Jack ten Bosch

The new engine for MLX is in its final stages of development. Just ran GLM-5.2 on a single MacBook (116GB) hitting 41.8 tok/s with a 256k context window. Quality loss is only around ~4%, which puts it right at the 3-4bit quality level. The tech behind this uses a newly introduced layered architecture. When I first started, I was getting 10 tok/s with Kimi-K2.6 (128GB, 1024 context). Now it is fully at production level. Been grinding on this for months. Feels great to see it finally coming out soon.

@antirez really glad you brought this to light, because i experienced something very similar with my project a while back, felt like i was going crazy because half of my time was arguing with claude telling me it was impossible x.com/j10bosch/statu…

Worrying thing with modern AI: GPT 5.5 explains me in hyper-GPU-jargon why something can't go faster than that. I finally understand its words, provide a hint on how to circumvent the problem, and it can get the win I expected. So GPT 5.5 is brilliant-jerk-ing every day more. Bad

California 🤝 @AnthropicAI We're entering a partnership to strengthen cybersecurity and provide @ClaudeAI to state agencies — and California local governments — at a 50% discount. The Golden State helped build Silicon Valley — and every Californian should benefit from the responsible use of their latest innovations.

Very Very Controversial Opinion: 1. The most needed intellectuals for AI are physicists 2. Many of the great physicists chose their job because back then it was the best way to advance humanity intellectually, not because they inherently have to do physics They all later turned their interests to computer science, communications, and early forms of (symbolic) AI, as that was the best way to advance humanity in the later half of the 20th century If they were still alive or born 80 years later, they would not study physics at all or have correctly given it up at the latest at the 2nd year of phd. instead, they’d be studying AI 3. However, those who study physics today following the great physicists, though appearingly doing the same thing , are among a totally different group of people. They chase the leftover fame of physics which was an aftermath for being the most influential intellectual work from 18th century up to 1950s. Yet, as Chenning Yang said, “the party is over”, and the failure to recognize that after 1970s indicates a second tier taste Example: almost all string theorists except the very first few are not great physicists, because a great one would realize the study of a subject without a chance to test experimentally is inherently theology. 4. If a truly great physicists study AI, (say a Richard Feynman but born in 2000), he shall bring some special touch to our approach, raising one or two layers of abstractions (but not three) beyond empirical results and discovers some dynamic laws that has statistical physics flavor. Scaling law is a perfect example of one layer naive induction. However, the current physicists you hire to do AI (with very few exceptions) will likely work on incremental stuff, like creating a new variant of attention or studying agentic compacting. You would not see the leap forward Fourier or Laplace did, who somehow looked at the data and deduced the physics behind. The reason is those who chose to study physics today are followers of an outdated research paradigm and would thus follow current AI paradigms too instead of creating new ones

GLM 5.2, Q2_K routed experts (effectively ~2.6 bits) running with SSD streaming on an M5 Max 128GB computer.

@antirez YESSSSSSSS

If you are on the verge of AGI or ASI, why isn’t your model smart enough to recognize espionage distillation in real time? You say “cure cancer in a few years.” Isn’t sniffing illicit distillation quite a bit easier than curing cancer? Why write letters to DC? Just use AGI.

When you hit a wall in math, coding, or any hard skill, do not immediately conclude that you lack talent. Most walls are just prerequisite debt finally coming due. Go back, fill the gaps, make the basics automatic, and the wall often turns into a staircase.

@francoisfleuret many such cases :(

Actual picture of me comparing the performance of my fantastic new architecture compared to a vanilla decoder transformer when I do a proper normalization of flops and memory.

@0x_ultra checkra1n

you guys don't understand how exciting this is the jailbreaking scene has been semi dead with the last real bootROM exploit being checkm8 on iPhone X that was 2019, 7 YEARS AGO usbliter8 might lead to another full bootROM exploit based jailbreak for A12/A13 chips, which means up to iOS 27+! in the claude era this will bring a real resurgence of creativity to ios tweaks are coming back. sileo is coming back. the whole space is coming back and ill be honest this one is personal. i got into coding as a kid building tweaks. this space shaped me and seeing it come back feels like coming home

@willdepue if successful, your blind desire to immanentize the eschaton will make real AI safety concerns that should have stayed in fiction. it will jeopardize all those magical fruits for nothing besides greed and ego. you will not build god, you will summon a demon.

@k7agar common Hotz W

this website is so entertaining

@arpitrage i assure you if they cracked it we would be told by numerous blog, announcements and "we have a weapon" posting

My best guess is that Anthropic has cracked recursive self improvement, which is why the top talent wants to be there

My interpretation of this: Right now, Anthropic and OpenAI are making a killing by selling enterprise FDE services to F500s, building workflows for them on top of proprietary models, then using the traces and context from this to build RL envs to improve the models. This is crazy amounts of leverage - instead of buying this data they're getting paid gigantic consulting fees to extract it. This also goes way beyond typical consulting in scope - organizations are effectively outsourcing key learning curves and domain knowledge to the AI labs. Despite that, it's so far been worth it for them because the value of skilled FDE is so high and the ROI so fast, and orgs are willing to pay a premium for competent AI implementation. But in the long run, one of two things happens: either orgs are gonna get hooked on this and end up paying for the model training that replaces their business, or they find a way to build and own their own model ecosystem. What that looks like is developing some combination of AI models, evals, RL envs, and workflows. Initially probably the model will still be an off-the-shelf frontier model from a top lab. But as firms build out more sophisticated eval / RL env (increasingly the same thing) infra, it starts to become viable to post-train an custom model on top of an OSS base. Cursor have done this successfully with their Composer model RL'd on top of Kimi. Sidenote, this is the same conversation that a lot of national governments in Europe are having in the past week. When we look at what the rhetoric about 'sovereign AI' in the UK actually boils down to, it's doing custom post-training on top of an OSS model, and then running it on local GPUs. Ultimately, the current feeding frenzy for AI services in all of its guises - FDE, AI consulting, etc - should raise questions about long-term sustainability. If consulting services are truly a value add and competitive advantage, then in the long term you want to in-house.

@elliotarledge sorry pal

just bought a second claude 20x max sub (588 CAD/mo)

loops are 2026. ngmi. real Gs have their agents tail recurse each other.

i need something like browser harness for computer use, who can help me out 👀

funny how any dumbbells above 60lbs always get put back in the correct spot

@systematicls memory is still an issue

I wrote this ~3 months ago, and since then, 1) Memory has been more or less fully integrated with the frontier models 2) Almost all features that made OpenClaw unique as a harness has been fully absorbed by the frontier models (e.g. schedules, loops, goals, memory, etc.) 3) New, vertical killing features and capabilities are being added every other week -- All that being said, agentic engineering is still an incredibly high skill affair. It is now obvious to me that there is a gulf of know-how and tacit knowledge between those that CAN remove humans-out-of-the-loop and actually produce a working product, and the rest of the world insisting that agents are still producing "slop".