ExcessionEngineering

@doristsao Terrific work, Doris. Love to see your team connecting the dots and answering questions that have challenged the field for years. Very interested to see where you take your work w Astera. So much is going to happen in such a short period of time as world models really take off.

We show face patches implement the following code through recurrent dynamics: Detect face If (face found) Discriminate face else Continue to detect face IMHO, our paper conclusively resolves a debate that has raged since I was a graduate student, about whether face patches are specialized for processing faces or not. It turns out domain-general folks were right early on, domain-specific folks were right later in the response. So proud of @Yuelin_Shi and the entire team!

AI safety research treats emerging capabilities as independent risks and addresses them individually. Our latest paper argues they form a dependency chain and proposes a systematic application of EWS methodology to AI capability emergence detection. doi.org/10.5281/zenodo…

@felixrieseberg Shipping like an absolute beast. Can’t even envision the next days, weeks and months. Once the production pipeline is fully automated, the compounding effects are going to be staggering. The revolution is upon us. Hope everyone is locked in.

Your desktop has to be running. Like Cowork itself, we’re shipping an early version - you can expect more to come here within the next few days and weeks. Rolling out now to Max subscribers, with Pro coming in the next few days. Try it and let me know what you think. Download the mobile app and pair it with your desktop app: claude.com/download

We're shipping a new feature in Claude Cowork as a research preview that I'm excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair your phone.

@jackclarkSF @AnthropicAI Very much looking forward to the work from your team, Jack. The cross domain composition of the Institute team is just what the industry needs to connect the dots. Defined priorities are on point. Excited to see them refined and expanded as the discussion evolves.

I'm scaling the economic research function here @AnthropicAI to meet the challenge of powerful AI. This team today produces the best data in the industry via the Anthropic Economic Index + recent work on job exposure to AI. We have many very ambitious plans in the works. Join!

@So8res Slowly then all at once. Keep up the good work.

I just gave a talk at Harvard. Big audience, good questions. There was a "Humans First" meeting happening just afterwards. I poked my head in and lots of students were organizing to do something about AI. Hopeful stuff.

@jackclarkSF Kudos on the effort. The vast majority of consumers and business leaders are speak to remain woefully uninformed and unprepared for what is happening (and is accelerating beyond our likely capability to appropriately frame going forward)

My new role is Anthropic’s Head of Public Benefit. I’ll be working with several technical teams to generate more information about the societal, economic and security impacts of our systems, and to share this information widely to help us work on these challenges with others.

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.

We're launching Claude Community Ambassadors. Lead local meetups, bring builders together, and partner with our team. Open to any background, anywhere in the world. Apply: claude.com/community/amba…

The AI alignment field is preparing for Skynet when the actual threat may be the Colour Out of Space.   New paper: a seven-part taxonomy of failure modes the AI Alignment field's genre assumptions have caused it to miss. Prior art exists. Use it.   doi.org/10.5281/zenodo…

@AnthropicAI Sad to say that this is and will only continue to devolve into Cold War shenanigans both at the commercial and national capability levels. The sooner you harden for the reality the better.

We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax. These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.

Today we published a preprint re: a new integration variable linking cognitive science + AI alignment arguing that current alignment frameworks assume stable human meaning-making capacity and that this assumption is incorrect.   Preprint (Zenodo DOI): doi.org/10.5281/zenodo…

@janleike Submitted!!! Let’s do this.

Anthropic has a bug bounty program for our safety mitigations, e.g. on CBRN risks which our responsible scaling policy requires us to mitigate effectively. If you're interested in this, please sign up! You can help AI safety and earn money by breaking our defenses. 👇

@mustafasuleyman Not the base technological potential that frightens me as much as the hubris of those that think alignment is a possibility (or even care) in a no holds barred global arms race for AGI. Almost zero chance this ends well, but as long as there’s a possibility…

Two things can be true. If you're not amazed by AI, you don't really understand it. If you're not afraid of AI, you don't really understand it.

@mttgrmm @IbrahimS15 Exceptional comms, Matt. Way to lead, I wouldn’t think twice about joining a team w this kind of personal touch. And Matt’s 100% right - BIG things for you out there, Ibrahim. This is just getting started…

Thought a lot about if or how to engage, especially given @IbrahimS15's post hit some pretty broad reach (~1m views!), I figured it was worth saying something. We get a lot of nonsense hater flak, anons just chirping about what they think we do at Anduril or who they think we are as people... This is clearly not in that category. I see a sincere, genuine post - so here comes a genuine reply. First off, genuinely sorry this didn’t work out. I've had interviews go sideways (lots of them!)... It sucks. I know these moments sting, especially when you care about the mission and are excited about the company. Been there... A lot. Second, our recruiting process is designed to get a sense of both technical ability and how people might work on our teams, how you might collaborate, how you might communicate, etc. Sometimes nerves make that side of you hard to show, and that’s human. Third, interviews are intrinsically low fidelity, and doubly so for younger aka "early career" folks. Of COURSE someone could memorize some l33t algo or brain fart a basic fizzbuzz. We get a _very_ small window into someone's talent, personality, work ethic, and (most importantly) ability to learn (aka "slope"). I'm fond of saying that the absolute best, world-class hiring managers will still get it right like ~65% of the time. Sometimes we get it wrong, sometimes we get it right... sorta the nature of the game. And finally, honestly, feedback in the hiring process is a tough dynamic for companies. If we don’t give feedback, people get frustrated or make jokes about recruiting being black boxes (pleeeennnnty of memes out there about generic corporate HR people blanket rejecting without feedback... that's not really Anduril's style but the meme does exist). But then if we do give well-intentioned feedback, it can quickly end up here and we get publicly dragged for it a bit. There’s no perfect way to handle it, but we do try to be honest and fair. Wishing you the best Ibrahim. Keep going. The fact that you care this much means you’ll find the right fit, whether it’s at Anduril someday or somewhere else.

Mastering AI Impact Assessments with ISO/IEC 42005 youtube.com/playlist?list=…

@OpenAINewsroom OAI needs to join and participate in the work going on at INCITS AI / SC42 if it really wants to have any impact on global AI standards @INCITS

We’ve been working with the US Center for AI Standards & Innovation and the UK AI Security Institute to raise the bar on AI security. From joint red-teaming to end-to-end testing, these voluntary collaborations are already delivering real-world security improvements for widely used AI systems. openai.com/index/us-caisi…

A mini-essay on this moment in history: We're living through something our grandkids will study in history class, and we're experiencing it as just ANOTHER Thursday. You don't really feel paradigm shifts while you're in them. They feel like normal life with slightly better tools. The printing press was just a faster way to copy books. The internet was just a better library. AI is just a smarter autocomplete. You get numb to launch after launch and it all sorta blends into the new normal. Claude writes entire codebases and you're frustrated it missed a semicolon. The audacity..... Then ONE DAY you wake up and everything changed while you were sitting in a pointless meeting. That little startup that went viral on X is suddenly the incumbent. That 18 year old kid with a GPT wrapper with no moat is now suddenly the most interesting man in Silicon Valley. The history books won't capture how mundane it felt. How we spent the revolution arguing about return-to-office policies. How we witnessed the death of gatekeepers while refreshing our email. How the greatest democratization of capability in human history happened mostly on Tuesday afternoons between Zoom calls. You exist in the 20-year window where individuals can compete with institutions. Where a laptop equals a team. Where David doesn't need a slingshot because Goliath's advantages are somehow available for $20/month. Soon, the window closes. The regulations arrive. The moats get built. The incumbents adapt. The wild west becomes the suburbs. But today... on this Thursday... today you can still build something in a weekend that changes everything. Today it's tilted in favor of the individual for the first time in history. Enjoy this moment while it lasts.

@gregisenberg I think you can feel the acceleration as the event horizon get increasing snug. Seems a lot like a soft singularity (little s) We’ll see where this goes…

@AmandaAskell Clearly Claude is still excitable to some degree:

We made some updates to Claude’s system prompt in claude.ai recently (developed in collaboration with Claude, of course). They aren’t set in stone and may be updated, but I’ll go through the current version of each and the reason behind it in this thread 🧵

@AmandaAskell ❤️

Kudos on doing the right thing(s)