What has Anthropic accomplished in biology so far? Not a lot.

hey im faye ♡ based in sf, looking for work to build my portfolio! ⟡ 3m views in 1 day ⟡ content creator (75k ) ⟡ prev hypebeast ⟡ led classical music nonprofit ($10k in grants and scholarships) ⟡ photographer & videographer shoot me a dm or send referrals

SITUATION DETECTED: Sam Altman, Dario Amodei, and Demis Hassabis have signed a joint open letter calling on Congress to mandate screening of synthetic nucleic acid orders, citing AI’s rapidly improving ability to assist with biological research as an urgent biosecurity risk.

Anthropic rejected me as well. They asked me to solve a problem that was obviously meant to be solved via mech interp. I said you clearly want me to use mech interp but I don’t believe in mech interp, would you like me to tell you how I’d solve this? They said, no, tell us how you’d solve it if you did believe in mech interp. Anyway I guess they didn’t like my answer lol

Biology already had more ideas than it could test in the lab, and the development of new AI models will only further stress that bottleneck. Since 2023, our work at Tetsuwan has been focused on fixing to the biggest problem in lab automation so that we can use it to blow that bottleneck open. Lab automation will be necessary to turn in silico ideas into in vitro results, but today, lab automation is prohibitively costly for most workflows. Moving water from point A to point B could be an instruction that contains as many as 25 instructions (liquid classes are hard...). Formalizing experiments and removing all of the tacit & implicit details that surround them is a process that is far too slow & expensive to be worth it for the vast majority of experiments. As @owl_posting writes, "Most experiments can be automated, but are not worth doing so." We realized this back in 2023. Lab automation does not work for most experiments, and this fact would serve as a massive obstacle to attempts to automate biology research. So, we created a standard language for wet lab experimentation (technically two). We call them the Procedure Description Language (PDL) & the Variable Description Language (VDL). PDL & VDL give us a common way to specify experiments and their context. Once an experiment is expressed in a concrete syntax, our compiler, Ariadne, turns that specification into a set of instructions for the robot. But users don't need to know this language or do any programming themselves to move their experiments onto automated platforms. Our software platform, ResearchOS, abstracts for lab automation knowledge, allowing researchers to easily & rapidly configure their experiments for automation- all they need to get started is a PDF. Our team has spent the last two years iterating with pilot partners to develop this new way to communicate with lab robots. In the past few months, we've shared our progress on ResearchOS (read more on our blog!), including: - Agentic experimentation... letting agents run their own experiments using lab robotics - Off-deck module support, including the integration of 6DoF arms - Support for driver libraries, allowing cross-platform code generation Autonomous science does not work unless lab automation works, and the biggest constraint on wet lab automation today is how we approach automation engineering. ResearchOS is the answer, and Tetsuwan's first step in building an autonomous lab that any researcher or agent can easily send their experiments to, and get the resulting data in return. Of cloud labs, Armer, Letronne, & DeBenedictis (2023) write, "The cost to enter is high (>$250k for general access to Emerald Cloud Lab, or >$100k to automate and run a single method at Strateos), and the contract lengths are long (one year minimum". Autonomous labs do not work unless the automation engineering works, otherwise the unit economics are horrendous. We've made the automation engineering work. In the future, you don't need your own research lab to experiment, just an internet connection.

Remember the security firm that Ubuntu hired to audit the (ill-advised, highly buggy) Rust-rewrites of all of the GNU Coreutils? Turns out that security firm is run by @gf_256, who: - Appears to be a man who thinks he's a woman ("trans"). - Uses an anime cartoon of a girl as his avatar. - Appears to have an OnlyFans page. I repeat: Ubuntu hired a "Trans" man, with an anime girl avatar and an OnlyFans page... to audit Rust code. It's hard to get more on-the-nose than that.

New paper: Can you prevent emergent misalignment with inoculation prompting, or by diluting bad data with good? Prior work suggests you can. 
We show the misalignment is still present but hiding. It is triggered by adding cues to prompts, evoking the bad data.