Current thinking in AI 4 Life Sciences, and starting ‘Fables’
Thoughts following a weeklong workshop on deep learning for science, and a plan to explore the tractability of AI for the life sciences through a series of “Fables”.
This week: I just got back from the deep learning for science workshop at my alma mater, and have had my AI-for-fisheries paper accepted with minor revisions (link forthcoming). I’m glad that paper is getting published, though I’m unsatisfied that we didn’t include a transformer case study in it. A specific transformer application, as part of a mentored Master’s student’s work, is coming shortly.
Here are some current thoughts I have about AI in the earth & life sciences, and what I plan to do about them:
The “not enough data” argument regarding AI and ecology is probably more fragile than we think, if we are interested in building models that predict and generalize well. (This is a big if, discussed below). Andrew Gordon Wilson (CS @ NYU) shared a provocative talk this week saying that when we target simpler models with the excuse that we only have a little data, we are lying about our beliefs in the underlying data-gen process. Even for sparsely sampled ecological questions we are better off fitting a soft-simplified higher-order model. This theme was echoed in various ways by several speakers.
Aside from data limitations, the bridge between AI and ecology is fraught because the business case is unclear. Not all ecology problems are about prediction or generalization. We use a simple GAM not because we fear a wiggly fit, but because we sincerely want to understand how moon phase affects reef fish spawning, or where Grey wolves like to hang out, and maybe use that information to do something in the real world. A vector of weights doesn’t give us this. We more often build ecological models as learners we can learn from. The typical AI engineer appears surprised by this.
In other words, not all life scientists have the same answer to the question “what do we want to do with this model?”. Even for the few cases where we do have clear goals, like industrial fisheries management, it’s far more likely that AI approaches will be first deployed within subcomponents of the complex modeling and management framework1 (data collection, fitting to individual observations, report writing). In short, a singular model to eat the entire ecological enterprise is not easy to envision.
This creates a tension, wherein it’s simultaneously observed that:
- an entire scientific domain is heavily moated from a targeted, mighty AI-driven methods overhaul, and yet
- multiple workshop speakers insist that “all scientific work will be handled by the models [LLMs, possibly w/scaffolds] within the next few years; plan accordingly”.
Following (a) to its logical conclusion could lead to either doing nothing (“it’s not worth it”/“where do I even start?”) or doing so slowly (the old-school way, that is, one paper for one species/region/technique at a time). If we put more belief in (b), maybe we should gin up some ambitious, even hasty curiousity about what the bridge might look like.
So…what will I do?
An experiment. For the rest of 2026 I’m going to produce ~10 “Fables”.
A “Fable” here has a double entendre: I plan to follow the advice of an Anthropic employee who presented at the workshop and assume that a prompt to Claude’s Fable model should get me reasonably far in solving an important, verifiable scientific question. The second meaning is, of course, the classical meaning of a “fable” as a short story that produces a moral lesson. The lesson here will update my beliefs about the tractability of these claims, in a domain I know well.
Given that fisheries science is my home domain, meaning the domain I know enough to validate, most of the Fables will come from that area. I’ll try to work in some other classical life-sciences/ecology problems. I will document the prompts and compute used to complete each.
Cover image: a Soviet Union stamp of a fish-related fable. Source
Footnotes
On the surface, the American enterprise of scientific fisheries management (stock assessment) has a singular goal: predict next year’s catch, and ensure we can keep catching. Conservation goals might be explicit or simply incidental to this goal. However, the real, bread-and-butter way that industrial fisheries are managed in the U.S. involves a complex series of top-down control rules, reference points, human review panels, and required report formats. This is a techno-social constraint. There is the easier route of using AI to step through the assessment hoops better or faster, but the shape of those solutions will be limited by the policy environment in which the stock assessment happens. For example, our paper shows a simple LSTM fitting growth curves better than other methods; a transformer could probably outperform on the non-stationary regime. This would materially improve forecasts, and frankly should be included in the projections process. But there is no current management framework that would accept the outputs of an RL policy for quota forecasting.↩︎