[Ey7NFZ3P0nzAe]

> They're not being trained autonomously with RL and a robotic wetlab. When that's a thing I'll start to take claims of biology risk more seriously

So if IIUC your point is "they're not good enough at biology right now because they're not trained on it so they're not a threat".

To which I want to answer: "they're not a threat now but I see *no* reason for models not to be trained on biology pretty darn soon unless people like you convince the world otherwise."

Thoughts?

[rdedev]

They are already being trained in biology right now? What he is saying is that they are being trained on what we know about biology currently. AI is going to hit that limit. And there is no way for the AI to gain more knowledge without actually doing lab experiments

[inciampati]

My point is more about conflict of interest in Anthropic, and certain techpeople types thinking that biology is a useful scapegoat because biologists don't use or understand this tech (they claimed 0.03% of users would be affected by biotech security stuff). So, arguing the world will fall apart because someone can ask your superduper powerful tool for a bad bad virus sequence, or how to do some technique in the lab, and get an answer that's plausible (oh but you never bothered to actually test if it's legit because you don't have the capability to do so).

As for the future... today the LLMs are "trained on biology", in that they read the textbooks, the research, the web.

They aren't trained on biology in the sense of being embodied, autonomously or semi-autonomously driving actual biological experiments. If you come from software, the timescale of these experiments is outlandish. Yes, I am partly saying the LLMs are not good enough today because they need to be embodied and trained for literal decades of lab time before there is even the _remote_ possibility that they could present a novel risk profile that is even a shadow of what the current fearmongering suggests the current models can enable.

And they aren't trained on biology in the sense that they've read the literature, but even 100T token training run only begins to touch the data scales that rather mundane bioinformatics operate at. True multimodal models that work on DNA and human language at high quality haven't yet emerged. We're talking new architectures which are going to arise after the next AI winter.

All of this ignores an even more fundamental point. Cost. If someone wants to make a bionuke, they don't need to use AI. They can set up the right evolutionary context and run quadrillions of parallel explorations of the design space. Directed evolution like this is cheap, well-understood, and insanely powerful. If you actually care about biosafety, we should be doing hard work to surveil gain of function research. Different flavors of LLM use are not going to be a differentiator for the foreseeable future.