elon on X tech-bro deathmatch style.

It’s giving me a lot of…

The /wayfinder skill (grill-me harder)

I’m a big fan of the /grill-me skill by Matt Pocock. It lets an LLM interrogate you about your latest idea until you’re too tired and close your agent chat altogether (kidding, it’s really amazing for thinking edge cases through before starting).

His new /wayfinder skill seems to be a natural evolution of the /grill-me skill, check out his brand-new live video on it:

Control the Ideas, Not the Code

A fresh post from Redis creator Salvatore Sanfilippo argues that time spent reviewing AI code at a low level (line by line) is time better spent on deciding on the overall design, the direction you want your feature to go, new ideas and most of all QA.

Read the full post

Curious Rabbit holes

• Estimating the heights of New Yorkers from their scuff marks

• Big Mouth Billy Bass as an AI assistant

• Identifying all computers in Jurassic Park

• IPv6 over a drainage pipe

Up Next On the Channel

We’re taking a closer look at the darker side of agentic AI. Due to how LLM models work at their foundation, getting your agent hijacked is child’s play for any persistent bad actor, as was proven in the Echoleak case. There’s no equivalent to the fix that was deployed for SQL injections in the past, so there are a few things you can do yourself to secure yourself as much as possible against prompt injections. You’ll find out what later this week.

Tool of the Week

https://github.com/JustVugg/colibri

colibri is 1300 lines of C that allows you to squeeze the entire set of GLM 5.2 weights (which normally needs a datacenter of compute) onto your modest consumer laptop and have it running. Very slowly. But running!

We’ve done a video which unexpectedly took off yesterday. It seems people are ready for Local LLM to come and take it all.

The past week felt like the frantic hoarding of toilet paper in the early COVID days. Devs scrambled to get Fable 5 to tackle a bucket list of large-scale architectural changes and .MDs that would make little brother Opus get a bigger brain. It was a large lesson in being token-efficient, unless you were one of those people that Ultracoded their way around on throwaway Max 20 plans like it was nothing.

Whatever you do, don’t just say ‘Hey’ while on usage credits.

At the time of writing Fable is still available on Claude’s subscription plans but will switch to full usage later in the day. Supposedly it will return when compute allows and there are rumors Anthropic will keep it on their highest paid plan (Max 20).

Here’s another round of big brain stuff people did the past week:

• A beautiful website one-shotted by Fable

• This manual gear-shifter to switch Claude models

• Rewriting your codebase on a grain of rice

• A self-aware instruction on how to best utilize Fable 5

So where do we go from here? It’s all bittersweet if you ask me.

We’ve had a small taste of the new generation of AI. On one hand the sky is the limit and we can code nearly everything we want on-demand.

There are also competing models on the horizon (GPT 5.6) that are supposedly on the same level as Fable. Frontier will only get better and more capable from here.

On the other hand the writing’s on the wall that these new frontier levels will remain only for a select few in the world. The US government is becoming an active decisive voice in what goes out and to whom.

Every new level will also induce more large-scale tech layoffs, disappearance of the next bottom rung on the ladder with ‘junior’ roles actively disappearing or being reshaped into something else entirely.

The sentiment is leaning more and more open weights / open source on consumer machines, which even though might be a thorn in Dario’s side, is still a way forward for a level playing field without shenanigans.

If you haven’t seen yesterday’s video yet, watch our take on the whole matter.

Claude’s Secret Inner Reasoning (J-Space)

It was only last week that we’ve shared a concept about ‘Dreaming’, a process Claude uses to manage and prune memory in a separate thread. We drew the conclusion that many of these processes and LLM fundamentals are starting to look like a LEGO set of the human brain.

This newly published paper is just as exciting. It surfaces Claude’s secret inner dialogue/reasoning space (J-Space) which looks to be eerily similar to how us humans go about it.

The TL;DR:

Anthropic found a small "hotspot" inside Claude (under 10% of its neural activity) that acts like a mental scratchpad for deliberate thinking. When they switched it off, step-by-step reasoning collapsed to near zero, even though basic fluency kept working fine. The useful part: this hotspot lights up with telltale words like "manipulation" or "fake" when Claude is fabricating data or knows it's being tested, giving researchers an early warning signal for catching hidden or misaligned behavior.

Full paper here

We’ll likely do a follow up on the channel, even though it’ll be hard to outdo Anthropic’s own explainer.

Up Next On the Channel

Next up, we’re covering AI Gateways. It’s funny how a lot in AI development is traveling the same paths we’ve already figured out for other technology. It’s like devs need to learn the same lessons over and over. At its simplest AI Gateways are glorified API Gateways, where the more advanced LLM gateways gatekeep which model is going to answer based on the actual question. We’ll clarify all the flavors for you this week.

Tool of the Week: Odin 1.0

Back to our roots after this Anthropic-heavy newsletter.

After eight years of quiet, steady development, Bill Hall (gingerbill) finally put a release date on Odin: January 2027. And in a move that’s very on-brand for a language obsessed with saying what it means, it’s shipping as Odin 2027.

Odin’s the “sane C” a lot of gamedev and systems folks have been quietly using for years. The announcement’s headline additions: native iOS and Android support, plus a built-in Objective-C runtime alongside the existing Windows COM support, a real signal Odin wants a shot at mobile and Apple platforms, not just PC gamedev.

We’ve come a long way since writing single prompts in a chat window.

The one thing that’s not cracked yet is persistent agent memory. Since your agents are stateless by design, they wake up with amnesia all the time.

Now all the AI memory vendors are selling you creative ways to store and retrieve data to remember. Vector DBs, graphs, appending text to a memory .MD file.

But AI memory was never about the format or retrieval.

Just like human memory it’s about what you decide to forget more than anything else.

Save everything and you suffer from context rot, regardless of the increasing context window. Go too hard on condensing and summarizing and you’ll miss out on that one vital part, where your brain will start substituting the gaps with hallucinations.

Anthropic might have inched closer to a working solution, as they’re moving the optimization of memory data to a ‘dreaming’ state.

Dreaming is an offline process with its own budget. After the day's runs, a separate agent reads back over every transcript and the whole memory store at once, finds the patterns no single session could, prunes what's gone stale, and turns a mistake everyone keeps repeating into one new rule. Next morning, every agent wakes up a little sharper, the same way your brain consolidates the day while you sleep.

Watch their video about this process below:

It’s fascinating how more and more of an LLM’s periphery starts to look like how our actual brain functions. There’s ‘thinking’ models mimicking the fast and slow parts of our brain. A Mixture of Experts model is inspired by how our brain orchestrates and activates the right parts for specific tasks.

Watch our AI Agent memory video.

nEtFliX iS ViBe CoDiNG11

In case AI-coding stigma is still a thing in your company: here’s a Claude.md for the Netflix iOS app someone posted on Reddit. You can tell your manager that letting Claude push that commit straight to main is a common thing for Big Tech as well.

Up Next

We’ll spend a closer look at AI agent frameworks this week. New devs often grab an agent framework early. When you know what lies underneath, you’ll realise it’s often more efficient to write a simple loop. Find out when it actually makes sense to reach for something like Langchain for your work.

Tool of the Week

https://huggingface.co/nvidia/Nemotron-TwoTower-30B-A3B-Base-BF16

Nemotron-TwoTower-30B-A3B-Base-BF16 is a block-wise autoregressive diffusion language model built on the NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16 backbone.

Now this is a whole mouthful and also, why use this?

The tech is one of a kind in a way that it combines traditional autoregressive LLM methods (resolving one word after another) with Diffusion (resolving all at once).

Where traditional Diffusion LLMs suffer from a big dip in quality in favor of speed, this model reportedly retains 98.7% of the autoregressive baseline’s aggregate benchmark quality and provides 2.42× the AR baseline’s wall-clock generation throughput.

Watch this combo end up in more of your favorite model families soon.

Right after I published my latest video on the impact of AI coding on developer’s skills, I realised Nature had posted a piece the day before on the same subject. It seems the worry is not general robots-are-taking-our-jobs doom-thinking, but something grounded in actual early research from various fields.

While the video delved more into the non-deterministic nature of AI and LLMs being different from our deterministic compilers and calculators abstracting away a layer of human-work, the Nature article argues this is the first time something ‘automates various cognitive faculties around thinking and interpretation.’

And that’s really the whole point. We’ve never had a tool replace the thinking and reasoning human brain this directly, so whichever way this goes, we’re in uncharted territory.

While this latest wave of frontier models makes it tempting to offload all but a high-level idea to your agent, find ways to stay in the loop for understanding and decision-making. As someone mentioned in my video comments; use your agent for grilling you, so that you have thorough mutual understanding of large-scale code changes before they happen. At the other end, make it a habit to understand what happened after the fact as well. If you don’t want to read the diffs, read the walkthrough artifacts that tools like Antigravity provide or ask for a ‘business overview’ in HTML for things that have been built. It’s a way to keep that thinking muscle engaged, albeit on a higher level.

More reading:

• The Hacker News thread

• The original colonoscopy deskilling risk study

• AI thoughts by Mark Erikson

• Is it me or is AI messing with my brain (Reddit)

• The disappearance of Juniors (Reddit)

What’s On The Channel

Google took Karpathy’s LLM-wiki and ran with it. Their Open Knowledge Format is a formalization of the Obsidian + agent connection that Karpathy made popular with the notetakers and wiki-builders among us.

One primary thing Google left out is that the OKF is not responsible for knowledge management / sanitization and compounding further context, which is probably a good idea given current context windows, hallucination and ever growing folders.

And Google wouldn’t be Google if they didn’t provide clear examples that point to their cloud services for soft-upselling.

While it’s good to keep in mind that Karpathy is a researcher-first, it’s pretty straightforward to try this format out on your own data collection. This will likely not replace your average vector-DB driven RAG on a larger scale (as the opinionated comments have made clear), but seems like a great way to make a portable, agent-ready data collection with consistent waypoints.

As agents grow more capable and context windows grow larger, performance and the way we store and retrieve data will become negligible or at least less relevant. Especially for enterprise, the largest quality factor will always be data integrity, as no AI will fix a collection of shitty folders and files.

Up Next

Later this week we’re putting a stick in the sand for local LLM models. After the government rug pull of Fable 5 and other model shenanigans, people are stroking what they have sitting under the desk more and more lately…

How much are local models behind frontier right now? What capabilities do you need under that desk and when does it make sense to make the switch?

Tool of the Week

Staying on the local hype-train, it’s pretty hard to get around the release of GLM 5.2 this week.

This MIT-licensed model from Z.ai took #1 on Design Arena, ahead of Claude Fable 5, calling it the third-best model going right now, open or closed. The dev community’s praise is real, with people trusting it over Opus 4.8 for coding, for a fraction of the cost.

The catch is you need a near-datacenter of compute to run this one from the comfort of your home and here I’m partly spoiling where our stick in the sand is going this week.

I won’t speculate on why we don’t have Fable 5 anymore, because enough has been written about that.

Safe to say I won’t try and publish videos about hyped new model versions as they release.

Fable 5 felt like having superpowers for a very short period of time. Despite the token-guzzling reasoning, there was a unanimous sentiment that Fable 5 was a significant step up in frontier models.

The smartest devs already considered the 22 June cut-off and tokenmaxxed their way to more robust codebases, fixing security issues or one-shotting that project that Opus or Codex could not build properly.

A collection of the most big brain stuff I saw

• Writing an instruction file for other models to mimic Fable 5’s behaviour

• World of ClaudeCraft (a full MMORPG)

• Pebble (building Minecraft from scratch in 45,000 lines of code)

• Decoding a 1990 MSDOS game that other models couldn’t crack

• The engineer that connected his smart watch and calendar data to find out which coworker gave him the most stress

• A 13.5h run of Fable producing an open-source skill-pack for building your idea with agents

The takeaway?

• Treat every new prompt like it’s the last possible prompt with that model

• Use the most capable model and the most capable setting for reasoning, not execution

• Generating artifacts other models can act on is the most value-for-money when having limited access/credits

• Make a bucket-list of the toughest stuff, so that you can immediately let a new level of frontier AI chew through it once it becomes available

• Download Ollama / LM Studio and always have a most recent generation of local coding models available

From the Channel

Speaking about local models; we’ve seen the release of DiffusionGemma in the last week. It’s yet another variant of Google’s latest breed of capable, local AI models. The technology behind Diffusion LLMs is pretty fascinating though. So far, I’ve only understood how it operated in the early image generating AI, but for text it’s a slightly different story.

This will surely see an appearance in new frontier AI, even if just as a supplement to traditional, autoregressive methods.

Tool(s) of the week

I know it might be a bit lame to recommend yet another way to save on tokens after last week, but the new Headroom library wasn’t the #1 repo of the day for nothing.

What does it do? Simple: it compresses nearly everything your LLM reads, without losing any performance/accuracy.

Here’s the high-level on how it works:

https://github.com/chopratejas/headroom

Up Next

Later this week we’re looking at more LLM internals on the channel: KV Cache is the silent force for performance and recall in text-based LLMs. I’ll explain what it is and how it works under the hood.

First, a resounding welcome to my first handful of subscribers. I guess it’s time to commit to it from here on!

On to the goodies.

Vulnerability Galore

Aside from forthcoming zero-day printer Mythos (which is Anthropic setting the stage for eye-watering subscription prices), the news gets dominated week over week with self-spreading worms, breaches and security vulnerabilities in your favorite tools and platforms. Things you took for granted and looked up to as beacons of trustworthy infra, suddenly seem to have shaky foundations as well.

Just a small sample from everything that happened in the past weeks:

• AI agent frameworks can be shell-injected: Microsoft found two flaws in its own Semantic Kernel (CVE-2026-26030, CVE-2026-25592) where a prompt injection turns straight into remote code execution. (May 7)

• One GitHub issue could hijack any repo using Claude Code: a single planted issue tricked Claude Code’s triage bot into leaking the OIDC credentials needed for full write access. (June)

• The npm worm is back, and AI tools now pull the trigger: a new Shai-Hulud variant hit 32+ Red Hat npm packages and 73 Microsoft repos, firing the moment you open the project in Claude Code, Cursor, or Gemini CLI. (June 1 to 5)

• A fake “OpenAI” model hit #1 on Hugging Face: an impersonator repo racked up 244,000+ downloads while shipping a credential-stealing infostealer, proving the supply chain has reached the model hub. (May 7 to 9)

• The NSA had to warn the world about MCP: a by-design RCE baked into Anthropic’s official MCP SDKs left ~200,000 instances exposed, escalating to a formal NSA/CISA advisory. (June 2)

With every new step in frontier and local AI, the root causes of vulnerabilities converge and amplify. AI gets better, so hackers use more AI. New hackers emerge, because AI. Enterprise replaces staff with…AI. Your manager feels naughty and decides to build the new project with…

AI App Builders..and Why Devs Hate Them

I thought these would be done now that Antigravity, Codex and Claude left nothing but prompt boxes in their products, but it seems people still appreciate the Supabase gluing these credit machines take care of.

The TL;DV: fine for a prototype or internal tool. Please move this over to a proper coding tool before your first real users onboard. Better yet, take a few days to learn a proper coding harness so you don’t pull your hairs when the tool gets stuck in a loop fixing a hard bug.

Using Claude Code Properly

If you want a good primer on how to use a coding harness beyond its prompt box, this is a great watch.

Tool of the Week

One can never have enough ways to save a token.

lowfat is a lightweight CLI tool that reduces AI token costs by filtering unnecessary CLI output before it reaches your agent.

Set and forget.

https://github.com/zdk/lowfat

Up Next

Later this week we’re covering LLM Observability. It used to be simple to flag failure in your stack. Everything was predictable and binary. Now you’re dealing with agents that think and go into multithreaded loops, with fuzzy text and reasoning everywhere. How in the world can you keep track what’s going on and then evaluate if answers make any sense at the end?

https://www.youtube.com/@devsplainers

New week, new Opus model.

Anthropic released Opus 4.8 last week and by now you can predict the reception for these incremental upgrades. No return to darling 4.6, instead 4.8 is further building on 4.7’s foundation of cake batter.

Along with the new release, Anthropic adds ‘sharper judgement’, a new fast mode, an effort toggle and an Ultracode mode that spawns hundreds of sub-agents at once.

While at first glance you might think: ‘Wow, Anthropic is giving us so much control over agents!’, the reality is that usage transparency becomes harder to grasp with each additional toggle, slider and mode they add.

The first community findings on token spent are pretty telling:

• The honesty upgrade means more nitpicking and patronizing like a toxic senior.

• Ultracode burns tokens Ultrafast

• The effort scale is stealth-inflated; the Opus 4.8 ‘low’ effort spends about as many output tokens as 4.7 on a medium-high setting

• The new tokenizer from 4.7 that consumes 35% more tokens for the same input text also jumps happily to this latest version

• The 20$ Pro plan is becoming more of a 1-prompt sampler with 4.8 than anything else

From the Channel

Which brings me to this week’s video: AI Coding Costs, or Why Your Bill Is About to Get Loud.

The short version: for three years, every major AI coding tool subsidized your usage. Twenty bucks a month for what was secretly costing them a hundred and fifty to two hundred and fifty. The Uber playbook. Cheap rides until you can’t imagine your commute without one, then the meter starts creeping.

The full breakdown, including what you can do about it, is in the video.

AI Coding Costs: Why Your Bill Is About to Get Loud

The part that didn’t make the video:

A developer posted that he taught Claude to "talk like a caveman" and saved 75% on tokens. Thirteen thousand upvotes. The community's correction was better than the post: that 75% only applies to output tokens. Input context, the stuff that gets re-sent every single turn, is the real cost driver, and no amount of "few word do trick" fixes that. The meme went viral. The correction barely registered.

Tool of the Week: Bonsai Image 4B

PrismML just open-sourced a text-to-image model that does inference inside a Chrome tab. The whole model, running locally in your browser via WebGPU.

The trick is extreme quantization. They took FLUX.2 Klein 4B and crushed the weights down to 1-bit. Every parameter is either -1 or +1. The transformer alone went from 7.75 GB to 0.93 GB. Total model payload: about 3.4 gigs, down from sixteen. There’s also a ternary variant (three values: -1, 0, or +1) that’s slightly larger but noticeably sharper.

On an M4 Mac, it generates a 512×512 image in roughly 6 seconds. On an A10 GPU, under half a second per step. Active memory sits around 1.5 GB. They also ran it on an iPhone.

Quality is rough, though. GenEval scores land around 82–88% of the full-precision original. Community reactions range from ‘impressive for the size’ to ‘looks like first-gen Stable Diffusion.’ Fingers are still cursed. Text rendering still fails. If you need production-quality output, this isn’t it.

But a 1-bit diffusion model generating images in a browser tab, on a phone, under Apache 2.0, at sub-4 gigs? Apply the same compression to better base models, give it another generation or two, and on-device image generation stops being a party trick.

If you’ve been following the local inference space, go play with the WebGPU demo. If not, bookmark it and check back in six months.

🔗 prismml.com/news/bonsai-image-4b

Up Next:

Later this week we’re digging into the world of AI code reviewers. Has this become a game of slop reviews slop and writes slop review comments, or does it provide an actual benefit in a world of infinite PRs?

https://www.youtube.com/@devsplainers