when you say automated I assume here you mean autonomous/ self-driven? We already have automated labs (i.e. We automate the workflows instead of doing them manually) - pharma has been developing them for a while but other industries are catching up just cause they’re so much more efficient at screening different conditions which would literally take weeks or months to humans to do by hand. Although you don’t need agents to run them, using automation doesn’t mean the task becomes “easy” - you still need to figure out a proof-of-concept experiment that you want to automate, this scaling is almost never linear, you get reproducibility issues, need to calibrate the robotic arms, in case of negative or volatile results you need to think about whether it’s the infrastructure, experiment, or something else, to name a fee issues (+ usual wet lab limitations). This is also in the context of having access to one of these robots, which more and more labs do, but it’s still not a trivial tool.
I think when you’re adding agents to this mix you’re adding even more of these variables that you need to account for (of course at the tradeoff that you are running even more experiments, in an even shorter span of time). People underestimate how sensitive scientific discovery is to the environment it’s undertaken in. It’a not like if you give your robotic arms more calibration, then all your experiments will magically work with high yields/ purity etc. (not saying that’s what you suggested, I just find this a very harmful myth introduced by people working in SDLs that never did any wet lab research in their life). There is already ongoing work in testing current capabilities of SDLs (see work from Ginkgo Bioworks, Emerald Labs, top of my head - but many academic labs published similar works in the past year). Most of them will use very specific use cases - that’s because having an autonomous lab do EVERYTHING for you would be simply too expensive and wouldn’t make sense on infra terms to establish (at least at the moment).
Back to your post: I think the current space in SDLs is still evolving and is heavily reliant on humans for enforcing the safeguards and right evals (the whole point of AI safety after all). I don’t think we’ll move away from that any time soon, and I would argue that regulation will play a big part. I think we need more people to take a bunch of wet lab experiments and translate them to automated/ autonomous workflows in order to better understand how this entirely new workflow really works (not sure it will happen and it would be expensive - one can dream). Otherwise, the agentic/ robotic capabilities still need to be improved and finetuned in order to access more experimental spaces - but needs to happen in a controlled way to avoid the risk you are describing. So assuming you are coming from a technical angle, I would say go for it. SDLs are not LLMs, and although there are more risks when leaving these agents on their own, there are also more safeguards that (hopefully) people can add in place.
I’m considering whether to work on accelerating automated scientific labs.
My main concern: if the time until the next future AI escape is roughly fixed (say like, hypothetically, 6 months), then faster progress on automated labs means there will be more scientific infrastructure available for that AI to exploit when it escapes, thereby increasing the chance of catastrophic risk.
Is that reasoning sound? Many people are still willing to accelerate automated labs, so I’m wondering if I’m missing something.
A few responses I’ve encountered, but am not convinced by:
TL;DR: Are there compelling reasons to reject the argument that we should avoid accelerating automated labs because doing so gives a future escaped AI more infrastructure to exploit, thereby increasing catastrophic risk? If there aren’t any, though, happy to not work on progressing automated labs.