Spooky, and sometimes brilliant, ideas from AI in science
This article from @Anil Ananthaswamy for @quanta magazine (my favorite science magazine!) delivers a concise description of my own experience of using AI for scientific discovery:
- Hallucinations are a feature, not a bug: LLMs can and do make things up, which in many situations is an impediment. But in science, when trying to expand your search space, coming up with crazy ideas can be the goal and things or concepts that don’t (yet) exist can be the point.
- LLMs can take you out of your comfort zone: when looking for new avenues, you are running away from “cognitive availability”, a heuristic or bias that makes you stick to what is familiar, easy to make sense of, things that you know. AI can generate ideas that are unfamiliar, hard to decipher at first, borrowing from fields that are alien to you.
- And by combining ideas, concepts and building blocks that have never been assembled in this way before, AI can generate conceptual novelty -it does not always stand up to scrutiny but when it does it can be deeply satisfying. Yes, it may be stochastically recombining or “recycling” existing stuff, but a lot of innovation comes from combining the existing into something new.
I have been using various models to test hypotheses or look for missing links and I have been blown away by the dots I was able to connect: it is as simple as asking: “what are possible links between X and Y?”. Even if no direct link exists in the literature, some models can find common nodes in the network of concepts, sometimes leading to an unexpected mechanism, confounder, common cause or effect. It has happened to me multiple times in ways that took me down productive rabbit holes and forced me to reconsider hypotheses.
Unsurprisingly, the one thing that is needed for this process to work is the ability to recognize what is interesting and plausible.