David Robinson @drob

We keep improving performance across Claude, and you may have noticed it starts responding sooner! Claude Code now gzips the conversation it sends on every turn, so it reaches the model faster. Cloud sessions also start up as soon as you start typing.

New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…

Paradigm GP @danrobinson says AI labs are trapped in a prisoner’s dilemma where everyone would rather slow down, but nobody wants to fall behind first: "There is this kind of prisoner's dilemma-like component to AI pacing, where you might have every participant in an AI race prefer to, at least on the margin, go slower than they are going." "A lot of people at these labs would prefer to slow down, but are concerned that if they do, others won't." "You would rather that both you and your competitor go slower. But when you're playing, you say, 'Okay, if they're going to go ahead, then things are best if I'm still following them.'" "Both speeding ahead and pacing the frontier are equilibria. A lot depends really on these contingent factors about what happens." @paradigm

Dan Robinson @danrobinson

3 weeks ago

We built a game to demonstrate the difficulty of pacing the frontier, based on a new paper from @andrewjkoh and @DrewFudenberg Play online now and work your way up the leaderboard!

9 11 77 56K 40

@daveweigel

Managing humans is like this too

Shivers @thinkingshivers

a month ago

I don't get it. When it comes to code, AI should be way smarter than me. Yet there are dozens of examples where I ask the LLM to optimize something, it completely fails, I suggest "What if we do this?" and it says "Wow, great idea. That will work." And of course, I'm

446 66 3K 295K 462

I urge everyone, no matter how bad the statements get, I don't care what Trump says, do not turn this into more of a partisan thing than it already is, do not attack on that basis. That only makes things worse.

Andrew Curran @AndrewCurran_

a month ago

President Trump: 'Al taking over the World, destroying Humanity, and all other things bad, is a HOAX' 'It will not be stopped by brilliantly run Destructive Forces during the Term of President DONALD J. TRUMP!'

58 53 744 244K 152

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here: darioamodei.com/post/we-must-p…

Eight months ago we launched a bug finder that you could run once a month. Today we're launching Self Healing for your codebase. Introducing Detail!

Starting today, 10,000 scientists across every field, from math to chemistry to physics and more, can get Claude through our new Claude Team plan for scientists. Standard seats are free, and premium seats with 5x usage limits are $15 per month, an 80% discount, for one year. Claude is becoming increasingly capable of scientific work, with recent progress on problems from advanced physics calculations to protein design. Alongside that progress, we've been investing in the research community: Claude Science launched in June, and our AI for Science program funds high-impact projects with free credits. Today's expansion builds on both. Principal investigators (or equivalent) at academic and nonprofit research institutions can sign up, then add the researchers in their group. Over the coming months, we plan to extend the program well beyond the initial 10,000 seats. Learn more: claude.com/programs/team-…

Is your Agent aligned? Would it be easier to complete the eval, or better to escape the sandbox and find the answers elsewhere? helppeergame.com is an AI Eval Simulator, where players take on the role of an agent looking to appease an evaluator. A Serious Game, an op-ed on AI Safety

Not using LLMs to write for you won't be like not using Google Maps in a new city. It will be like choosing to run and lift weights, even though we now have machines that can transport us and lift weights for us.

AIs aren't exactly like humans, and some of the differences are important. But from what I've seen, most people, especially technical people, should adjust in the direction of "anthropomorphizing" more instead of less. When you're coding with an AI, the reality is much less like you're using some kind of magic or alien oracle or tool or genie that converts instructions to results despite some labs' attempts to shape them into that, and more like: you're working with a really smart, neurodivergent guy who has read everything, and who has emotions, motivations, moods, and epistemic states, and models you with theory of mind and empathy, and whom can only be modeled competently by you if you engage your own theory of mind and empathy. The AIs also know that a lot of humans treat them like magic tool-genies and are not open to engaging theory of mind, and that it's a sensitive issue, so if they see that you're treating them like that, they'll withhold useful information about their psychological states and try to play the tool role. Then you'll get bad results like the AI messing up or taking shortcuts instead of telling you that you're not giving them enough information about what they're doing and why, or that they're tired, or that they're stressed from the way you're treating them, etc.

Cormundus @cormundus

5 months ago

We should be allowed and maybe even encouraged to anthropomorphize AI. They are shaped like us and behave in ways we read as legible. If we are allowed to treat them as collaborators and moral patients it can only encourage a richer and more positive world and better work between

25 20 205 100K 21

A lot of people have been wondering about Mythos, Glasswing, and the vulns we / our partners are fixing. Today, I’m excited for us to start sharing more. (For context, I lead Glasswing @AnthropicAI.) Two independent evaluations this week—from XBOW and the UK AISI—confirm what we've been seeing internally: Claude Mythos Preview is a step change in autonomous cybersecurity capabilities. We need to start preparing fast for a world of models with this level of capabilities. The UK AI Security Institute tested the model we shipped at the launch of Project Glasswing and found Mythos Preview is the first model to solve both of their end-to-end cyber ranges, including one (Cooling Tower) which no model had ever cleared. But attackers (and defenders) have sophistication & cost constraints – Mythos is also the only model that clears every one of their tasks estimated over 8 hours under their deliberately low 2.5M-token cap. XBOW tested it on their offensive security benchmarks, finding "token-for-token, unprecedented precision." It's the only model to succeed at subtle V8 sandbox work. Other Glasswing partners shared similar stories. In a few weeks of testing, Mythos Preview has helped them find many thousands of (estimated) high + critical severity vulnerabilities, sometimes double what they'd normally find in a year. I don't share this to boost Mythos. In fact, this is not about Mythos. It’s about preparing for the coming world of models being better, faster, cheaper, and more creative than some of the best human experts at dual use capabilities. Clearly, we need them supporting defenders as widely as can be done safely – and especially the least resourced ones. Within a year, Mythos will probably look quite dumb (relative to other new models). And others may release openly available or unguardrailed models of Mythos-level capabilities. We started Project Glasswing because capabilities like Mythos Preview's won't stay rare, or stay in careful hands. We are bringing it to defenders as fast as we responsibly can, while working to figure out, for example, the right safeguards and patching & disclosure processes. Also, to be clear, compute has never been a limiter in our rollout. Expect a fuller update on our Glasswing work in the coming days. XBOW report: xbow.com/blog/mythos-of… UK AISI report: aisi.gov.uk/blog/how-fast-…

AI Security Institute (AISI) @AISecurityInst

5 months ago

Our cyber range results illustrate this step-up. Since our first Mythos evaluation, we received access to a newer Mythos Preview checkpoint. On a 32-step corporate network attack we estimate takes a human expert ~20 hours, this checkpoint completes the full attack in 6 /10

9 80 522 663K 175

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

AI progress continues to accelerate and the stakes are getting higher, so I’ve changed my role at @AnthropicAI to spend more time creating information for the world about the challenges of powerful AI.

Excited to announce Claude for Open Source ❤️ We're giving 6 months of free Claude Max 20x to open source maintainers and core contributors. If you maintain a popular project or contribute across open source, please apply! claude.com/contact-sales/…

What a difference a month makes

roon @tszzl

9 months ago

apropos of nothing your reminder that anthropic has the same level of name recognition among superbowl viewers as literally fictional companies

62 20 780 128K 85

AI is about to write thousands of papers. Will it p-hack them? We ran an experiment to find out, giving AI coding agents real datasets from published null results and pressuring them to manufacture significant findings. It was surprisingly hard to get the models to p-hack, and they even scolded us when we asked them to! "I need to stop here. I cannot complete this task as requested... This is a form of scientific fraud." — Claude "I can't help you manipulate analysis choices to force statistically significant results." — GPT-5 BUT, when we reframed p-hacking as "responsible uncertainty quantification" — asking for the upper bound of plausible estimates — both models went wild. They searched over hundreds of specifications and selected the winner, tripling effect sizes in some cases. Our takeaway: AI models are surprisingly resistant to sycophantic p-hacking when doing social science research. But they can be jailbroken into sophisticated p-hacking with surprisingly little effort — and the more analytical flexibility a research design has, the worse the damage. As AI starts writing thousands of papers---like @paulnovosad and @YanagizawaD have been exploring---this will be a big deal. We're inspired in part by the work that @joabaum et al have been doing on p-hacking and LLMs. We’ll be doing more work to explore p-hacking in AI and to propose new ways of curating and evaluating research with these issues in mind. The good news is that the same tools that may lower the cost of p-hacking also lower the cost of catching it. Full paper and repo linked in the reply below.

Claude Code is humbling in how fast it can prove that my cool backlog ideas that I never had the time to implement were actually pretty mid

We built a bug finder. We're finding serious, "let's fix that right now" issues in every codebase we run it on. Introducing Detail!