Skip to content
Multigrid

AI in science & discovery

Where machine learning has genuinely changed how research gets done, described by the method and by the verification step that still has to happen.

There is a version of this subject that is a list of breakthroughs, and it is almost useless. Breakthrough lists go stale, they compress a prediction and an experiment into one sentence, and they give a reader no way to judge the next claim they encounter. This cluster is built the other way round: each page describes a problem that was intractable, the thing a model actually computes, and who has to check the answer before it counts as knowledge.

That last part is the load-bearing one. A model does not confirm anything. It produces a hypothesis — a structure, a candidate compound, a forecast, a classification — and somewhere downstream a crystallographer, a chemist, a clinician or a proof checker either accepts it or does not. Where that downstream step is fast and cheap, machine learning has transformed a field. Where it is slow and expensive, the field looks much the same as it did, no matter how good the predictions have become. Most of what is confusing about AI in science is explained by that one ratio.

AI in Scientific Research: How to Tell Where It Is Actually Working

An evidence ladder for AI-in-science claims, from a model output to a confirmed result, and the reason some fields have climbed it and others have not.

4 min read

Protein Structure Prediction: What It Solved and What It Did Not

How structure prediction works, why evolutionary covariation is the signal it runs on, and the specific questions a predicted structure still cannot answer.

5 min read

AI in Drug Discovery: What a Model Can Move and What It Cannot

Where in the drug pipeline machine learning actually acts, why compressing the early stages barely moves the overall odds, and how to read a milestone announcement.

5 min read

Materials Discovery With Machine Learning: Predicted Is Not Made

What a materials model actually predicts, why its training labels are simulations rather than experiments, and the rungs between a stable structure and a real compound.

4 min read

AI Weather Forecasting: Where Learned Models Beat the Equations

How a learned forecast model differs from a physical one, why it wins on cost and headline scores, and why the loss function is what limits it.

5 min read

AI in Mathematics: Why Formal Verification Is the Interesting Part

Why wrapping a language model in a proof assistant changes its failure mode completely, and what that buys — and does not buy — for research mathematics.

5 min read

Literature Review With AI: A Workflow With Verification Built In

Splitting a literature review into search, screening and extraction, and putting the right check on each so a model never becomes the source of a claim.

5 min read

AI in Genomics: Sequence Models Applied Outside Language

Why self-supervised sequence modelling transfers to DNA and protein, what the resulting scores actually mean, and why a variant score is not a diagnosis.

4 min read

Surrogate Models: Replacing an Expensive Simulation With a Learned One

When it pays to train a model on a simulator's outputs, how to choose the points you train on, and why the answer must always go back to the simulator before it counts.

5 min read

AI in Astronomy: Classification at Scales Humans Cannot Reach

How classifiers triage survey data streams, why the selection function is the part that decides whether a result is publishable, and where simulation-based inference fits.

5 min read

Self-Driving Labs: What It Takes to Actually Close the Loop

The four components an autonomous experimental loop needs, why characterisation rather than synthesis is the bottleneck, and how to tell how much of a loop is really closed.

4 min read

Reproducibility of AI-Assisted Science: The New Failure Modes

A taxonomy of the leakage and versioning problems machine learning introduces into a method section, each with the check that catches it.

5 min read

AI-Generated Papers and Journal Integrity

Why detectors are the wrong instrument for this problem, what the artefacts actually look like, and where publisher policy has settled.

5 min read

Peer Review With AI Assistance: Confidentiality Comes First

Why the objection to using these tools in review is about disclosure of an unpublished manuscript rather than about output quality, and where assistance is uncontroversial.

4 min read

Data Analysis With LLMs: Where It Breaks

The three separate places a model-assisted analysis can be wrong, the statistical errors it makes fluently, and a workflow that makes each of them visible.

5 min read

Hypothesis Generation: Can a Model Have an Idea?

What novelty would have to mean before the question is answerable, the two mechanisms that could produce it, and the study design that would settle it.

4 min read

AI in Medical Diagnosis: The Evidence Ladder a Model Has to Climb

Why a strong retrospective score says little about clinical benefit, what changes between a dataset and a clinic, and what each rung of evidence actually establishes.

5 min read

Benchmarks for Scientific Reasoning: What a Score Establishes

How expert-written, search-resistant question sets are constructed, what a high score does establish, and the four gaps between answering questions and doing science.

4 min read

Open Science and Open Models: Reproducibility Needs Weights

Why a study whose method is a call to a hosted model is not archivable, what openness actually has to mean for reproduction, and how to handle the trade-off honestly.

4 min read

Will AI Compress the Research Cycle?

The strongest version of the acceleration argument, the arithmetic that bounds it, and the specific bottlenecks that no amount of prediction quality touches.

5 min read

Other topics

AI in science & discovery · Multigrid