Lila says its platform designed an in vivo CAR-T mRNA that beat a benchmark in monkeys; its conference abstract and its blog give different figures
What happened: In an ASGCT 2026 abstract (12 authors, all at Lila), one 0.5 mg/kg dose in monkeys gave near-complete B-cell depletion through day 7, and in mice 0.25 mg/kg beat the benchmark at 1.0 mg/kg with no significant toxicity. Lila's 2026-08-28 blog post says the therapy outperformed the leading benchmark, took three scientists under six months and cost about a tenth of a conventional programme. All of it is preclinical and company-reported.
Why it matters: The two Lila sources do not agree. Time: three months in the abstract, under six in the blog. Day 14: B cells back to about 50% of baseline in the abstract, depletion maintained in the blog. Benchmark: back to baseline in the abstract, past it in the blog. Neither names the benchmark or gives animal numbers or statistics, so the claim to beat the leading therapy cannot yet be checked.
Anthropic previews a common interface for agents and lab hardware
What happened: The Model Hardware Standard proposes a shared layer for instrument control, state, measurements and errors. In a Genentech proof of concept, Claude drove a liquid handler, robot arm and plate reader through it and tuned pipetting flow rates against an expert's transfers.
Why it matters: A shared interface could turn one-off instrument integrations into reusable infrastructure, a prerequisite for scaling autonomous labs.
Co-Scientist crosses from hypotheses into physical experiments
What happened: An execution-grounded extension connects Gemini agents to real experiments, including CVD protocols for two-dimensional materials and wet-lab validation of biological predictions.
Why it matters: The important shift is from proposing hypotheses in silico to adapting research plans to laboratory constraints, execution and empirical feedback.
Tetsuwan Scientific runs 9,143 robot dispenses chosen with Claude and learns a pipetting model that predicts precision about 12% better than the maker's datasheet
What happened: Using Anthropic's Model Hardware Standard on an Opentrons Flex and a Byonoy plate reader, Claude proposed batches of a dye tracer, 1,508 measured conditions in all. On held-out sessions the learned model beat the datasheet on 33 of 45, about 12% better (sign test p about 0.003), and about 17% on the most replicated data. About a third of the variation turned out to be evaporation at the plate edges.
Why it matters: A lab robot that knows its own error can plan a protocol a scientist can trust, which is what a cloud lab sells. The gain is modest, measured on one robot with one tracer dye in the company's own lab and reported in a blog post, and the company says most of the effects it expected to find did not show up.
StoneWise and Peking University ran a generative model that learns from each round of wet-lab results through three drug-optimisation campaigns
What happened: Across three targets the team made 13 compounds each, feeding every round's measurements, failures included, into the reward of a structure-guided 3D generative model. On ROR1, three cycles took oral bioavailability from 3.51% to 35.64%, about ten times, with potency held. On NLRP3, the concentration needed to inhibit the liver enzyme CYP2C19 rose 20 times, from 1.59 to 30 micromolar. On NSD3, a hit bound about 18 times more tightly. Retrospectively on six tasks, REINVENT4 recovered no known active analogue; RCDO found at least one on every scaffold.
Why it matters: A generative model usually proposes molecules once and is judged on how good the list looks. This one is rebuilt after each round of synthesis and assay, and the rounds it learns from are the ones a medicinal chemistry team runs anyway, so the compounds that failed on potency or on drug-like properties count as training signal rather than waste. StoneWise sells molecule-design software to drugmakers, and the campaigns it reports are the kind its customers run.
Evidence and limits: Every campaign is the authors' own, with no outside team repeating one and no arm where chemists optimised the same series without the model; the gains are against the parent compounds, and the paper puts no number on the time saved. The abstract's 40-fold oral exposure is not in the text, which gives about ten times on bioavailability and leaves the exposure figures to a supplementary table, unread. The code is promised on acceptance of this preprint.
Diogo Machado had GPT-4o pick among 139 Austrian grant proposals; it was less likely than the funder to back those reviewers rated novel
What happened: Diogo Machado, of the consultancy Technopolis and KU Leuven, gave GPT-4o all 139 single-field 2021 Austrian Science Fund (FWF) proposals in four sciences, two at a time within a field, asking only which to fund. Funding as many per field as FWF did (51 in all), the model matched the human decision on 87. Of the four criteria FWF's reviewers scored, only novelty predicted disagreement (odds ratio 4.80), and one way: the more novel a proposal, the likelier people funded it and the model did not.
Why it matters: Funders are starting to try AI in grant assessment: UKRI says it will compare AI evaluations with past funding decisions, and an EPSRC-funded network has let models reject proposals unread. Here agreement looked respectable while the disagreements fell on the proposals reviewers found most novel, so agreement with past decisions is not enough of a test. Machado argues AI-assisted review should also be checked against what funded work went on to produce.
Evidence and limits: One funder, one year, one model at temperature zero, and a prompt that asked only which proposal to fund, without FWF's criteria. The paper calls the sample small, did not test other prompts or ways of pairing, and does not identify why the model differs. Novelty is as FWF's reviewers scored it.
Google moves its AI Responsibility team out of DeepMind
What happened: The roughly 90-person unit is moving into Google Global Affairs at the start of September, covering frontier-model risks, CBRN evaluations, user behavior and the psychological effects of chatbots.
Why it matters: Not a layoff but a governance experiment: does moving frontier-risk research closer to policy strengthen coordination, or weaken its independence? Google says mission, staffing, compute and DeepMind access remain.
BixBench3: on 20 computational biology tasks run from raw data, 13 frontier models score 0.00 to 0.48
What happened: Edison Scientific's benchmark gives an agent a study's research question, its analysis plan spelled out, and the raw data, then grades the 138 outputs it produces, such as differential expression tables, against those the original study reported. Scores fall from 0.36 to 0.10 when raw data passes 100 GB, and from 0.36 to 0.24 at three or more steps. A task averages 6.8 hours and $43; the top scorers were not the biggest spenders.
Why it matters: It measures whether an agent can carry out a whole study's analysis from raw data, and says where agents break: large data and long chains of steps. The agent runs a prescribed pipeline and never chooses the analyses; scores rank models against published results, on 20 studies, in the company's own benchmark.
Transfyr launches with a $25M seed led by General Catalyst to record how lab work is actually done; ARPA-H's founding director is a co-founder
What happened: General Catalyst led; Lux Capital, Breakout Ventures, Factory, Neo, SV Angel, MVP Ventures, Underscore VC and Lyda Hill joined. CEO Anna Marie Wagner led AI and corporate development at Ginkgo Bioworks; Renee Wegrzyn, its Chief Innovation Officer, was ARPA-H's founding director. The Cambridge, Massachusetts company puts sensors and multimodal models into labs to capture what scientists do, the conditions and the instrument readings, and runs its own wet lab for training data. Hemant Taneja, General Catalyst's CEO, spoke for the round; Kevin Weil and David Baker are named advisers.
Why it matters: The bet is that what AI and lab robots lack is not more papers but the unwritten part of an experiment: the steps that work only in one person's hands. The release says Transfyr already runs evaluations for the largest frontier AI labs, without naming them, and shows no model or measured gain. The people are the news: the first director of ARPA-H, the US government's health moonshot agency, now builds lab-data infrastructure, with OpenAI's former head of science as an adviser.
MIT trains a crystal-and-text model on AI-written descriptions of 126,000 crystals; trained on sentences from papers instead, its property search does no better than chance
What happened: Yang Jeong Park, Elsa Olivetti, Ju Li and colleagues pair 126,000 Materials Project crystals with descriptions a GPT model wrote from each one's computed properties, and train a crystal encoder and a text encoder to match them. Asked for a property in words, it retrieves crystals whose computed values follow the request; trained instead on sentences from 4 million papers, it retrieves at about random. In multiple-choice tests it names composition 95.4% of the time but crystal system only 34.0%, level with a small text-only model.
Why it matters: Searching materials by a sentence needs text tied to structure, and the literature does not supply it: the paper text here covered about 400 compounds against 126,000, and ties loosely to any one structure. Writing the text from computed properties addresses both, which is the paper's case for synthetic data over more papers. In a search for a solid lithium electrolyte, the top 30 included two compounds other work had already flagged, Li6PClO5 and Li4SiO4.
Evidence and limits: Everything is computed: the descriptions are written from Materials Project's calculated properties, and retrieval is scored against the same values. Ionic conductivity is not in the descriptions, so the electrolyte search used band gap alone, and four of its top 30 have a computed band gap of zero. The paper-text baseline drew on 4 million papers that copyright keeps from being shared, so a rerun needs its own publisher licences. No crystal was made or measured.
Claude and two mathematicians push the elliptic curve rank record to 31
What happened: Levent Alpöge (Anthropic) and cryptographer Ava Howell used an unreleased internal Claude model to construct an elliptic curve of rank at least 30, then 31 three days later. The lower bounds are unconditional.
Why it matters: The last jump, rank 28 to 29, took 18 years; this one took three days. But disclosure is thin: no paper, no account of the search.
OpenBind-0 releases a fully open co-folding model and finds four more years of PDB data did not help everywhere
What happened: OB0, trained on PDB data to June 2025 and released under Apache 2.0, rivals AlphaFold3 and Protenix on standard benchmarks, but top-25 success falls from 92.2% on EV-A71 to single digits on flexible RdRp targets.
Why it matters: Clear gains came only where new structures filled unfamiliar regions: data value depends on which gap it fills, not on its size. This is OpenBind's own evaluation.
Microsoft Research pretrains Orbformer, one neural network wavefunction for many molecules, and reaches chemical accuracy on breaking bonds at a cost that competes with classical methods
What happened: Adam Foster, Frank Noé, Jan Hermann and colleagues at Microsoft Research AI for Science pretrained one wavefunction model on 22,350 equilibrium and stretched small-molecule structures, then fine-tuned it on unseen molecules; code and weights are public. Earlier pretrained neural network quantum Monte Carlo covered at most 700 structures. Its Diels-Alder activation energy matched experiment within chemical accuracy (1 kcal/mol), and on five bond-breaking curves of up to 48 electrons it was on or ahead of classical methods' cost and accuracy.
Why it matters: Breaking bonds, which many reactions and much transition-metal chemistry involve, is where standard quantum chemistry methods fail, and the methods that work are costly and tuned by an expert for each system. Pretraining spreads that cost over many molecules: fitting a 20-point curve in one go was about 20 times cheaper than point by point, and pretraining cut the cost of reaching chemical accuracy 16-fold on ethane. For large-scale use, the authors see it as a way to make training data for cheaper reactive force fields.
Evidence and limits: All results are the authors' own calculations. Only the Diels-Alder energies, N2 dissociation and small-molecule atomisation energies are compared with experiment; the five bond-breaking curves are scored against the authors' own costly reference runs. The gain shrinks for molecules unlike the training set: on L-alanine, pretraining reached 1 kcal/mol only about 10% faster.
The Baker lab's SAPP and DMX turn protein-design testing into a 48-hour batch
What happened: SAPP expresses, purifies and characterizes hundreds of designs in one 48-hour cycle with about six hours of bench work; DMX recovers individual constructs from cheap pooled DNA, cutting DNA cost an estimated 5-8x.
Why it matters: The bottleneck in protein design has moved from generating candidates to building and measuring them. It remains an IPD-internal screen, not a closed loop.
Bioptimus gets 1,750 colorectal cancer patient profiles from Cancer Research Horizons to train and test M-Optimus
What happened: The S:CORT dataset joins Bioptimus's STELA data programme as its first UK cohort, with imaging, genomic, transcriptomic and clinical data per patient. No financial terms were given.
Why it matters: A model company's edge here is patient data it can train on and also test against. Bioptimus says S:CORT will serve as an evaluation benchmark, but no results were released.
An individualized cancer vaccine whose neoantigens an algorithm picks meets its Phase 3 endpoints
What happened: At a prespecified interim of INTerpath-001 (1,137 melanoma patients), intismeran plus KEYTRUDA beat KEYTRUDA alone on recurrence-free and distant-metastasis-free survival. Only topline results are public.
Why it matters: The algorithm earns credit only as part of the whole system the trial tested; no hazard ratio is public yet, and overall survival is immature.
511 AI-designed or AI-ranked antibodies went into a blinded challenge and were measured under one set of protocols: at picking the strongest binders, all but one method did worse than choosing at random
What happened: AIntibody, modelled on the CASP structure contest, took submissions from 29 organisations on three tasks: improving affinity from phase-1 sequencing output, ranking the strongest binders inside three sequence clusters, and designing new loops for sequences outside the library. Outside labs made and measured every antibody: surface plasmon resonance for all, the slower KinExA for the strongest, and a five-assay panel for whether one can be made into a drug. A few submissions came in under 100 pM; one cluster's own control antibody was already at 36.6 pM.
Why it matters: The gains show up where there is data to learn from. On improving antibodies inside a library the submissions did well: the winner, Aureka, reached 95 pM, about 2,000 times the parent antibody, though statistically tied with the best antibody the lab's own experiments produced, and a consensus sequence built without machine learning came third. On deciding which clone to test, all but one method was beaten by picking at random. Of the designs made for sequences outside the library, 51 of 168 did not bind at all.
Evidence and limits: One challenge, one target, one set of assays. The scoring let high affinity offset a biophysical failure: the strongest out-of-library design, at 2.9 pM, failed the chromatography test. Blinding rested on the organisers' integrity rather than on any check in the data, which the authors say the next round will fix. Entering meant disclosing the method, which the authors think kept some groups away.
Generalist AI's GEN-1.5 learns a new robot task from one demonstration, succeeding 59% of the time with no retraining
What happened: Generalist AI says GEN-1.5, a robot model that reads 30 seconds of video and sensor data and sets arm motions 100 times a second, can be shown a 3 to 12 second demonstration of a new task and then do it with no further training. Across 10 short tasks (unzipping a pouch, twisting off a jar lid) it succeeded 59% of the time this way, and 83% after 10 training steps on 5 minutes of demonstrations. It was pretrained on recorded human activity in homes, warehouses and factories.
Why it matters: A self-driving lab needs robots that can be set to a new bench step quickly, and today each new task means collecting data and retraining. If showing a robot a task once is enough to get it started, moving a robot to a new protocol would take a recording of the step rather than a new dataset and a training run. This is the company's third model in nine months, after GEN-0 in November 2025 and GEN-1 in April 2026.
Evidence and limits: Every number is the company's own, from a blog post with no paper, code or weights. The post does not say how many trials each task got, and the tasks are short household manipulations it calls simple, none of them lab work. The 59% carries a spread of 10 points across tasks.
Network Bio launches with $50 million from Section 32, Thiel Bio, Founders Fund and others to train AI on patient tissue from academic biobanks
What happened: Network Bio, co-founded by chief executive Asad Ali Ahmad, Raphael Potter and Hani Goodarzi, links academic medical centres' biobanks, Mass General Brigham, Penn and Colorado Anschutz among them, under shared standards, so tissue, paired blood, molecular data and patients' later outcomes can train disease models. Investors named are Section 32, Thiel Bio, Founders Fund, Breyer Capital, Blue Venture Fund and JSL Health Capital, with no lead or stage given. It also announced a collaboration worth over $30 million with an unnamed Fortune 100 healthcare company.
Why it matters: Patient tissue with recorded outcomes is the training data most AI models of disease lack, because any one hospital holds too few samples of a given disease. Network Bio's bet is that several biobanks processed to one standard can supply it. Mike Pellini of Section 32, the board's chairman, and Morgan Cheatham of Breyer Capital spoke for the investors; the data generated go back to the partner hospitals as well as to paying customers.
Evidence and limits: The $50 million, the investor list and the $30 million collaboration are the company's own announcement; the partner is unnamed and no terms beyond the value were given. The company says its models already find signals in diseases they were not trained on; the published work it cites is on cell-free RNA in blood, not on the new tissue network.
Anthropic reports Claude-designed protein binders working in the lab
What happened: External wet-lab partners validate binders for 14 of 15 targets in a model-assisted design workflow.
Why it matters: External wet-lab results make this more than a model demo and test whether a general system can compete inside a specialized design workflow.
ASI-Bench tests agents without step-by-step research guidance
What happened: Average performance drops sharply when agents must choose their own research methods rather than follow prescribed steps.
Why it matters: It isolates a central weakness of AI scientists: executing a method is easier than choosing the right method and knowing when to change course.
OpenAI's LifeSciBench: 750 expert-written life-science tasks, with OpenAI's own GPT-Rosalind the best model at a 36.1% pass rate
What happened: 173 PhD experts wrote the tasks, each with a rubric, graded by GPT-5.5. Of five models, GPT-Rosalind scored 0.576 and passed 36.1%; Grok 4.3 passed 13.0%. No model passed 171 tasks (22.8%).
Why it matters: A hard, unsaturated test of life-science judgment. But OpenAI wrote it, built the top model and used GPT-5.5 as grader; answers are single-turn, and public release of tasks may be limited.
Enveda reports ENV-308, a pill built to mimic an exercise hormone, was well tolerated in 88 healthy volunteers
What happened: ENV-308, found with Enveda's PRISM model, mimics Lac-Phe, a molecule the body releases during intense exercise. In the Phase 1 trial, the company reports no serious adverse events, discontinuations or dose interruptions across the dose range, and lower circulating leptin, most in people with the highest baseline levels; it gives no leptin figures. Enveda, whose founder and chief executive is Viswa Colluru, plans Phase 2 in people stopping GLP-1 drugs.
Why it matters: Another drug Enveda found with its AI model reaches people, and the trial shows only that it is safe enough to keep testing. Leptin is an exploratory marker with no numbers given, the weight claims come from animals, and the readout is a company release, not a paper.
GenBio AI, co-founded by a Nobel laureate, releases a preview of AIDO Cell, a model that simulates a human cell from DNA to whole-cell behaviour
What happened: The company says AIDO Cell is multi-scale and stateful: a researcher can perturb the cell at one level, and later perturbations build on earlier ones. The preview covers two cell lines, K562 and HepG2, and in one preliminary study modelled how imatinib acts in leukemia cells. Co-founder David Baker says the cell is not solved; co-founder Eric Xing calls this a preview, with early access and more accurate versions to follow.
Why it matters: A virtual cell is the model drug discovery keeps asking for, and this one comes from a company founded around a Nobel laureate. What is public is a description and one qualitative example: no accuracy figures, no comparison with other models and no prospective experiment. For now it shows that a system exists, not that its predictions hold.
Evidence and limits: Everything comes from GenBio's own release as SynBioBeta carried it; genbio.ai returned 403 and no paper or technical report was found. The imatinib example is described, not shown with data.
LiteFold's LiteMol-1 puts small molecules, peptides, macrocycles and PROTACs in one generator built for agents
What happened: One masked-diffusion model generates and edits several molecular classes in sequence space; with tree search, peptide chemical validity rises from 0.77 to 0.98. Full PROTAC generation stays around 25% valid.
Why it matters: For agents, cheap and editable calls may matter as much as single-call accuracy; the evidence is computational only, and the training recipe is not fully disclosed.
Lanyon AI, a Princeton lab building formally verified scientific AI, emerges from stealth with a $10.6M first round led by Dimension
What happened: Lanyon AI says it closed a $10.6M first round led by Dimension, with Industrious Ventures taking part; its own blog adds the angel Siqi Chen. Co-founders Jonathan Gorard, Ammar Hakim and Jimmy Juno come from Princeton and its plasma physics lab. Its agent has a model propose a formal specification, and symbolic code then writes the solver and its proof together. Aerospace, propulsion and nuclear are the first targets.
Why it matters: Dimension's partner says next-token prediction ends at "mostly right", which does not clear the bar for flight controls or nuclear systems; that is the bet. Venture money is backing a team that moved from academic plasma physics and formal methods to a company, aimed at simulation where an error has a cost. The founders are people to follow.
Evidence and limits: The round is the company's own account: the release gives no valuation, and the blog dates the close to July. Claims that Lanyon is correct by construction and far cheaper than frontier models rest on its own benchmark notes, which no outside group has run; the company says matching a user's intent remains an open problem.
Inherent trains a 27B agent that directs frontier coding agents
What happened: Faraday calls a frontier coding agent as a tool; by the team's own judge it outscores Claude Opus 4.8 and GPT-5.5 on 73% of training-split tasks and 60% of held-out ones.
Why it matters: It separates scientific judgment from execution: the small model decides what to try, the large one writes the code.
AI antibody-discovery company Aureka Biotechnologies closes a $100M Series B in tranches, taking its total to nearly $200M
What happened: Granite Asia alone funded the first tranche; an unnamed industry investor led a later one, with HighLight Capital joining and existing backers Qiming Venture Partners, Matrix Partners China and NRL Capital following on. Founder and chief executive Weian Zhao leads the company, which says the money trains its next generation of biological foundation models, after AuraIDE, and expands Lab-in-the-Loop, the screening platform whose results go back into training. Yinghui Kuang spoke for Granite Asia. Valuation not disclosed.
Why it matters: Investors are paying for the claim that an antibody model needs its own lab feeding it data to keep improving. The company says it earned tens of millions of dollars in revenue over two years, with pharma partners, and that its open model, OpenDDE, ranks well in outside tests; it names neither the partners nor any drug.
Claude improves a longstanding bound related to the Riemann hypothesis
What happened: An unreleased Claude research model raises the known lower bound for the fraction of Riemann-zeta zeros on the critical line from 41.6% to 67.2%, with a formally checkable proof.
Why it matters: Unlike Olympiad performance, this is claimed novel mathematics produced while exploring an unsolved problem, without claiming to solve the Riemann hypothesis itself.
Discovered Materials, a YC startup using AI agents to look for chip-cooling materials, raises a $9M seed led by Lightspeed India Partners
What happened: Hemant Mohapatra, the Lightspeed partner who led the round, and co-founder Advaith Sridhar spoke to TechCrunch. Y Combinator, Peak XV Partners and angels Paul Graham, Gokul Rajaram and Thariq Shihipar also invested. Sridhar and Akash Ramdas, a Stanford materials-science PhD, use Anthropic models in their own harness to propose candidates and physics models they trained to screen them. Valuation not disclosed.
Why it matters: Another bet that agents can search for materials faster than a PhD student's 20 guesses a day, aimed at heat in AI chips. The company says it has found materials that match ones chipmakers use but gives no details, and its own lead investor says filtering and making the materials, not finding candidates, is the bottleneck.
Purdue's torch test: Radical AI's alloy RAI-939 gained 0.25% mass where the Apollo-era alloy C103 lost 32%
What happened: Radical AI says Purdue torch-tested RAI-939 against C103 above 2,500 C and calls it 125x better at surviving the environment. It came from 300-plus materials made in 16 weeks.
Why it matters: An outside lab tested a named alloy. One minute of torch on one sample is not a flight spec, and no Purdue report was read.
WeatherNext Cyclones gains roughly a day of forecast lead time
What happened: A Nature paper reports state-of-the-art forecasts for cyclone track, intensity and wind structure, with at least a day of average lead-time advantage and ensembles of up to 1,000 members.
Why it matters: This is unusually strong AI4S evidence because forecasts are compared with mature operational systems and then continuously checked against the physical world.
AI verifiers start auditing the scientific record, flagging errors across whole paper corpora
What happened: Nature reports cases, a company audit and three preprints: an AI prediction led to a 75-year-old chemistry database value being confirmed wrong, and 83.2% of one checker's flags were confirmed by the papers' authors.
Why it matters: Verification is becoming AI4S infrastructure too, but these are leads, not verdicts: on one 2025 benchmark of errata-level errors, the best model found at most 21.1%, at 6.1% precision.
Why it matters: The personnel move concentrates rare systems, model and science expertise in an independent company built around closing discovery loops.
BASF's catalyst arm licenses Orbital Industries' CurieOS to look for automotive emissions catalysts
What happened: BASF Environmental Catalyst and Metal Solutions will run CurieOS, the agent platform Orbital Industries built around its Orb simulation model, on catalyst discovery and development for vehicle emissions. Orbital Industries says it had used the same platform to find the PFAS-free coolant developed for its data-centre brand. Jonathan Godwin, its chief executive, James Gin-Pollock, its chief technology officer, and Saeed Alerasool, BASF's senior vice president and chief technology officer for the unit, each spoke in the release. No financial terms, duration or targets were disclosed.
Why it matters: Until now the record showed Orbital Industries using its platform only on its own products. A catalyst maker licensing it is a different kind of evidence from a benchmark: it says an industrial buyer thinks it worth trying. Every performance claim in the release (100,000 atoms on one GPU, ten times faster than the nearest model) cites the Orb-v3 preprint already on this record, so nothing new is shown about the model, and no catalyst has come out of the deal.
AI Proteins publishes the lab workflow that turns DNA into purified protein in three days, and made 87 of 96 test proteins with it
What happened: Edson N. Cárcamo Noriega, Chris Bahl and colleagues at the company describe a protease built to cut a protein free while it is still stuck to the purification beads, by adding a long, floppy linker so the enzyme can reach it. Around it they built a five-step workflow for liquid-handling robots. Of 95 varied natural proteins plus a GFP control, 87 of the 96 gave more than 100 micrograms per 2 mL of culture at over 90% purity. Plasmids, the protein set, scripts and data are public.
Why it matters: AI can now propose far more protein designs than labs can make and measure, so the bottleneck has moved to production. A company whose business is designing miniproteins has published the part of its pipeline that makes them, with the plasmids at Addgene and the robot scripts on GitHub, so other labs can make and test designs at the same pace and add to shared training data.
Evidence and limits: The test set is natural, single-domain, disulfide-free proteins already known to express in E. coli, and none was designed; some needed an added two-residue spacer before the protease cut them. All the work is the company's own. The code and data are licensed for non-commercial use.
MinerU.Chem adds a chemistry layer to PDF parsing, turning papers into traceable molecule and reaction records
What happened: Shanghai AI Lab chains five modules after general document parsing (relevance filter, structure detection, identifiers, structure recognition, reaction diagrams) and deploys it publicly; each record links back to its page.
Why it matters: Finished PDF parsing had left the chemistry as pixels. The strong numbers are per module; document-level accuracy is unreported.
A Multiply Labs robot cluster matches skilled operators on cell-therapy steps, with a lower cell count at low concentration
What happened: Multiply Labs, Stanford's Laboratory for Cell and Gene Therapy Manufacturing, which has a collaboration agreement with the company, and co-authors from the Università degli Studi di Milano and instrument makers ran a modular robotic cluster against manual work on the same instruments. Buffer exchange on a CTS Rotea gave 83-95% recovery by hand against 81-133% by robot, and viability above 93% against 83-97%, with no significant difference over three to four runs a side; T-cell isolation on a CytoSinct 1000 was comparable, with more spread.
Why it matters: Autologous cell therapies are still made by hand, one patient at a time, and that cost is why they reach few people. What decides whether a robot can take over a validated process is whether the same instrument gives the same numbers, and here it mostly did. Mostly: at low concentration the robot counted significantly fewer cells than the operators (5.63 against 5.91 x 10^5 per mL), the CytoSinct needed its wash and elution lengthened, bag-to-bag transfer was robot-only, and the cells' phenotype and function were out of scope. The authors call it an R&D prototype.
Lanyon AI's own benchmark has five frontier models write PDE solvers with Lean proofs, and finds they often prove something other than the code they wrote
What happened: Lanyon had Kimi K3, Claude Opus 4.8, Claude Fable 5, GPT-5.5 and GPT-5.6 Sol write finite-volume solvers in C with Lean proofs, three trials each, for advection, Maxwell, Burgers and 2D Euler. In the linear tests every detailed-prompt run that finished was scored faithful; under a terse prompt some proved a degenerate case or used native_decide. On 2D Euler, five of fifteen detailed runs gave a wrong solver. Lanyon's own solvers took 7 to 23 seconds and 600 to 2,940 output tokens.
Why it matters: The test targets a failure this record tracks: a proof that type-checks but is about different code. Under the terse prompt, both Claude models proved a first-order, CFL=1 special case for the second-order solver they had written, and some GPT and Kimi runs used native_decide to skip kernel checking. It shows what comparing a proof against the code can catch, and how much the prompt decides.
Evidence and limits: Lanyon wrote the prompts, rubric and scoring, and picked the models; its own solvers are single reference runs, counted in its own DSL tokens. No outside group has run the benchmark. The Lean files are public with no sorry or native_decide, but prove statements about real numbers while the C uses double. Their hyperbolicity and wave-stability theorems are tautologies: the wave-speed estimate is written identically to the eigenvalue, and nothing ties either to the Jacobian.
Owkin's model reading heart ultrasound flags which early-stage hypertrophic cardiomyopathy patients worsen, checked on 48 patients in a second hospital
What happened: Owkin, with cardiologists in Birmingham, Palma and Paris and funding from Bristol Myers Squibb, segmented the heart chambers in archived ultrasound videos of 260 patients without symptoms yet, and trained a model to predict worsening symptoms or a cardiac hospital stay. It scored a C-index of 0.66 on 212 UK patients and 0.67 on 48 Spanish ones. A simpler ratio the model pointed to, left atrium against left ventricle area, scored 0.67 and 0.72. Above its cut-off, 21% of the Spanish patients worsened within three years against 12% below.
Why it matters: The paper says it was still unclear how well a heart ultrasound predicts which symptom-free patients with this inherited heart thickening will get worse; the authors offer the result as a way to manage these patients better at an early stage. The model points to a simple ratio the pipeline measures automatically from the standard ultrasound view hospitals already record, which is easier to check than a black-box score.
Evidence and limits: Retrospective and exploratory by the authors' own account: the analysis was not prespecified, the outside check rests on 48 patients (C-index interval 0.46 to 0.88), both cohorts are European, and it was scored by the team that built it. The cut-off may need recalibrating before use on values a clinician measures by hand.
