The Signal Report // AI IN THE LAB
AI in the Lab.
Where machine learning genuinely helps research, and where 'AI discovered X' is doing a lot of the lifting. All reports →
31 in this beat · page 1 of 4
What the AI X-ray scientist actually did
An "AI X-ray scientist" ran an experiment at SLAC, and the coverage says self-driving labs are next. The Nature Machine Intelligence paper shows something narrower and still useful: an off-the-shelf model aligning one crystal from a detailed written procedure, with a human relaying every command, in sessions the authors say were not systematically repeated.
Read the report → Report 176 · AI in the Lab · 2026-10-04 · 7 minWhat 37,000 AI agents actually proposed
A Stanford "virtual biotech" of AI agents, published in Science, proposed a B7-H3 antibody-drug conjugate for lung cancer, and the coverage said the swarm found a drug. A B7-H3 ADC has been in human trials since November 2019. The 37,000 agents were mostly reading trial records. What the AI really added is a fibroblast hypothesis nobody has tested at a bench.
Read the report → Report 170 · AI in the Lab · 2026-10-01 · 8 minWhat 75% accuracy meant in AI mind-reading
"AI recreates imagined images with over 75% accuracy" was a two-choice test where a coin scores 50%, graded in the same feature space the algorithm was tuned to match. A 2025 reanalysis of the released code found the showcased images came from one participant and its "Bayesian" step barely changes the output.
Read the report → Report 165 · AI in the Lab · 2026-09-27 · 11 minWhat Claude actually found in phage DNA
Coverage asked whether AI had found the next CRISPR. Anthropic's own 40-page technical report describes agents that noticed an unannotated repeat array beside a phage enzyme, human lab work showing the array is expressed as short RNAs, no cas genes, a function the authors call unknown, and ten reruns that missed the array every time.
Read the report → Report 159 · AI in the Lab · 2026-09-23 · 10 minThe model doesn't know it was retracted
Asked whether 161 of the loudest retracted papers in science had been retracted, three open-weight models said no 82 to 88 percent of the time. Run over 34,070 valid papers they rarely cried wolf, but one of them called clear methods errors in 97 to 100 percent of everything it read.
Read the report → Report 152 · AI in the Lab · 2026-09-20 · 10 minThe benchmark nobody outside can run
Five drug firms fine-tuned an open protein-folding model on 20,167 private crystal structures and reported beating every public model tested. The training data, the test set and the model are all private, so nobody outside can reproduce it. The consortium says as much, in a limitations section the coverage did not carry.
Read the report → Report 146 · AI in the Lab · 2026-09-14 · 8 minWhat the AI scientist actually found for AMD
Robin, FutureHouse's multi-agent AI, proposed the glaucoma drug ripasudil for dry macular degeneration, and the work is now in Nature. The evidence is a 1.89-fold rise in bead uptake by cultured retinal cells, three wells per condition, under a protocol humans wrote and changed. What the AI did, what people did, and the disease model and randomized trial the paper says are still ahead.
Read the report → Report 141 · AI in the Lab · 2026-09-11 · 12 minDid AI solve Navier-Stokes?
The Clay Mathematics Institute's official description of the problem is four statements, and two of them allow an external force while two forbid it. The three preprints you can actually download prove forced blowup for the porous medium equation, for Boussinesq and for Euler, and the same official document says Euler is not on the prize list. Meanwhile the rules require journal publication plus a two-year waiting period, so the answer to the headline is a calendar, not an opinion.
Read the report → Report 136 · AI in the Lab · 2026-09-08 · 9 minAI predicted the crystal. Not the disorder.
Computational materials discovery proposes tidy crystals in which every site holds one element. A third of the experimentally known inorganic record is not like that. A classifier trained on the ICSD puts a number on how much of two big predicted databases would come out of a furnace disordered, and attaches a warning to the number that travelled.
Read the report → Report 130 · AI in the Lab · 2026-09-05 · 10 minThe force field that passed the benchmark
Machine-learned force fields get reported by their error on a held-out test set. The TEA Challenge 2023 took five of them, gave them identical data and identical hardware, ran twelve million-step simulations each, and counted how many survived. The ranking on the metric and the ranking on survival were not the same ranking, and the reason is a cutoff radius.
Read the report →