AI FOR SCIENCE · TIMELINE
The 14 entries the timeline holds for 2023, newest first. Each was read at source. The full timeline, with filters →
What happened: A GPT-4-based multi-agent system searches documentation, plans reactions, writes control code and operates a cloud laboratory.
Why it matters: The LLM becomes an orchestrator for a real experimental workflow rather than only a scientific text interface.
What happened: The system evolves LLM-generated programs against an automated evaluator and reports new cap-set constructions alongside improved bin-packing heuristics.
Why it matters: A reusable discovery pattern appears: models can search broadly when incorrect ideas are rejected cheaply and objectively by execution.
What happened: Raphael Townshend's team trained the model on in-house chemical-mapping data, which the abstract puts at millions of RNA sequences. Small probe networks on its embeddings reach state-of-the-art accuracy on RNA prediction tasks after a little extra data.
Why it matters: A startup measuring its own training data is the model's real edge. The abstract names no benchmark, so how far ahead it is stays unchecked, and nothing here is a designed RNA tested in a lab.
What happened: DeepMind reports 2.2 million candidate crystal structures, including 380,000 added to a set judged most stable.
Why it matters: Materials screening reaches industrial scale while sharpening debate over what should count as a discovered material.
Evidence and limits: Anthony Cheetham and Ram Seshadri (UC Santa Barbara) examined a random sample of the predictions in Chemistry of Materials in 2024 and found "scant evidence" of compounds that are at once new, credible and useful. Many predictions put similar ions on separate sites, an order that the high temperatures of real synthesis would scramble.
What happened: The autonomous lab runs 353 experiments in 17 days and realizes 36 of 57 target inorganic materials.
Why it matters: Prediction, literature-derived recipes, robotics, measurement and active learning operate as one physical loop.
Evidence and limits: Josh Leeman, Leslie Schoop, Robert Palgrave and colleagues re-examined all 43 products in PRX Energy in 2024 and concluded that no new materials had been made: the automated fitting of the X-ray patterns was unreliable, and the predictions often ignored disorder, where elements share sites. The authors' 2026 correction drops "novel" and confirms 36 of 40 syntheses.
What happened: John B. Ingraham, Gevorg Grigoryan and colleagues, all at the company, describe a generator steered by symmetry, a target shape, a protein class or a text prompt. They made 268 unconditioned and 42 conditioned designs in E. coli without first filtering them by structure prediction. All 172 of the first unconditioned set scored above the negative control in a pooled solubility screen, and western blots confirmed 19 of the top 20. Two crystal structures matched their designs to about 1 Å. The code is Apache 2.0.
Why it matters: The record holds Generate's IPO, its Phase 3 asthma antibody and its deals, but not its main published method. Chroma puts the design conditions into the sampling itself, so one model takes a symmetry, a shape or a text prompt where otherwise each would need its own tool, and its code is open for others to test.
Evidence and limits: What was measured is that designs express, fold and match their models; no design was tested for a function. Every author works for the company, seven designs were purified and two crystallised.
What happened: The graph neural network outperforms ECMWF HRES on more than 90% of 1,380 verification targets and produces a ten-day forecast in under a minute.
Why it matters: Learned forecasting proves competitive under a mature, continuously reality-checked evaluation regime.
What happened: Giulia Zavagno, Christopher Saunter and David Weinkove of Durham and Magnitude describe WormGazer: fixed cameras imaging many Petri dishes at once, moving nothing, scoring how many worms move and how fast. Beside a manual lifespan assay, 390 worms per condition per method, seven days of imaging showed all three sulfamethoxazole doses beating control on healthspan; the lifespan arm took 40 days to reach mean survivals of 20, 21 and 21 days against 18. The age-1 mutant spent 93.5% more time moving than wild type over days 2 to 7.
Why it matters: This is the published method behind a company the funding board describes as reading worms out with machine learning, and the paper has none: movement is found by thresholding which pixels changed between frames, with a 10 um/s cut-off to drop noise. What it buys is time (a healthspan readout in a week instead of a survival curve in six), and the comparison was run by the authors, two of whom are shareholders and directors of the company.
What happened: The Schmidt-backed nonprofit begins building scientific agents, initially focused on biology and scientific literature.
Why it matters: AI-for-science agent development becomes the mission of a dedicated research organisation rather than a side project.
What happened: The model classifies most possible missense variants as likely benign or pathogenic, vastly expanding the small experimentally interpreted set.
Why it matters: AlphaFold-derived representations move from structure prediction into genome interpretation and disease research.
What happened: A diffusion model generates novel protein backbones, with hundreds of designs tested experimentally.
Why it matters: The central question shifts from “what shape does this sequence have?” to “what sequence and structure should exist?”
What happened: Huawei's 3D neural network reports stronger deterministic forecasts than ECMWF's operational system while running more than 10,000 times faster.
Why it matters: A mature scientific field with strict operational evaluation becomes one of AI4S's clearest tests outside biology.
What happened: Reinforcement learning finds faster low-level sorting routines that are added to LLVM's libc++ implementation.
Why it matters: An AI-discovered algorithm moves from a paper into infrastructure used by developers at scale.
What happened: Demis Hassabis takes charge of the combined unit; Jeff Dean becomes Google's Chief Scientist.
Why it matters: Talent, compute and model development are concentrated inside one organisation that also houses major AI4S programs.