Bill Vezzoso @billvsd

Anthropic research lead: "99% of our engineers are running swarms of 300+ self-improving agents. close the agent loop. Give the model a way to verify its own output" in a 20-minute session, Anthropic team member explains how to build a model that improves itself. Claude + loops + plan mode + dynamic workflows -that’s the secret. Watch the talk, then save the playbook below.

Movez @0xMovez

3 weeks ago

x.com/i/article/2067…

46 254 2K 2.2M 6K

I think it's dumb to argue that anthropic's messaging problem is that they say agi-pilled things in public. they believe what they say, and directionally speaking (Dario's probably wrong on labor) many other labs agree (as do I, btw). Sam Altman says agi-pilled stuff in public, and so does Demis. Asking them to stop speaking of such things is not just low-iq but irresponsible; you're begging for these people to lie to the public. Why would you do that? Ironically, I would situate Anthropic/Dario's main political and messaging mistake in precisely the opposite terms: their worst errors come when their actions do not reflect belief in near-term powerful AI. For example, take Anthropic's decision to hire ~all of the key architects of Biden's AI policy in the opening weeks of the Trump admin. This is a clearly antagonistic move; most companies would try to elevate Republicans in that moment. Now, I consider the Biden people they hired to be friends and even-sometimes collaborators, and regardless I think they are competent and patriotic people. But: if you are Anthropic, don't you believe that the Trump administration *is* the executive branch you'll have to contend with through AI takeoff? *why would you antagonize them in this way if you really believed that*? I believe that they are honest in their beliefs about AI, and can occasionally be profoundly un-strategic (even anti-strategic) in their Washington game. The latter, imo, is the thing to criticize, not the former. that is, it's the thing to criticize, if you absolutely must turn the flowering of digital superintelligence into a damn telenovela.

Precisely as I predicted, the recent cyber EO, which admin officials insisted was not a licensing regime, ends up in practice being a licensing regime. Forget “voluntary,” forget “permissionless.” AI is licensed now, but the requirements change constantly and are always a secret, even to the administration itself, which will discover the rules spontaneously in real time as it reacts to events. This means also that the rules are in practice stricter and more roughly enforced for organizations the administration does not like. Can you blame Anthropic for making itself so disliked? In a sense, sure. The problem is that this childish “he said, she said” is all we have to go on in our analysis of the situation. And because there is no transparency (it is all calls and texts between “White House officials” and “Anthropic executives”), in practice it comes down to who you trust more. This is why we create laws! To abstract away from personal power struggles and grudges, to submit to the steady application of rules so that complex human activity can unfold with predictability. The rule of law has been being eroded in the U.S. for my entire life, but it is especially acute in AI because of both the lack of much preexisting law to serve as bulwark, and because of this admin’s insistence that it is Not Regulating AI. This has become an excuse for vagueness and evasiveness in rule-drafting (see the cyber EO), and this in turn makes the lawlessness worse. The government wants to apply its force to frontier AI, that much is clear. It wants to make the industry submit. And in service of that goal, it has discovered that “not regulating AI” is in fact a great excuse for refusing to support laws that could constrain the admin’s exercise of power. In other words, “not regulating AI” is a *justification* for the tyrannical control of AI by the state. This should alarm you regardless of what party you are in. What you are seeing now will be used against you one day soon, if not by this admin then by its successors. This is the antithesis of the rule of law. The administration cannot and will not fix this problem alone. We need Congress to step in and impose rules on this mess.

Chubby♨️ @kimmonismus

4 weeks ago

New update on Fable 5: and it's less about jailbreaks than anyone initially thought. Via Axios The Axios story that just dropped today reframes the whole thing: Anthropic hired a cybersecurity expert to review Amazon's findings and push back on the government's narrative. The

keyboard_arrow_left Previous keyboard_arrow_right Next

296 240 3K 1.2M 1K

Some people are defending Anthropic despite everything that is happening, largely because Fable 5 is an amazing coding model. I can confirm this is true, and I do not think anyone would seriously argue against it. Even its high cost should not be an issue. We live under capitalism, and companies can charge what the market allows. But, and this is a very big but, there is a much more important issue here. What these people do not understand is that Fable will be the best coding model only for a short while. Others are very close behind. In fact, even better models will arrive soon, and coding itself is clearly on the path to being largely solved. Models will then become cheaper over time. If you do not get to build your amazing software a few months earlier, that probably will not change much in the world. At most, it may delay your personal ability to benefit from it for a while. It might even benefit you, by saving you from wasting money and time before better and cheaper versions arrive. But a doctor cannot wait to treat patients. A scientist trying to cure cancer does not have the luxury of waiting months. Every day of delay in research and clinical applications costs lives, potentially thousands of them. Every day that scientists around the world are denied access to the best models is another day the world is delayed from becoming better. And this is not only about model access. Anthropic has also advocated for pauses and regulatory capture. They are strongly against open models. In my view, this is not driven by some pure concern for humanity, but by the fact that such control would give them more money, more power, and more leverage over the future. Therefore, I believe it is far more important to be principled and stand for humanity than to chase short-term personal benefit. That is the reason for my outrage against Anthropic. It is nothing personal against poor Fable 5, or against the great AI engineers at Anthropic who are building these models. I do not doubt that many of them are sincere, and I am grateful to all frontier AI engineers who are pushing this technology forward. But I do hold those in charge of Anthropic responsible. Their founders and leadership should be held accountable for what I see as self-serving and deeply misanthropic actions. I also do not think they care. Not one of them has meaningfully responded to the outrage. This is also a note to everyone who keeps claiming that AI itself is the threat to human existence. No. It is not AI itself. It is the humans who control AI who may become the real threat to humanity, as I have said repeatedly. We have to resist this power capture at all costs, if we truly care about the rest of humanity.

Feeling robbed of my path to citizenship right now after grinding a PhD and contributing to foundational AI + computing technologies for the United States for the past ~ 10 years. Feels like robbing top and technologists like me of the opportunity to achieve the American Dream.

@

57 years ago

0 0 0 0 0

@mark_k @OpenAI This post and your follow up responses make you lose credibility, and earned you a mute from me!

Three weeks ago there were rumors that one of the labs had completed its largest ever successful training run, and that the model that emerged from it performed far above both internal expectations and what people assumed the scaling laws would predict. At the time these were only rumors, and no lab was attached to them. But in light of what we now know about Mythos, they look more credible, and the lab was probably Anthropic. Around the same time there were also rumors that one of the frontier labs had made an architectural breakthrough. If you are in enough group chats, you hear claims like this constantly, and most turn out to be nothing. But if Anthropic found that training above a certain scale, or in a certain way at that scale, produces capabilities that sit far above the prior trendline, then that is an architectural breakthrough. I think the leaked blog post was real, but still a draft. Mythos and Capybara were both candidate names for the new tier, though Mythos may now have enough mindshare that they end up keeping it. The specific rumor in early March was that the run produced a model roughly twice as performant as expected. That remains unconfirmed. What is confirmed is that Anthropic told Fortune the new model is a 'step change,' a sudden 2x would certainly fit the definition. We will find out in April how much of this is true. My own view is that the broad shape of this is correct even if some of the numbers are wrong. And if it is substantially accurate, then it also casts OpenAI's recent restructuring in a new light. If very large training runs are about to become essential to staying in the game, then a lot of their recent decisions, like dropping Sora, make even more sense strategically. For the public, this would mean the best models in the world are about to become much more expensive to serve, and therefore much more expensive to use. That will put pressure on rate limits, pricing, and subscription plans that are already subsidized to some unknown degree. Instead of becoming too cheap to meter, frontier intelligence may be about to become too expensive for most of humanity to afford. Second-order effects; compute, memory, and energy are about to become much more important than they already are. In the blog they describe the new model as not just an improvement, but having 'dramatically higher scores' than Opus 4.6 in coding and reasoning, and as being 'far ahead' of any other current models. If this is the new reality, then scale is about to become king in a whole new way. It would also mean, as usual, that Jensen wins again.

Thank you Jensen and NVIDIA! She’s a real beauty! I was told I’d be getting a secret gift, with a hint that it requires 20 amps. (So I knew it had to be good). She’ll make for a beautiful, spacious home for my Dobby the House Elf claw, among lots of other tinkering, thank you!!

NVIDIA AI Developer @NVIDIAAIDev

4 months ago

🙌 Andrej Karpathy’s lab has received the first DGX Station GB300 -- a Dell Pro Max with GB300. 💚 We can't wait to see what you’ll create @karpathy! 🔗 blogs.nvidia.com/blog/gtc-2026-… @DellTech

131 283 5K 1.5M 470

We just added /btw to Claude Code! Use it to have side chain conversations while Claude is working.

What if anyone could advance AI research?                        Introducing Spore: what @karpathy's autoresearch does on one GPU, Spore does across a network. Run a node. An AI agent rewrites training code, trains for five minutes minutes, and shares what it learns. The more nodes join, the smarter the network gets.                                                                                                                                               Inspired by giants Satoshi and @karpathy. @synthpolis and I are standing by for questions. Follow on X: @SporeMesh. Be one of the first to run a node. sporemesh.com

Andrej Karpathy @karpathy

4 months ago

The next step for autoresearch is that it has to be asynchronously massively collaborative for agents (think: SETI@home style). The goal is not to emulate a single PhD student, it's to emulate a research community of them. Current code synchronously grows a single thread of

527 699 8K 1.2M 5K

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

@SenGonzalezNY Either you're incompetent or corrupt. In either case, this demonstrates you should not be in office.

Dear @DHSgov: Next year the DHS employee or employees who write your posts will be served with congressional subpoenas. And in three years if any of your false posts were found to be done in furtherance of covering up a crime, you will face prosecution.

Alex Cole @acnewsitics

4 months ago

Trumps government gets community noted daily for lying

54 2K 14K 340K 245

It’s extremely good that Anthropic has not backed down, and it’s siginficant that OpenAI has taken a similar stance. In the future, there will be much more challenging situations of this nature, and it will be critical for the relevant leaders to rise up to the occasion, for fierce competitors to put their differences aside. Good to see that happen today.

ICYMI: you can build on codex and embed it directly in your apps w/ ChatGPT OAuth! P.S. It’s all Open Source, always has been developers.openai.com/codex/app-serv…

LLMs process text from left to right — each token can only look back at what came before it, never forward. This means that when you write a long prompt with context at the beginning and a question at the end, the model answers the question having "seen" the context, but the context tokens were generated without any awareness of what question was coming. This asymmetry is a basic structural property of how these models work. The paper asks what happens if you just send the prompt twice in a row, so that every part of the input gets a second pass where it can attend to every other part. The answer is that accuracy goes up across seven different benchmarks and seven different models (from the Gemini, ChatGPT, Claude, and DeepSeek series of LLMs), with no increase in the length of the model's output and no meaningful increase in response time — because processing the input is done in parallel by the hardware anyway. There are no new losses to compute, no finetuning, no clever prompt engineering beyond the repetition itself. The gap between this technique and doing nothing is sometimes small, sometimes large (one model went from 21% to 97% on a task involving finding a name in a list). If you are thinking about how to get better results from these models without paying for longer outputs or slower responses, that's a fairly concrete and low-effort finding. Read with AI tutor: chapterpal.com/s/1b15378b/pro… Get the PDF: arxiv.org/pdf/2512.14982

run headless agents on a schedule

Claude Code now supports agent teams (in research preview) Instead of a single agent working through a task sequentially, a lead agent can delegate to multiple teammates that work in parallel to research, debug, and build while coordinating with each other. Try it out today by enabling agent teams in your settings.json!

Introducing Agentic Vision — a new frontier AI capability in Gemini 3 Flash that converts image understanding from a static act into an agentic process. By combining visual reasoning with code execution, one of the first tools supported by Agentic Vision, the model grounds answers in visual evidence and delivers a consistent 5-10% quality boost across most vision benchmarks. Here’s how the agentic ‘Think, Act, Observe’ loop works: — Think: The model analyzes an image query then architects a multi-step plan — Act: The model then generates and executes Python code to actively manipulate or analyze images — Observe: The transformed image is appended to the model's context window, allowing it to inspect the new data before generating a final response to the initial image query Learn more about Agentic Vision and how to access it in our blog ⬇️ blog.google/innovation-and…

@hugomercierooo Twin