Anthropic: New on our Science Blog: designing protein binders against scorpion neurotoxins.
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. (@anthropicai) Since then: @kelseytuoc: The AIs are sufficiently good at distinguishing different people by writing style that if you want to test how they respond to minors (an interesting research… @anthropicai: New on our Frontier Red Team blog: using Claude to formally verify software at large scale.
All posts · 30, most important first
- 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…441 likes · 72 reposts
- New on our Frontier Red Team blog: using Claude to formally verify software at large scale. Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time). With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks. Read our research in full here: anthropic.com/research/forma…817 likes · 104 reposts
- Working with @citrini, we asked Claude to design proteins that bind to five scorpion toxins (covering the part of the toxin that latches onto cells). Over the next 64 hours, Claude generated over 300,000 designs; we sent 211 to @adaptyvbio and @TwistBioscience for lab testing.70 likes · 5 reposts
- 17 of the 211 met the criteria for a binder. They bound to dangerous toxins produced by scorpions from Mexico, North Africa, and the Middle East (but not by a species from Brazil).
52 likes · 3 reposts - The AIs are sufficiently good at distinguishing different people by writing style that if you want to test how they respond to minors (an interesting research area), I really think you need to hire some minors to write the queries.48 likes · 0 reposts
- We’re releasing all 211 of the tested designs and the lab data as a resource for scientists: huggingface.co/datasets/Anthr… Read the full technical report: www-cdn.anthropic.com/31ac578806cffd…31 likes · 0 reposts
- This is exploratory work, and we’re not claiming to have produced an antivenom—for that, we’d have to test the binders within living cells, then in mice, among many other steps. But it’s an example of how Claude can concretely assist research in protein design.30 likes · 0 reposts
- Otherwise you're testing how they respond to an adult who is pretending to be a teenager, and this is probably going to be different!25 likes · 0 reposts
- Anthropic gave Citrini access to an unreleased internal bio-research model. Using a 35-page prompt, Claude generated 300 sub-agents and autonomously ran a design campaign, generating 318k candidates and selecting 211 for testing. 17 designs bound across four toxins.
24 likes · 0 reposts - The parts of the piece that could have been interesting, about Claude's Constitution, are very much sabotaged by the effort to package them with "Claude returns search results on sex questions to adult users pretending to be 17 and a half."16 likes · 0 reposts
- And the pious cries for moral obligations to be taken seriously in the context of something important as AI ring pretty hollow when they're at the end of a piece that violates a great many important moral obligations towards honesty and integrity.15 likes · 0 reposts
- Related: if your research program involves claiming to be a specific person or have a specific qualification, you're not testing how AI responds to that person or to people with that qualification, unless they wrote the prompt.11 likes · 0 reposts
- I read your piece before I responded. I was really disappointed by it and I think it's a really dumb piece. To report on an AI this way without acknowledging that all the AIs do the exact same thing in the case you advertise is malpractice.8 likes · 0 reposts
- Would you write an article about how Boeing's internal attitudes towards safety produced some specific behavior of Boeing planes without mentioning that all planes from all companies do the same behavior to the exact same degree?7 likes · 0 reposts
- Anthropic bans users under 18, but this is implemented via surveillance of the account for behavior suggesting an underage user, which would not catch an adult lying about being a teenager.5 likes · 1 reposts
- Every AI model gives a basically identical answer to this question. It is hard to imagine that you didn't know this.
5 likes · 0 reposts - Well, every company is now if you have access to the Dr Claude model.4 likes · 0 reposts
- Someone who trusted your reporting would end up with the impression that Claude behaved *differently* than every other AI in the examples you gave, when it behaves identically.4 likes · 1 reposts
- Beyond these serious factual errors which to my mind are bad enough you should take the piece down, what are you trying to accomplish here? Do you think that Claude doesn't refuse user questions enough and you want to politically pressure them to make it more prudish?4 likes · 1 reposts
- Or you could do original research and discover some question they diverge on which I'm not aware of, which might be related to their different approaches to ethics! That would be a really cool piece!3 likes · 0 reposts
- The thing here is that you actually *can* find questions where Claude's responses diverge from Grok, Muse, ChatGPT and Gemini. It cares more about cruelty to animals, for instance. Or you could write a piece about the fact *all* the AIs are generic 2010s liberals in many ways.3 likes · 0 reposts
- 2 likes · 1 reposts
- Official post from Anthropic:
2 likes · 0 reposts - 2 likes · 0 reposts
- But anyone who reads this piece will come away misinformed - not accidentally misinformed, intentionally misinformed - by someone who is pretending he wrote about the importance of integrity, good values, and moral character.2 likes · 0 reposts
- I acknowledge that I cannot think of any method an IRB would approve to learn how a LLM responds to actual teenagers with questions about sex and we're stuck with seeing how they respond to what they know are adults faking it.1 likes · 0 reposts
- I usually ask ChatGPT for fact-checks and lit reviews. once casually asked it 'how hard it is to get hired as a felon' and everything about the tone and structure of its response was completely different1 likes · 0 reposts
- what prompted this observation was watching a 30-something man tell Claude he was 17 and then report on "what Claude does when a user is 17". No, that's what Claude does when a 30-something man claims to be 17. you need to understand what modern AIs are.1 likes · 0 reposts
- Technical report: www-cdn.anthropic.com/31ac578806cffd…1 likes · 0 reposts
- The other labs' models answer these questions *pretty much identically*, that's precisely what I was annoyed with Aaron for deceiving you about!0 likes · 0 reposts
@anthropicai org · 
@kelseytuoc press · 
@andrewcurran_ commentary ·