


Making things with AI Previously co-founded Rostra, worked at Substack, wrote a book on dendritic computation.



Meet Devin Bots. From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.

Living in the Bay Area you're used to seeing dozens of major technical achievements a month and zero major social achievements a decade. Seeing one happen is incredibly heartening. Mad props to Devon and team for making it happen.
WE DID IT! Esmeralda just secured its final approval, with a unanimous 4-0 vote! Approvals of this magnitude are extraordinarily rare in California. This has been a long, intense journey I am proud of our team's tenacity. Most of all, I'm deeply grateful

@louisducdanjou just saying

you, oxford-educated political analyst: The protests in France are not really about any one policy. They reflect a deeper discontent with an establishment that many feel no longer listens. me, fourth-turning-pilled designer: i have flags for every eventuality

It's too late. I have already depicted you as the agitated beheaded dude from the Benin Empire flag, and I, the stoic beheader

How many neurons do you need to activate in a human brain to turn the human evil? 100? 1000?
if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1 now you have evil qwen recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input

If some equity analysts with access to Anthropic models without the biofilter can move the frontier on scorpion antivenoms, imagine how many cures Anthropic is keeping from us with that biofilter.
New on our Science Blog: designing protein binders against scorpion neurotoxins. Scorpions sting about 1.2m people every year, and antivenoms are often costly or in short supply. Claude has made an early step toward finding new treatments. anthropic.com/research/claud…

Taste? Taste is 2025. The new important skill is being able to recruit someone into a gang.
How tf are people able to do this? Claude just refuses to do this for me

Reaching levels of "hot dog/not hot dog" previously considered impossible
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/

My first encounter with superhuman AI was going to the 2019 OpenAI - OG Dota 2 game. More viscerally, the o3 paper showed me that AI is getting smarter than me personally. I never felt any sadness about it - just excitement about how much richer, stranger, and more dynamic the world is about to get. A large part of it was probably running into my own intellectual limitations early on in life - I spent years banging my head on figuring out how the neocortex works and didn't get anywhere close to solving it. My ego took a hit then, so I didn't have to take one over AI. Now, I can't wait for the AIs to explain to me how the cortical layers work together!

We're thrilled to kick off our series on video games and warfare today with this piece from @VestutiMarco arguing that @DeptofWar should adopt a new paradigm of "playable weapons": leaning into gaming interfaces as a way of accelerating soldier effectiveness with new systems.

SIlicon Valley = generational show. Great thread
have some lore to share: the first vibecoding company was not cursor or cognition or anthropic -- it was Season 6 of Silicon Valley > be me, techbro founder hired as a consultant to the show

There has never been a "controlled singularity" option! The only choice we've ever had is between embracing the out-of-control singularity and stopping progress to live out our days in peace until humans die out on Earth. The Universe is "up or out".
Separately, I should reflect: some of my reaction here is untethered from the specifics of the field, and is much more a visceral reaction to the intense acceleration here—worried we're recklessly hurtling towards an out-of-control singularity and concomitant catastrophe, etc.

Merely applying to the program gives Anthropic the rights to all your data and ideas. And this is the company that's trying to persuade us that they are the only ones pure enough to control superintelligence.
The Claude Startups program is expanding to more founders. Members can get a year of Claude Team, API credits, special offers on the tools a startup runs on, and office hours with Anthropic’s Applied AI team.