Sure, there’s been lots of discussion about AI taking jobs, but let’s face it: until your job shows up on the list, you are unlikely to get that excited. Well, yesterday, my job — being an academic researcher or, as I like to term it, `scientist’ — was added to the list in a big way.
This was in the form of a new paper, “The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery.”
We introduce The AI Scientist, which generates novel research ideas, writes code, executes experiments, visualizes results, describes its findings by writing a full scientific paper, and then runs a simulated review process for evaluation. In principle, this process can be repeated to iteratively develop ideas in an open-ended fashion, acting like the human scientific community.
Well, hang on just a minute. That’s my day job. But it is surely OK. The amount of computing power that is required to do what I can do for a can of Coke and a burrito is surely prohibitive.
We demonstrate its versatility by applying it to three distinct subfields of machine learning: diffusion modeling, transformer-based language modeling, and learning dynamics. Each idea is implemented and developed into a full paper at a cost of less than $15 per paper.
Oh crap. With inflation, my lunch costs more than that.
Now, it is time to get serious about what to do about this AI thing. Let’s go through the stages, shall we?
Denial: this was a paper just about machine learning. That’s like computer stuff. Sure, that can be automated. They were hardly thinking anyway and might be considered mere engineering. But I deal with trying to understand the real world. No AI is going to be able to do that.
Anger: Who do these people think they are? Just letting computers write papers and submit to journals and fool people into thinking it is the work of real people when it is just code. They denigrate the real hard work of real hard scientists who took really hard decades to gain the skills necessary to push the real hard frontier of real hard knowledge.
Bargaining: Look. Maybe what we should agree to do is this. Yes, you can develop AI Scientists, but they have to work with a real human scientist as a co-author before anything gets submitted. This is just to make sure everything is on the up and up.
Depression: Why do I bother? I struggle to come up with ideas, craft them into research questions, develop arguments, assemble evidence, write dispassionately in a paper and run the gauntlet of the academic publisher process, and all some teenager in a basement has to do is press a button and can do the same thing for the School Science Fair.
Acceptance: You know what, if it’s that easy to generate scientific knowledge, I guess we don’t have to do it anymore. Who even needs to read these papers and have them in libraries and such? We can just ask a question, and the AI will answer whether it exists in the scientific literature or not. The whole system can be dismantled.
That last bit is the actual point I wanted to make here. It is interesting to consider what happens if the AI Scientist becomes a thing and really does what it promises across many scientific fields. The answer is that it really does make the existing system look pointless.
This was something we envisaged in Prediction Machines when we talked about online retailing moving from shop-then-ship to ship-then-shop with AI implying that retailers could just send you stuff when they predicted you needed it rather than you shopping or searching for things and then having them sent to you.
The same would occur for science. The only reason we do what we do now — discover-then-find — is because it takes so long to do the research that basically, we scientists are guessing what knowledge people might need and working on it so that when they need it, it is there for them to find. An AI Scientist would upend that and move to find-then-discover. Because it is so quick to answer a question, science would no longer need to answer questions in advance. Instead of asking, “is there a paper that tells me how X works?” people would ask, “tell me how X works” and receive a paper with that information. Then, I would write the result in a friendly substack. Problem solved.