Becoming a Super-Scientist

I am a scientist with lots of ideas, all the time. But I make a lot of calculation errors, and I am a slow coder. Most of my ideas go untested.

On the flip side, I can easily and quickly verify whether a calculation is correct, an equation makes sense or a piece of code does what it is supposed to. In other words, I am great at critiquing (in keeping with my French origins, I guess) but mediocre at producing.

Until recently, these evaluative qualities were not that useful, except when assessing a student’s work. Using students as generators (a sort of generative human intelligence) can be a clever strategy, but only in proportion to your students’ cleverness. I was always blessed with great students but depended on them to test the whole range of weird ideas I had all the time. Adrian Goedeckemeyer, Roujia Wen, Ian Van Buskirk, Xiaofan Liang, among others, have the scars to prove it.

But the latest iterations of LLMs, GPT 5, Claude Sonnet 4.5, Future House, changed everything: I know which questions to ask (all these weird ideas to test) and I can curate the results, test and run code, all in a few minutes or perhaps one hour. One year ago, that would have taken me days, weeks, or more. Here are some examples of what I was able to do in the last 6 weeks:

1️⃣ Develop the mathematical framework, simulate and test a non-ergodic theory of population aging.

2️⃣ Prove that certain evolutionary dynamics on continuous trait spaces are information-geometric flows when the fitness function satisfies some “conjugate” property with respect to an underlying statistical manifold.

3️⃣ Create low-dimensional hyperbolic embeddings of very high-dimensional datasets using a novel approach.

4️⃣ Create, simulate and test a mathematical health outcomes and economics model of chronic inflammation blockade to understand the conditions under which a public health intervention could make sense.

5️⃣ Complete a self-propelled particles model exploration project that had been in limbo for 5 years.

6️⃣ And tens of other projects, some mini, some bigger, most terminated because they were clearly not good ideas. I simulate everything.

These growing superpowers are and will be available to anyone. But I believe that creative scientists have an edge: they know how to frame questions and they know what looks good. They will produce the opposite of AI slop: high-quality research augmented by AI. The problem will be to create the right infrastructure with the right incentives to select and disseminate such research: the increase in AI slop my vastly outrun the increase in quality work and the current publication-industrial complex is not the solution.

Kevin Weil, Mike Krieger, Anastasia Borovykh, Andrew White, Danilo Jimenez Rezende, Vivek Natarajan, Sam De Brouwer, Sebastian Uribe