Timeline

What happened: Google announced TPU 8i Max, which puts two TPU 8i chips on one package to double compute, on-chip memory and memory bandwidth in the same rack space, and TPU 8tx, which pairs TPU 8t silicon with AMD EPYC hosts so data and model weights stay protected and isolated even while in use. With the training TPU 8t and inference TPU 8i from April, that makes four TPU systems in a year. Amin Vahdat, Google's chief technologist for AI and infrastructure, says TPUs now have nearly 100 customers, universities among them.

Why it matters: Neither system is a new chip: 8i Max packages two existing 8i chips together, and 8tx reuses 8t silicon. Google says that is the point, since a new system takes months and a new chip years, so new configurations can follow what customers ask for. Two inputs change. What can be bought: four systems in a year, one for fine-tuning on private data that stays protected while the chip works on it. Chip design: Google says AI now works on its own hardware, AlphaChip in TPU design and Gemini producing layouts in hours rather than weeks.

Evidence and limits: Every figure is Google's own, measured against its other TPUs, with no independent benchmark. The post gives no absolute specs, prices or availability dates for the two new systems, names no science customer, and shows no example of an AI-made layout.

What happened: Claude Fable 5.2 was asked, with the board design and manufacturing company Diode Computers, to design a USB speakerphone, a voice-assistant hub and a handheld virtual pet. From short prompts it wrote the specifications and schematics, placed the parts on each board and routed the copper, modelled a 3D-printable case, wrote the firmware and listed the parts. Diode built the boards straight from the files and the three prototypes ran with no fix to their circuitry. On Diode's in-house PCB Gym it scored 69%, against 25% for Fable 5.

Why it matters: Models that touch hardware usually stop at chip design or a benchmark score. Here the output is a file a factory built from, and the test is whether the boards worked. For a lab that needs a custom sensor board or instrument controller, layout, routing and the manufacturing checks are the weeks of an electrical engineer's time that stand between a design and a working device; Diode, which does this work for a living, says it did not have to go through the usual rounds of adjustments with the files Claude produced.

Evidence and limits: This is Anthropic's own post with Diode, a commercial partner, which ran PCB Gym, its own unpublished benchmark; the score quoted here is against an earlier Claude. All three are prototypes: no electrical validation, antenna tuning, thermal testing or certification, and none designed for volume production. Claude was connected to no test equipment or factory machine, so Diode's engineers built and tested everything, and the prompting was minimal rather than absent.

What happened: Mathew Chow and Tao E. Li had one agent (OpenAI's GPT-5.6-sol) propose code changes on training cases while a second accepted or rejected each on separate test cases, by speed and accuracy. On ten other molecules the excited-state (TDDFT) step in PySCF took 84-94% less time, with nearly identical energies; a separate job cut the ground-state calculation 5-28%. LAMMPS water simulations took 19-22% less time with identical trajectories, and a serial molecular dynamics code became CPU and GPU versions two to three orders of magnitude faster.

Why it matters: Computational chemists often write their own code, and the first version is usually serial, as the authors note; speeding it up is skilled work that takes time from the science. Here each change had to pass a second agent on cases the first had not seen, and outputs were compared with the original code. In TDDFT the gain was in the method, not only low-level tuning: on one molecule the solver needed 9 iterations instead of 45. Faster codes mean more simulation for the same compute.

Evidence and limits: One lab's preprint, scored by its authors. Repeat runs of the same model ended at different scores; the LAMMPS gains did not hold in pressure-controlled simulations, where one part got slower; the GPU code ignored the request for mixed precision; and the agent-built SCF strategy did not reliably reach the lowest energy or the shortest time. The paper does not say any change was merged into LAMMPS or PySCF. OpenAI supported the work through its academic researcher programme.

By Mathew Chow, Tao E. Li

What happened: Cancer Grand Challenges, founded in 2020 by Cancer Research UK and the US National Cancer Institute, gave Marinka Zitnik's group at Harvard Medical School nearly $7M to build AURORA, an agent that reads papers and databases, calls tools such as AlphaFold and proposes hypotheses and experiments. Three teams get about $1.3M each to use it: PROMINENT on why some mutated cells turn cancerous, NexTGen on CAR-T for children's solid tumours, MATCHMAKERS on how T cells recognise cancer.

Why it matters: A research funder is paying outside teams to put an AI co-scientist into research already under way, and the evaluation is part of the grant: as the teams test its hypotheses in the lab, Harvard will compare the system's reasoning traces with their results, which would show how often its suggestions hold up. Harvard says it is the first Cancer Grand Challenges project on AI and the first to span several funded teams.

Evidence and limits: Nothing has been tested yet: the project launched on 5 November, and neither release gives an end date, a measure of success or a plan to publish the comparison. Neither names the language models AURORA runs on. Harvard gives the money as nearly $7M plus $4M for the teams; Cancer Research UK puts the project at £8m.

What happened: OpenAI and Ligo, an enzyme-design startup, released BioBasicsBench: Ligo scientists turned data from their own experiments into 100 tasks with scoring rubrics (the examples shown include a gel, colony plates and a handwritten plate label), each reviewed by another Ligo scientist, an outside PhD-level wet-lab scientist and a final internal check. At xhigh, the highest setting run on all seven models, GPT-6 Astra earned 66.7% of available points, GPT-6.1 Sol 62.0% and GPT-5.6 Sol 43.8%; 37 model and setting combinations were run, five attempts per task.

Why it matters: Reading which colonies carry the right plasmid, or whether a control plate is valid, is routine work an agent beside a biologist would do, and tasks taken from a lab's own runs keep the ambiguous images that figures chosen for a paper leave out. The failures shown are of that kind: Astra sorted colony PCR lanes into the right groups, then took the larger product as success and picked colonies still carrying the old plasmid.

Evidence and limits: OpenAI built it with Ligo, scored it and compared only its own models; OpenAI says higher scores do not yet mean models can make reliable judgments in the lab. No release of the tasks is reported, and it tests reading data, not doing bench work.

What happened: The New York Times reports that Radical Numerics has acknowledged it did not work alone on the 20 October announcement: Tabula carried out much of the research, including making the phage cocktail the patient received. Eric Nguyen, its chief executive, wrote to the paper that the two worked closely for seven months and that the contract required both to keep the collaboration secret. Tabula says it led the phage therapy. Both agree Radical designed some of the cocktail's phages and Tabula supplied others.

Why it matters: Radical Numerics did not name Tabula when it announced the treatment on 20 October, and Tabula did much of the lab work. So who made an AI-designed therapy is contested by the two companies that made it. George Church held roles at both; he left Radical's on 5 November and told Nguyen that the argument from confidentiality was very weak on common sense and basic ethics. The Times sets it beside two autumn disputes over whether AI companies used other researchers' work, on an OpenAI proof and on Anthropic's enzymes; both companies denied it.

Evidence and limits: The Times's report was read through a licensed translation; nytimes.com refuses automated readers here. It says Church left the board; where he sat is the one thing the accounts differ on. Neither company has published the contract, and the Times writes that the research was done and the treatment given. Tabula says it will give its own account later; the patient's outcome is still undisclosed.

What happened: A language model labelled 207,049 research sessions on Claude's consumer plans, from 2 August to 30 September 2026, by field and task. Data analysis took 33.7% of Claude sessions and 48.1% on Claude Code, against an estimated 27.6% of researchers' time; data gathering about 12% against 28.6%; ideas 4.2% against 7.1%. In a survey of 528 scientists, mostly in the life and health sciences, those two lagging tasks ranked as the most important. Johannes Hermle, Ben Jones and Lukas Althoff led it.

Why it matters: Like Google's study in September, it puts the gain in work done at a screen, not at the bench. Physical steps take 38-41% of researcher time in chemistry, materials and the biomedical sciences, against 4% in the median field, so the report's model caps their output at three to four times what it would be without AI even if AI did all other work, against up to 15 times in computer science and statistics. On Claude Science, the version wired to research databases, idea generation takes about twice its share on Claude.

Evidence and limits: Only consumer-plan sessions are counted, a model rather than a person assigned each one its field and task, and researchers' time shares come from another model. The survey reached active AI users. The speed-up figures are scenarios from an economic model, not measured output, and this is Anthropic's report on its own products.

By Johannes Hermle, Ben Jones, Lukas Althoff

What happened: The startup, which uses AI agents to look for chip materials, posted five problems written with outside physicists and chemists, Stanford's Felipe Jornada among them: a predictive theory of cuprate superconductors (up to $400,000); cheap but accurate reaction energies, and how chip interfaces get their electrical properties from processing ($200,000 each); what a laser pulse does to a material, and how heat crosses solid interfaces ($100,000 each). Four are scored on published results; the chemistry one on a hidden computed benchmark.

Why it matters: Mathematics has prize problems that show whether AI is closing in; TechCrunch notes materials science has no such set. These come with fixed numbers to hit, from superconductors to the interfaces inside chips. Co-founder Advaith Sridhar told TechCrunch the aim is goalposts that show whether models are improving in this field. On the company's own benchmark, models proposed over 500 new materials but a plausible way to make only one.

Evidence and limits: That frontier models given days on each problem made little headway is the company's account, with no runs published. The company alone judges a solution, may solve a problem itself and close it unpaid, and takes a perpetual, non-exclusive commercial licence to every submission, winning or not.

What happened: Paolo Di Achille, Ming-Zher Poh and Google Research colleagues, with Stephen Juraschek of Beth Israel Deaconess, pretrained a network on about 492,000 hours of unlabelled watch data, then trained a classifier on 663 people with 24-hour blood pressure monitoring. In a prospective study at five US sites, 196 adults with no known heart disease or blood pressure drugs wore a Pixel Watch 3 for a week. Against the 24-hour monitor it found 65.8% of those with hypertension at 90.0% specificity; office cuff readings at the first visit found 55.3%.

Why it matters: More than half of adults with hypertension do not know they have it, and a watch many already wear could tell them to get checked. The model needs no cuff to calibrate it, and the study scores it against the 24-hour monitor that guidelines treat as the reference, where earlier wearable studies often used home cuff readings. The authors frame it as a prompt toward a cuff-based diagnosis, not a diagnosis.

Evidence and limits: One study of 196 people, funded by Alphabet, with every author but Juraschek a Google employee. It used one watch, left out people already treated or with heart disease, and missed a third of cases. The authors say randomised trials are still needed to show it finds people who would otherwise go undiagnosed and improves outcomes.

By Paolo Di Achille, Ming-Zher Poh

What happened: In its third-quarter results Ginkgo said it sells two things built on one robotic platform: lab systems made of its reconfigurable automation carts and Catalyst software, bought by Novo, Pacific Northwest National Laboratory, the NSF and ARPA-H, and experiments run in Nebula, its own lab of over 100 carts. Novo's lab, announced in September, is the first such sale to a large drugmaker; a subcontract under ARPA-H's GIVE programme, worth up to $17.5 million, puts the carts into RNA-drug manufacturing. Revenue was $28 million, down 6% on a year earlier.

Why it matters: It shows who is paying for labs that AI can drive: a drugmaker buying robots for its own scientists, government programmes building shared cloud labs, and AI companies buying experimental data. Ginkgo says frontier labs such as OpenAI use its cloud lab to generate training data, and that 22 customers signed for its drug-property testing in the quarter, 20 of them new. Jason Kelly, Ginkgo's co-founder and chief executive, said the US government now treats autonomous labs as national infrastructure.

Evidence and limits: The figures and customer counts are Ginkgo's own; the price of Novo's lab and the size of the AI-data work are not given. A sale shows demand for the equipment, not that it speeds up Novo's research.

What happened: OpenADMET, an open-science effort on drug-safety prediction, measured how strongly compounds inhibit four cytochrome P450 enzymes, which break down most drugs, and withheld the values for 750 test compounds until the challenge closed on 3 November. In the direct-inhibition track, 271 participants took part and 103 entries made the final board. Inductive's entry ranked first, an unnamed entry close behind sat in the same statistical tier, and NVIDIA's entry came third. Inductive, led by co-founders Josh Haimson and Ben Birnbaum, calls it its fourth OpenADMET win in a row.

Why it matters: A drug that blocks these enzymes can raise other drugs to dangerous levels, so chemists want to flag the risk before a compound is made. Here models were scored on measurements produced for the test and hidden from every team, which is a harder check than a public benchmark. Inductive, whose models are trained on an ADMET dataset pooled from industry partners, ranked first again, in a field of 271 participants.

Evidence and limits: The organiser puts the top two entries in one tier, not statistically distinct, so first place is a rank within a tie. Inductive's own method report says the 750 test compounds are close analogues of 75 training compounds, an easier test than new chemistry. It scores predictions against measured data; no drug programme was run on them.

What happened: Vaults are large hollow protein barrels in cells across the eukaryotic tree, and nobody knows what they do. The team searched about 15 billion protein sequences, found relatives of the human vault protein in 29 bacterial and 4 archaeal phyla, and put 24 into E. coli. Seven made vault-like particles under a first microscope and cryo-electron microscopy confirmed four, one read as a ring on that pass. L05, from a phage, builds a shell of 96 copies at 3.2 angstroms, enclosing 25% more volume than the human vault.

Why it matters: This is structural biology out of Anthropic's own biology lab, not a design campaign sent out to be assayed, with the microscopy run at Thermo Fisher's cryo-EM centre and Berkeley's EM lab. The loop is what to read: people chose the proteins and the analysis calls, and Claude searched the sequence databases, wrote each protein into DNA for the host cell, recommended the column and the shortcut of loading ruptured cells straight onto it, and ran the cryo-EM processing. Isolation went from about three days to 90 minutes, a first model 34 hours into processing.

Evidence and limits: What vaults do is still unknown, and the authors say their work deepens the question. The particles were made only in E. coli, not shown to form in the phages and bacteria they come from. A small molecule sits in every wall copy and is unidentified: coenzyme A is a shape match. The report was not yet posted on bioRxiv when Anthropic published, the time savings are the team's own estimates against no baseline, and a provisional patent is filed.

By Januka S. Athukoralage, Nicholas T. Perry, Matthew G. Durrant

What happened: People familiar with the matter told Bloomberg that Alphabet's drug-design company is seeking roughly $5B in new money at about $50B. The financing has not closed, no investor is named, and Isomorphic did not respond. Bloomberg's report of 8 October gave at least $40B, possibly as high as $50B, and no amount. Max Jaderberg, its president, said at Bloomberg Tech in London on 2 November that its work is still preclinical; Demis Hassabis, its CEO, said in January that first trials should start by the end of 2026.

Why it matters: At about $5B, the round would be some 2.4 times the $2.1B Series B of May, for a company whose own drugs have not yet been tested in people. Investors would be pricing the drug-design engine and its pharma deals before any clinical result, and the size shows how much private money one AI drug-discovery company may now draw ahead of its first trial.

Evidence and limits: The only source is unnamed people speaking to Bloomberg: Isomorphic has confirmed neither the amount nor the valuation, no investor is named, and the round may change or fall away. Jaderberg's own word is that the work is still preclinical.

What happened: Johan de Jong of Columbia started the Stacks Project in 2005, and hundreds of mathematicians have grown it to about 7,500 pages. Axiom Math's coordinated agents translated all of it into Lean, linked each of its permanent tags to the matching formal declarations, and released the library open source with a side-by-side reader. The company says the work turned up dozens of statements that lacked a hypothesis or did not hold as stated, and records counterexamples to them.

Why it matters: Axiom says modern algebraic geometry has been far harder to formalize than fields with mature Lean libraries: one modern proof leans on long chains of commutative algebra, schemes and cohomology that had no Lean counterpart. With those foundations importable, a person or an agent checking a new result can start from verified definitions instead of rebuilding them, which matters as AI writes more proofs than people can read. Where the group classification a week earlier was one theorem, this is a library meant for reuse.

Evidence and limits: Size, faithfulness and the errors found are Axiom's account: its own agents audited whether each formal statement matches the text, with targeted audits of key definitions and theorems, and no outside review is reported. De Jong asked on 21 October that AI-generated contributions not be sent to the project itself, so the fixes stay in Axiom's repository.

What happened: In Nature, Mingyu Wu, Zhaoguo Wang, Chunhai Fan, Haibo Chen and colleagues describe a lab built around AI agents from the start. From a user's request, the agents plan and run DNA and RNA work (synthesis, transcription, amplification, sequencing) across more than 20 instrument types and over 11,000 steps. In four tests it matched expert-tuned yields for enzymatic DNA synthesis, ran concurrent requests 3.2 times faster, stored a text in DNA and read it back exactly, and found and confirmed a KRAS G13D cancer mutation that clinical testing had detected.

Why it matters: The authors say existing self-driving labs sit mostly in chemistry and materials, chase one goal through a fixed workflow on a few instruments, and use AI for one step. Molecular biology needs many different machines and long chains of steps. If agents can plan and schedule a whole nucleic-acid lab, it can take requests from several users at once, which the authors call science as a service. Milad Abolhasani of North Carolina State University, who refereed the paper, told Nature's news team the system stands out for its autonomy and for how readily it takes in different equipment.

Evidence and limits: All four tests were set and run by the authors in their own lab. The yield comparison is against expert-optimized synthesis; the 3.2-fold gain is for running requests concurrently, not for any one experiment. The mutation test confirms one known mutation, a demonstration rather than a diagnostic study.

By Mingyu Wu, Zhaoguo Wang, Fei Wang, Chunhai Fan, Haibo Chen

What happened: In Nature, Heqin Zhu, Peng Xiong, S. Kevin Zhou and colleagues at the University of Science and Technology of China train a language model on RNA sequences together with how their bases pair, with the pairing taken from several structure predictors at once so no single one's bias dominates. Its 3D version, Zfold, performs comparably to AlphaFold3 on CASP15, CASP16 and RNA-Puzzles. The model also sets new results on 2D structure benchmarks and carries over to newly released RNA families. Dataset, weights and code are open.

Why it matters: What an RNA molecule does follows from how it folds, and far fewer RNA structures have been solved than protein ones, so models trained on sequence alone learn little about folding. Building base pairing into pre-training is one way around that shortage. With the 21 million pairs and the weights released, other groups can retrain the model, test it on their own RNAs or build on it.

Evidence and limits: The results the abstract reports are scores the authors computed on existing benchmarks; it reports no new experiment testing a structure the model predicted. The pairing it learns from is itself predicted, not measured. On par with AlphaFold3 is the paper's summary across three test sets, not a claim of beating it.

By Heqin Zhu, Peng Xiong, S. Kevin Zhou

What happened: In Nature Methods, a team led by Pengtao Xie at UC San Diego, with James Zou of Stanford among the authors, describes RobPicker, which uses a small set of cleaner labels to correct wrong or missing labels in the training data and to give rare protein species more weight. On five real cryo-ET datasets and the CZII benchmark it scored higher than three top-ranked Kaggle pipelines; averaging its beta-galactosidase picks gave a 5.2 Å map, against 7.8 Å from the ground-truth picks.

Why it matters: Cryo-electron tomography images proteins inside intact cells, but each particle must first be found, using labels that are slow to make by hand and often wrong when made automatically; small, scarce complexes are the ones models miss. RobPicker asks a lab to curate a small, cleaner validation set instead of re-annotating everything, and once trained, its inference was about 160 times faster than the authors' estimate for PyTom template matching on the CZII test set. For groups with many tomograms, the hand work moves from cleaning every training label to curating that subset; the rest can stay noisy.

Evidence and limits: Every comparison is the authors' own, and two of the five datasets are their unpublished data; no outside group has reproduced it. The sharper map held for one protein: thyroglobulin maps came out about equal (8.3 Å against 8.5 Å). The authors say it does not model differences between labs and acquisition settings, and the cross-organism test used one bacterial species, judged by map resolution rather than scored picks.

By Youwei Liang, Ramtin Hosseini, Pengtao Xie

What happened: ALCHEMI Toolkit 0.3.0, released by NVIDIA on 4 November, adds distillation, which trains a small student model on a large teacher model's predictions, to its open-source (Apache 2.0) Python framework for running AI atomic simulations in batches on NVIDIA GPUs. It also adds LoRA fine-tuning, enhanced sampling for rare events, nudged elastic band searches for reaction paths and tools for predicting molecular crystal structures. It runs MACE, AIMNet2 and UMA models, or a user's own.

Why it matters: Machine-learned interatomic potentials let AI stand in for slow quantum calculations in materials and chemistry simulation. An NVIDIA scientist working on AI for science writes that the most accurate of them are often too large to run at the scale many simulations demand, and that distilling one into a compact model makes those workflows practical on far more systems and longer timescales. NVIDIA writes this layer itself and gives it away, and its GPU code is built on CUDA and NVIDIA's own Warp library: groups that adopt it train, distil and sample these models on NVIDIA GPUs.

Evidence and limits: A software release with no results: the changelog and README report no accuracy or speed for any distilled model, though the release adds tools to measure a student model's accuracy and stability. That large models are too big to run at scale, and that distillation fixes it, is one NVIDIA scientist's account in a LinkedIn post.

What happened: Sam van Puyenbroeck, Tim Van Den Bossche and Tine Claeys at VIB-UGent and colleagues ran eight deep-learning de novo sequencing models, among them Casanovo, InstaNovo and π-PrimeNovo, on three public proteomics datasets, with database-search identifications as the answer key. On the Orbitrap dataset, 64.2% of spectra were read right by at least one model, the best alone 50.4%. Three rescoring and refinement tools gave little or no gain. Most errors did not depend on the model and traced to spectra missing fragment ions; the models' wrong answers themselves rarely agreed.

Why it matters: These models are how proteomics would find peptides no database lists. If their errors come from what a spectrum lacks rather than from the model, a better architecture alone will not fix them; the authors point instead to information outside the fragments, such as retention time and ion mobility, and to better fragmentation in the experiment.

Evidence and limits: The answer key comes from database search, so the test covers only peptides a database already holds. Performance fell sharply on timsTOF data, because the models were trained on Orbitrap spectra.

By Sam van Puyenbroeck, Tine Claeys

What happened: 10x Genomics, which makes the Chromium single-cell sequencing instruments, launched Sentira Single Cell on its 10x Cloud. A researcher asks a question, or lets the agent propose one from the data; the agent drafts an analysis plan for the researcher to review, writes and runs the code, and keeps the methods and code in an editable notebook. It runs on Anthropic's Claude models. It is paid in credits, and 10x told GEN a complete analysis takes about 60.

Why it matters: Single-cell work usually stalls at analysis: each lab's question needs its own code, and many biologists who produce the data cannot write it. Putting an agent inside the instrument maker's own cloud is meant to let Chromium users take that step without a bioinformatician. Michael Schnall-Levin, 10x's chief technology officer, told GEN that analyses that took weeks could take hours. It extends the Sentira platform 10x announced at the end of September, whose spatial version is in early access.

Evidence and limits: The release shows no test of the agent's analyses against a bioinformatician's, and 'weeks to hours' is the company's estimate. Schnall-Levin told GEN the agent is not perfect and its output needs review before it goes into a paper.

What happened: Erik Demaine and six coauthors, writing as the MIT Hardness Group, note that the OpenAI and Anthropic algorithms break ETH but stay consistent with Richard Stearns and Harry Hunt's 1990 Satisfiability Hypothesis (SH). They prove SH equivalent to Weak ETH, which rules out 3-SAT in 2^(n^(1−ε)) time, and give a framework under which most ETH-based lower bounds survive under SH, losing an arbitrarily small constant in the exponent.

Why it matters: Decades of conditional lower bounds rest on ETH. If the refutation stands, this paper says most of them can be kept by swapping the assumption, at a small loss, rather than thrown out; the authors set that against a WIRED headline calling the research irrelevant. This paper, like notes it cites by Michael Lampis and UC San Diego's ETH Group, already builds on the claim being true before anyone outside has confirmed it.

Evidence and limits: The paper assumes the refutation is correct and does not check it. Its results are human proofs in an unreviewed preprint, and the authors say OpenAI's GPT 6.1 Sol helped with the writing.

By Josh Brunner, Erik D. Demaine

What happened: Arda Uzunoglu, Eric Nalisnick and colleagues at Johns Hopkins, with Mehrdad Farajtabar of Apple, extend a theory of why models hallucinate to discovery: both draw on facts seen once in training. They fine-tuned gpt-oss-20b and gemma-3-12b on sets whose share of such facts they set from 0 to 100%. One task asks for a formula with given properties, scored on materials new in the Materials Project's September 2025 snapshot; the other edits a DNA sequence to a target activity. In both, exact matches rose steadily with that share.

Why it matters: It points to how training data are sampled, not only model size or reasoning, as a lever on how often a science model proposes something true and new. Showing the same examples several times, which keeps diversity fixed, made the trend disappear, so diversity alone does not explain it; more reasoning raised hit rates without changing the trend. The same lever raises false answers too: on held-out data from the training distribution, discoveries and hallucinations rose together.

Evidence and limits: A discovery here is exact recovery of a known answer held out by date or at random, scored in software; nothing was made or measured. Two small open models and two tasks; the theory's bounds rest on parameters the authors say are hard to estimate. An unreviewed preprint.

By Arda Uzunoglu, Eric Nalisnick

What happened: Yuxuan Du, Dacheng Tao and colleagues at Nanyang Technological University turned 60 published studies in quantum simulation, algorithms and measurement into tasks: rebuild the method from the paper and reproduce one stated result. Checking agents' attempts against the released code, with experts confirming, found problems in 18 of the 60, from missing details to one study whose paper and code could not be reconciled, which was dropped. On the other 59, Claude Code with Opus 5.5 succeeded about 93% of the time; six third-party agent frameworks scored 19 to 48%.

Why it matters: It separates two reasons agents fail to reproduce published science: the agent, and the paper. Once the gaps between papers and their code were repaired, the strongest systems reproduced most results, and spending more on inference did not reliably buy more success. The authors' own agent, which stores fixes experts verified and reuses them, lifted the low-cost DeepSeek V4.1 Flash from 68% to 85%.

Evidence and limits: Computational results rerun in software and judged against each paper; nothing was measured on quantum hardware. The authors chose the 60 studies and built the agent they compare; its stored fixes came from 5 of the 59 tasks, which its 85% includes. One field, and an unreviewed preprint.

By Yuxuan Du, Hanqing Zhao, Dacheng Tao

What happened: Deep Principle, the Hangzhou company behind the Mira research agent, further trained Z.ai's open model GLM-5.3 on chemistry, materials and life-science data, feedback from its own scientists, its lab notebook records and the step-by-step histories of its own materials searches, failed attempts included. The weights are on Hugging Face under the MIT licence, behind a sign-up gate. In the company's tests it beats GPT-6 Astra on the FrontierScience-Research benchmark and on two of three in-house research tasks.

Why it matters: The discovery results reported this year by frontier labs came from closed models; this one can be downloaded, inspected and adapted, and it was trained on how a lab's research decisions turned out, not only on published knowledge. The test closest to discovery is a search: choosing new metal complexes over five rounds, seeing only the results of its own picks, the model's best energy gap, averaged over runs, reached 4.318 eV, against 3.689 for GPT-6 Astra and 3.353 for Kimi-K3.

Evidence and limits: Every score is the company's own, on benchmarks it chose and internal tasks it wrote, and the public-benchmark numbers appear only in a chart. The post reports the complex search as evaluations, with no complex made in a lab, and it covers one property. GPT-6 Astra still leads on assigning structures from NMR spectra, by 3.33 points.

What happened: Joseph Shunia's test decides primality with one main congruence, (1+X)ⁿ = 1+Xⁿ in a cyclotomic quotient ring, in L³² bit operations for an L-bit number. Choosing its parameter needs a theorem by D. Cureton and J. Thorner that no Dedekind zeta function vanishes past real part 31/32, which extends OpenAI's 4 October zero-free theorem for Hecke L-functions and was checked in Lean with AxiomProver. The paper says GPT-6 Astra developed and wrote the proofs, from Shunia's conjectured test, notes and direction.

Why it matters: OpenAI's theorem was extended by D. Cureton and J. Thorner, checked in Lean with another company's prover, and then used by GPT-6 Astra to build an algorithm. An AI-produced result is serving here as an input to new work, two steps on, not only as something to check or rewrite.

Evidence and limits: It is not faster: Lenstra and Pomerance's unconditional test already runs in L⁶, and this one reaches L⁶ only with a stronger zero-free bound or GRH. No outside check of the paper's own proofs, which the model wrote, is reported; the Lean check covers only the Cureton and Thorner theorem it rests on.

By Joseph M. Shunia

What happened: Fabio Dias of the University of Surrey and Yvens Lima of the Federal University of Ceará take OpenAI's proofs as correct and ask how large a formula must be before its bound beats trying every assignment. The preprint states no constants: with a hidden constant of 2 the crossover is near 10³² variables, with 3 near 10¹²⁶. From the tensor-contraction proof's own constants, the time it certifies on bounded-density formulas stays above 2ⁿ until log₂(n+m) passes 10⁹.

Why it matters: If the proof holds, ETH is still refuted; what changes is what the result means. It guarantees no faster count for any bounded-density formula that could ever be stored. That separates the claim about theory from any claim about solving real formulas faster, and gives the review a number to check.

Evidence and limits: The authors say they do not assess whether OpenAI's proof is correct. Their limits on the clause-product route hold in a cost model they built to match what the preprint's evaluation pays, and the parity form and tensor contraction escape them. The work was done with Claude Opus 5.5 through Claude Code, refereed by a second instance, and is an unreviewed preprint.

By Fabio Dias, Yvens Lima

What happened: Stepan N. Ostarkov, Michael G. Medvedev and colleagues at the Zelinsky Institute in Moscow gave models 60 labelled molecules per task. Prompted with the examples, GPT-6 Astra beat the best chemistry model on solubility, hydration energy, lipophilicity and two quantum-chemistry targets: R² 0.85 against 0.53 on orbital energy, 0.97 against 0.63 on hydration. Chemistry models won only on BACE1 enzyme inhibition, the control. On orbital energy the best of them needed 1,000 molecules to pass Astra's score on 60.

Why it matters: Most chemistry projects have fewer than a hundred measurements, which is exactly the case chemistry foundation models were built for. Here a general model prompted with the data did better on the property-prediction tasks tested, outside the drug-like classification the chemistry models were trained around. If it holds, a lab with a small dataset should try a frontier model before fine-tuning a chemistry one.

Evidence and limits: All six benchmarks are public, so the general models may have seen the answers in training. With labels shuffled they still scored well, and the authors flag likely memorisation for Claude Opus 5.5. Each task used 100 molecules over 15 random splits; this is a preprint.

By Stepan N. Ostarkov, Gregorii A. Semakin, Ivan A. Bespalov, Michael G. Medvedev

What happened: Timur Bazhirov at Exabyte (Mat3ra), with co-authors at Berkeley Lab, Stanford, SLAC, NVIDIA, UC Berkeley and other institutions, released Q3, a density functional theory engine built from scratch. Claude Code, Gemini and Cursor agents wrote its 82,200 lines of C++ and over 1,800 tests in weeks; the scientists set the interfaces, equations and tolerances. On silicon and copper test cells its total energies agree with Quantum ESPRESSO's within 0.014 mRy. It is open source under Apache 2.0.

Why it matters: DFT calculations are the main source of training data for machine-learned materials models, and the codes that run them are decades-old Fortran that is slow to extend. If agents can write and test a new implementation in weeks while scientists check the physics, a new method can reach a fast, validated code much sooner. Q3 already runs on GPUs at NERSC and writes its results in BerkeleyGW's file formats for many-body calculations.

Evidence and limits: Accuracy and speed are checked against one reference code on small cells of up to 54 atoms that the authors chose, and outside review so far is black-box checks of its numbers. The paper gives no count of human hours, so 'weeks' cannot be checked.

By Vsevolod Biryukov, Timur Bazhirov

What happened: Flagship Pioneering combined the platforms of three of its companies, which BioPharma Dive says had raised $374M between them, with its in-house AI unit, Pioneering Intelligence. Adrian Rawcliffe, formerly CEO of Adaptimmune, is CEO; Flagship senior partner Alec Reynolds chairs the board, and Noubar Afeyan is a co-founder. It starts with programmes in blood disorders such as myelofibrosis, in metabolism, including partnered MASH work, and in immunology. Cellarity's sickle cell drug CLY-124 goes to a separate company.

Why it matters: Each company held one kind of evidence: Cellarity read the state of whole cells, Etiome followed how a disease changes from stage to stage, and Empress showed what a molecule actually binds. Flagship is betting that the three together, plus data on how cells respond when a gene is silenced or a chemical is added, can tell which targets cause a disease, and so which drugs will work in patients before a trial. Reynolds framed the target as the roughly half of drugs that fail in Phase 2.

Evidence and limits: No new funding was announced. The 150-million-cell foundation model, which the release calls the industry's largest, and the more than 30 prediction models are the company's description; no result from them is shown.

What happened: Physicists at National Taiwan University, Shanghai Jiao Tong University, SLAC, the University of Washington and others pretrained a model of whole collision events on 500 million simulated events, then tuned it for four analyses. In simulated searches for new heavy particles and exotic Higgs decays it beat XGBoost, TabPFN and SPANet, most clearly with little data, and in simulation it measured top-quark entanglement 37% more precisely than a published benchmark. In real 2016 CMS data it found the known Upsilon meson at 7.6 sigma, against 6.4 for the published method.

Why it matters: Collider physicists today train a separate model for each analysis. EVENET tests whether one pretrained model can serve many: one large pretraining run, on 512 GPUs at NERSC, then cheap fine-tuning for each new question. The gain is largest in searches for rare signals, where each new-particle hypothesis comes with only thousands of simulated signal events. The pretraining data and three of the four downstream datasets are public on Hugging Face under the MIT licence.

Evidence and limits: Only one test uses real collisions, and it finds a particle already known; the other three run on simulated events. The authors say the model does not yet carry detector and theory uncertainties through, and that use in a live LHC analysis is the decisive test.

By Ting-Hsiang Hsu, Vinicius Mikuni, Yuan-Tang Chou, Yulei Zhang

What happened: Fanxu Meng, Kedar Hippalgaonkar and colleagues at UC Berkeley's Singapore centre, Nanyang Technological University, the National University of Singapore, Monash University and Lawrence Berkeley National Laboratory built SKIER: AI models call analysis procedures the researchers wrote, and plain code does the calculations. On 49 records from the group's own published study of methanol oxidation on platinum, all 11 AI configurations finished the data processing with the shared procedures and none without. The models then proposed three mechanisms and tests to tell them apart; the team ran the tests.

Why it matters: AI models are increasingly asked to interpret experiments and propose mechanisms, and their claims are often hard to trace to the measurements behind them. SKIER splits the work: researchers fix how each measurement is analysed, the model chooses and reasons, and each conclusion points back to its data. In the case study the loop closed once: isotope-labelling, residence-time and oxygen-exchange tests argued against two of the three mechanisms and placed C-O bond cleavage before formic acid forms.

Evidence and limits: One reaction system, from the authors' own earlier study. The 100 single-choice questions on which GPT-5.6-sol scored 100% were expert-designed, and the paper does not say by whom. The authors note that intermediates below the detection limit and pathways they did not consider are not ruled out. A preprint.

By Fanxu Meng, Kedar Hippalgaonkar

What happened: Edison Scientific, the FutureHouse spin-out led by CEO Sam Rodriques, announced an agreement to deploy the Edison AI Platform and Kosmos across Pfizer's research and development. Kosmos will draw on Pfizer's data in oncology, internal medicine, immunology and other areas, and Edison's forward-deployed scientists and engineers will build it into specific workflows with Pfizer's research teams.

Why it matters: This is the third Kosmos deployment Edison has announced this year, after Incyte in May and Population Health Partners in June. Here the AI scientist reads a large drugmaker's internal records, including programmes that did not work, which published papers rarely report. Rodriques describes the aim as letting what each programme learned inform the others.

Evidence and limits: Only Edison announced it: no Pfizer release was found, and no Pfizer person is quoted. No money, duration or scope is stated. It is a deployment, not evidence that Kosmos improves Pfizer's decisions.

What happened: Mistral released the weights of Mistral Large 4 ("Le Chonk"), a 1-trillion-parameter model with 52 billion active, under a modified-MIT licence, free to commercialise except for model-as-a-service providers. It reports a tie with the top open-weight model GLM-5.3 on a coding-agent index (48%) at 47% fewer tokens, and a 96.6% refusal rate on malicious cyber requests, the highest it measures among open models. It also says that in live traffic it saw a potential for the model to help conduct real attacks, and is changing its guardrails.

Why it matters: Mistral's pitch is the open frontier: a model researchers and companies can download and run on their own machines, trained in Europe for what it calls sovereign AI. The open weights are the part that bears on open science tooling. The self-disclosed finding is the one to note: a lab releasing a capable model's weights while saying the same model could help attack live systems, and shipping anyway with guardrails it is still revising.

Evidence and limits: Every benchmark figure is Mistral's own, and it says a full technical report is still to come. The uplift finding is Mistral's own observation from its live traffic, with no number attached.

What happened: Opentrons Labworks, a New York maker of open benchtop lab robots, closed a $40M Series E led by the healthcare investor Avego, with BroadOak Capital Partners; SoftBank Vision Fund and Khosla Ventures are existing backers. It says it has placed over 11,000 robots in more than 80 countries. CEO James Atwood says the money goes to robotics, software and data tools, AI-run experiments and new markets; James Flexner, Avego's managing partner, says every AI-driven discovery still has to be tested in a real lab.

Why it matters: The company's case is that models now propose experiments faster than labs can run them, so the robot that does the pipetting is where the queue backs up. Opentrons has a large installed base and an open programming interface, which is why outside teams have already had Claude drive its Flex (Tetsuwan Scientific) and GPT-5.6 Terra drive its OT-2 (Duke's PEARL). The round is money betting that accessible, open robots like these become the usual way AI models reach a wet lab.

Evidence and limits: A financing, not a result. The robot count is the company's own figure, and the release gives no valuation, no total raised and no measure of how much of its robots' work is run by AI today.

What happened: Cascade will treat blood immune cells from 12 healthy donors with cytokines, chemokines and drugs, alone and in combination, and read them at several time points: 100 million single-cell RNA profiles from Parse's GigaLab and more than 11,000 profiles of over 1,000 secreted proteins from Nomic's Omni 1000, across more than 150,000 perturbation measurements. GSK, Pfizer, Sanofi, Graph Therapeutics and Turbine are founding partners. Data generation has started, and the set will be licensed.

Why it matters: Models of how cells respond to drugs, the virtual cell, are short of data that follow a response over time and measure more than RNA. Cascade measures the RNA inside each cell and the proteins the cells release, from the same samples and time points, so a model can be tested on whether cells' responses predict the proteins they secrete, or held out on unseen donors and perturbations. GSK, Pfizer and Sanofi signing on before any data are released is a sign of how much drugmakers want this kind of data.

Evidence and limits: An announcement: no data have been released yet, and the companies give no date, price or partner terms. The set is licensed, not open.

What happened: Lingxin Cheng, Jack Shireman, Christina Kendziorski and colleagues at the University of Wisconsin-Madison built SMART, about 20 Claude agents that read a published single-cell RNA study, list its findings, then write and run analyses testing each one in another dataset. Across 121 studies of six cancers it wrote 726 reports: 46% of the cross-cancer reports found a transfer, against 75-88% within one cancer. In 10 cases checked by hand, 92-97% of its outputs were correct, and fed summaries with every finding reversed, it rated all 15 low or middle-low.

Why it matters: A published finding is rarely tested outside the setting it came from, because checking it means rereading the paper and redoing the analysis. SMART makes that check routine and points to what is worth following up. ORACLE, a 23-gene lung cancer risk signature, is cited by 161 studies and none tested it in another cancer; after SMART flagged it, the authors found high scores linked to worse survival in public liver (hazard ratio 1.92), pancreatic (1.73) and mesothelioma (3.04) cohorts.

Evidence and limits: Every check reuses existing public data; no new experiment was run. ORACLE showed no significant link to survival in ovarian, brain or colorectal cancer. The team audited its own system, and the preprint is not peer reviewed.

By Lingxin Cheng, Jack Shireman, Christina Kendziorski

What happened: Conor Igoe, Siddharth Narayanan and colleagues at Edison Scientific and MIT CSAIL built 96,000 episodes from drug-response RNA data, CRISPR gene-dependency screens and GWAS results. A model reads a run of related results and states the probability that a held-out result supports a hypothesis. GPT-5.6-sol and Claude Opus 5 rank true above false (AUROC 0.737 and 0.752) but are underconfident, with Brier scores no better than a flat 50%. Qwen3-30B post-trained on scoring rules is called Ibis0; its classifier version reaches a Brier of 0.177 and AUROC 0.820.

Why it matters: An agent that chooses experiments needs to know how far a result should move its belief, not only which answer is likely. The benchmark measures that, and the training improves it. In a simulated task of picking the next experiment, Ibis0 paired with the paper's information-gain estimator got about 2.5 times the information from its first pick that the best prompted frontier model did; prompted alone it did no better. Code, data and models are released.

Evidence and limits: The episodes are structured summaries of published data, with yes-or-no hypotheses balanced at 50%, and experiment choice was tested against a simulator, not in a lab. The experiment-choice gain shows in two of three domains. One model family at one size was trained, on the company's own benchmark.

By Conor Igoe, Siddharth Narayanan

What happened: Ali Saffar Shamshirgar, Johanna Rosen and colleagues at Linköping University, Tallinn University of Technology and Aalto University trained a classifier to recognise the layered structure of MXenes, thin metal carbide and nitride sheets, and ran it over the Materials Project and OQMD. With database tags and literature checks, it turned up 33 compounds reported as made but never recognised as MXene candidates. The team then made seven MXenes with a self-sustaining combustion reaction that needs no furnace and finishes in minutes, and 11 new ones built on seven rare-earth metals.

Why it matters: Most MXenes are still made by etching with hazardous acids such as hydrofluoric acid, in slow multi-step processes. Here the model's job is not to invent materials but to find ones already in the literature under other names, and those finds pointed the team to rare-earth carbides with halogen surfaces, a region where almost no MXene had been made. The combustion route is a candidate for making them at scale.

Evidence and limits: The classifier only sorts structures already in the databases; whether each of the 33 really behaves as an MXene is not tested. The new phases are multilayer powders, not delaminated sheets, band gaps were measured, but the magnetic states come from calculations. One group's work.

By Ali Saffar Shamshirgar, Johanna Rosen

What happened: Prerana Chandratre, Lutz Brusch, Simon Syga and colleagues at TUD Dresden University of Technology paired OpenAI language-model agents with Morpheus, the open-source multicellular simulator their group develops, and had them build, run and revise models of 15 published studies with no human in the loop. With GPT-6.1 Sol, the setups that could run the simulator averaged 83-86% of hidden, study-specific criteria, in about 105 to 110 minutes and US$3.74 a model. Five studies had no published model, and the agents' own mechanisms scored about as high.

Why it matters: Building a virtual tissue model, turning observations of cells into a simulation that reproduces them, takes biology, mathematics and programming at once, so it is slow, specialist work. The deciding factor here was letting the agent run its own simulations: without that, scores fell to 0.55 with the larger model and 0.11 with the smaller. The larger model cost 10.8 times as much per run and scored 0.85 against 0.66, the price-for-quality choice a group adopting such agents would face. Because the agent writes only model files that the simulator runs, a modeller can check what it built.

Evidence and limits: The scorer is itself a language model, one of the two the agents used; it agreed with hand scores by three of the authors at r = 0.93 on an earlier version of the benchmark. Costs are computed at OpenAI's list prices. The authors set the criteria, the models reproduce published behaviour rather than predict new experiments, and the full run transcripts are not public.

By Prerana Chandratre, Lutz Brusch, Simon Syga

What happened: The New York company buys or licenses Phase 1 and Phase 2 drugs from other companies and runs their development itself. Its new Drug Development Accelerator Fund has $600M committed by research foundations, endowments, hospitals and philanthropies; Formation Bio is the general partner through a subsidiary and puts in money of its own. Each drug sits in its own company. CEO and co-founder Ben Liu, chief business officer David Steinberg and chief scientific officer Michael Ehlers spoke in the release.

Why it matters: It is another way for an AI company to pay for drug development: Formation Bio raises money against individual drugs, and investors buy into the drugs, not the platform. Liu said it opens paths to liquidity without selling the parent company and should lower its cost of capital over time. The claim to test is that one shared, AI-supported development team takes each drug to a clinical result more cheaply, within 2 to 4 years of buying it.

Evidence and limits: This is a fund for drug programmes, not an equity round in Formation Bio. No limited partner and no drug the fund will buy is named, and the release shows no case where the platform shortened a programme.

What happened: Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, announced by Koray Kavukcuoglu (SVP, Google DeepMind). It reports Carbon first on version 5.0 of the Artificial Analysis Intelligence Index at 59, and leads on three cybersecurity benchmarks it cites: finding software flaws (SEC-bench Pro, 88.2%), exploiting them (ExploitGym, 54.8%) and patching them (CWE-bench, 71.0%). It resists every attempt on an indirect prompt-injection test (0.0%), against 0.7% for the earlier Argon. Pricing matches Argon at $2 and $10 per million input and output tokens.

Why it matters: Carbon continues the Gemini 4 line the record already tracks, but where Argon shipped with science case studies and a free-access plan for scientists, Carbon's stated focus is cybersecurity and long agentic tasks. Google is gating it: cyber defenders first through its Fairwind programme, developers and consumers later with no date, and it says it is in the US government's voluntary pre-release testing. The rollout and the government testing are the governance step to watch as these models gain offensive and defensive security skills.

Evidence and limits: Every benchmark figure is Google's own, with no independent check and no technical report cited. The internal coding and secure-code gains are on its own teams and tests. Access is restricted, so nobody outside the vetted partners can test the claims.

What happened: Eunbi Rha, Mingda Li and colleagues at MIT built QUANTA: an agent that draws on literature and past runs, plans transport measurements on a Physical Property Measurement System and revises the plan as data arrive, while a rule-based layer checks every command against hard limits. On the Weyl semimetal TaP it recovered from an amplifier overload, chose a current that passed its noise checks and rejected the faster averaging setting, though none fully passed. On Co2MnGa it tested a nonlinear Hall signal's origin in four steps.

Why it matters: Measuring a quantum material means many small calls, such as which current, how long to average, and whether an odd signal is the sample or the setup, and that know-how mostly lives in people. Here an agent makes those calls on a real instrument and real samples, with the reasoning for each step logged. It also stops where the data stop: it left the Co2MnGa signal's origin open, electronic or thermal, because the fit could not tell the two apart.

Evidence and limits: One instrument, two samples, run by the team that built the agent, with no comparison against a researcher doing the same campaign. People still mount samples and change wiring, and high-risk actions need human sign-off. The agents ran on OpenAI's Astra Ultra model. It cleared the overload by widening the amplifier's input range, without finding the cause. Neither campaign produced a new physical finding. A preprint.

By Eunbi Rha, Mingda Li

What happened: Juan B. Pérez-Sánchez, Varinia Bernales, Alán Aspuru-Guzik and colleagues at the University of Toronto, with co-authors in Chile and Spain, had El Agente Quntur, a team of agents built on OpenAI's GPT-5.6 models, choose the active space, the small set of orbitals a multireference calculation treats in full, for 32 excited states of 31 molecules. Against published CASPT2 values the agents' energies were off by 0.239 eV on average, against 0.459 eV for the best of six fixed Active Space Finder protocols.

Why it matters: Choosing the active space is the step in these calculations that still takes an expert, and it is why some chemists say the methods are not a proper model chemistry. Here the agents did more than pick orbitals: they checked which excited state a calculation had actually found, noticed orbitals lost during optimisation and changed course. The prompts, agent traces and calculation files are in a public replication dataset, so the decisions can be audited.

Evidence and limits: The reference is another calculation, CASPT2 from earlier papers, not a measured spectrum. For four cases the agents computed a different excited state from the reference, which the average does not penalise, and case by case the agents and the best fixed protocol each won 16. The 32 cases used 764 million tokens, searches included, and repeat runs of one prompt took different routes.

By Juan B. Pérez-Sánchez, Varinia Bernales, Alán Aspuru-Guzik

What happened: BigHat Biosciences and Anthropic ran two lead-optimisation campaigns on BigHat's Milliner platform, about 1,010 designs over seven rounds in all, four on one molecule and three on the other, with Claude in the loop. Reading the raw instrument curves, they found a thermal measure on each scaffold did not measure what its name implied: on the nanobody, the clumping-onset temperature. They fitted a physical model of the curves, chose new measures and changed the protocol. Both leads beat their parents on binding, heat stability and clumping.

Why it matters: A closed-loop lab optimises the number it is given, so a measure that stands in badly for what a drug needs cannot be rescued by a better model or more data. The gain here came from people and a model reading the instrument signal and changing the experiment, not from more rounds of search. Published lab-in-the-loop work mostly takes the reported assay value as a correct label; this campaign did not.

Evidence and limits: Both companies' own report, not peer reviewed, every measurement in BigHat's lab: two molecules on one platform. The report says each lead's values rest on one or two replicates and that confirmation runs come first, and the new measures were chosen inside the campaign that produced them. The leads are antibody fragments, not drugs, and nothing is tested in animals or people.

By Samuel Stanton, Peyton Greenside

What happened: Over three rounds of design and measurement with AI agents, the company built a 45-amino-acid signalling tail for a CAR, less than a third the length of the 4-1BB-CD3zeta domain in approved therapies. Against that domain it raised receptor levels on the cell surface and antigen-driven IFN-gamma release, with none detected when antigen was absent; it kept killing going longer after mRNA delivery, carried four receptors on one mRNA dose, and depleted more B cells in donors' blood cells after lentiviral and lipid-nanoparticle delivery.

Why it matters: Shortening the tail was expected to cost function, because more activation motifs were read as stronger signalling. The short tail made more receptor per unit of mRNA, which is why the team could put four receptors on one dose, and the report suggests compact tails may lower the dose needed to make CAR T cells inside the body. It is also a worked case of agent-guided design: the agents proposed designs and revised them as measurements came in.

Evidence and limits: The company's own technical report, not peer reviewed, and all of it in cells: human blood cells from a few donors, no animal test. The report says how compact tails perform in a living animal, and how many cells carry all four receptors, remain unknown; exhaustion was not measured, and whether agent-guided design beats expert design, or needs fewer rounds, is untested against a matched baseline.

By Liam Bai, Wujie Wang, Jason Yi

What happened: Enveda tested CDP, its Claude-based checker for drug manufacturing batch records, on four executed records from one clinical campaign: 1,061 largely handwritten pages. Human reviewers had raised 11 actionable issues; CDP raised nine as direct matches and two judged intended matches, across its runs rather than in one pass. Single runs on Claude Opus 4.7 added 7 actionable comments no reviewer had raised. A review holding CDP's comments took 9.4 times less time.

Why it matters: Batch-record review is required before a batch of a drug is released, and Anthropic's post says Enveda's manufacturing chemists had been spending up to 75% of their time on paperwork. This is the regulated-paperwork end of drug development, which rarely gets measured at all, and the yardstick here is what human reviewers caught on real records rather than a score on a benchmark.

Evidence and limits: The company's own report, not peer reviewed: four records from one campaign, checked against reviewers' comments, two of the four reviewer arms being outside consultants. No CDP comment cited content the records do not contain, and each run also posted one misread at Major, and the human reviewers could consult earlier-step records CDP was not given. Whether the 7 extra comments matter is Enveda's judgment. The time figure comes from a review holding CDP's comments in advance.

By Andrew Martin, Brandy Schenck, Daniel H. Wee, Viswa Colluru

What happened: On its Science page Anthropic said it has acquired Mithril Medicine, which was developing treatments for two rare childhood muscular dystrophies, and that its founder, Cody Tranbarger, has joined the company to continue that work and start further programmes. The same post reports research partnerships with six organisations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate:Biomedicines and The Column Group, which is setting up a company, Pleiades, to run drug programmes with Claude.

Why it matters: A model supplier has become a drug developer. The partnerships put Anthropic's own scientists inside other companies' programmes, and the acquisition gives it programmes it owns, so it now carries the risk and would hold the result. The post also says Anthropic plans to scale this work up substantially.

Evidence and limits: No terms were disclosed for the acquisition. Anthropic says that if the candidates continue to hold up it expects a drug to be ready to start human trials as soon as late 2028, which is its own forecast, and the mouse results behind it are in a technical report it published itself, not peer reviewed; that report studies two tool molecules in a pilot study, not a candidate. Pleiades has no disclosed programme yet.

What happened: Omalizumab, the only approved biologic for food allergy, neutralises circulating IgE but does not effectively remove the kind bound to FcεRI on mast cells. Guided by its research team, Claude ran 22 phases toward a two-armed antibody that would prise it off. Of 91 disruptor designs screened at one concentration, 20 beat every negative control; eight were purified and titrated, three of them removing more than omalizumab from 25 to 400 nM. Three of 57 anchor designs bound IgE, at apparent affinities of 3.4 to 235 nM.

Why it matters: Pulling bound IgE off the receptor needs a precise geometry, which display and immunisation libraries cannot specify, so designing from a structure is the point here rather than a convenience. Activity survived as many as eight substitutions from the parent, and hits came from every design scope and method the campaign tried. Divergence says the two arms came in weeks rather than the months screening takes.

Evidence and limits: Divergence's own report, not peer reviewed, all biophysics on a sensor chip: no cells, no animals, the two-armed molecule not yet made. In the titration the report says the variants did not appear to improve on the parent sdAb026, the best of them matching it at most concentrations, and Anthropic's post summarising the work does not mention the parent. The anchor affinities may be confounded by avidity, and the method needed an existing disruptor's sequence and solved structure.

By Ezra Erives, Emilio Ferrara, Bowen Jing

What happened: The model generates video together with music, ambient sound and interactive dialogue, which DeepMind calls its first real-time world model to generate audio natively, and it takes natural-language edits while a simulation runs, to change the weather, move the camera or add a person. It builds on the Gemini family, with Gemini Omni anchoring real-world appearance, and every output will carry a SynthID watermark.

Why it matters: DeepMind's stated purposes include developing training environments for robots and agents, and testing AI systems in different environments before taking them to the real world. One clip is generated from the prompt "Universal Robots UR5 holds a pipette and dispenses liquid into a lab plate, triggering tiny electric sparks", the only lab scene in a post whose testers are creative industries, journalism, education, gaming and engineering.

Evidence and limits: No benchmark, score or comparison with Genie 3 appears: it is clips, captions and an internal safety test with no result given. The two-minute consistency figure in press coverage is in neither post; neither gives a session length, and the post says object permanence can drift the longer a simulation runs. Trusted testers only, no public access.

By Raia Hadsell, Diego Rivas, Hang Qi

What happened: In a bioRxiv preprint, Paride Antinucci, Dominic Hall and colleagues at Biographica, a London crop-AI startup, feed a model a gene's DNA plus the measured levels of transcription factors, the proteins that switch genes on, in each tissue or condition. On held-out data from six plant species its median per-gene correlation across contexts is 0.831, against 0.486 for the same model without that input, and 57-246% above three fine-tuned genomic foundation models. It then designed edits of up to 20 DNA letters in soybean and tomato.

Why it matters: Many crop genes affect more than one trait, so an edit that helps in one tissue or condition can cause unwanted changes in others. A model that predicts expression per context lets a breeder rank candidate edits by the change wanted in one place and the change everywhere else, before any plant is edited. Biographica sells gene-editing targets to seed companies.

Evidence and limits: The edits exist only as predictions: none has been made in a plant and measured. The test data come from the public PlantExp collection and were scored by the company's own team, and the preprint is not peer reviewed.

By Paride Antinucci, Dominic Hall

What happened: Deputy Prime Minister Gan Kim Yong announced MARVEL-AI at SWITCH 2026: a platform led by A*STAR, starting with S$150 million of government money, in which AI models pick candidate materials and design experiments and self-driving labs make and test them. It is anchored at A*STAR's Institute of Materials Research and Engineering, whose executive director, Xian Jun Loh, also leads MARVEL-AI. Named partners include ST Engineering, Heraeus Electronics and Foxconn Industrial Internet on materials, and CuspAI and Google Cloud on the AI side.

Why it matters: The speech names the gap it is meant to close: models such as Google's GNoME and Microsoft's MatterGen propose materials, but a material still has to be made, tested and manufactured before industry can use it. MARVEL-AI puts the models and the labs in one national programme and opens its equipment, testing and researchers to companies, startups and SMEs included. CuspAI, which designs materials with AI, will connect its system to the platform's automated labs.

Evidence and limits: This is an announcement: the speech and reports give no split of the S$150 million, no timeline and no material yet. The figure of up to 470 alloy samples a day, more than 400 times conventional throughput, describes an existing lab process Enterprise Singapore cited, not a MARVEL-AI result.

What happened: Adam Lahouari, Jutta Rogal and colleagues at New York University and the Flatiron Institute studied acridine, a small molecule with at least eight crystal forms. Melt experiments solved the structure of form VIII for the first time. Their CrystalMath method kept 15 of 341,683 generated structures without computing any energy, and the 15 include all seven forms with known structures, form VIII among them. A MACE potential fine-tuned on acridine, with full free-energy calculations, reproduced the switch from form II to form III as temperature rises that experiments show.

Why it matters: One molecule can crystallize in several forms that behave differently, which matters for drugs, organic electronics and agrochemicals, and crystal-structure prediction with machine-learned potentials is how the field hopes to rank the forms before making them. Here the ranking step needed tuning: four off-the-shelf foundation potentials put the forms in the wrong order, and simpler free-energy approximations missed the observed crossovers. The study also grows acridine from the melt with 18 additives and finds the two strongest effects match computed defect energies.

Evidence and limits: One molecule. The predicted crossover between the two most stable forms sits near 290 K, against about 318 K measured, some observed transformations disagree with the free-energy curves, and form IV was left out of them. The additive link rests on two of the 18 additives. A preprint.

By Adam Lahouari, Jutta Rogal

What happened: Edgar Olehnovics, Matteo Salvalaglio and colleagues at University College London, AbbVie and the University of Colorado Boulder train a normalizing flow that maps molecular-dynamics snapshots of a crystal onto a simple reference state, which gives its free energy directly, without the chain of intermediate simulations the Einstein crystal method needs. On succinic acid and two drug molecules, Veliparib and Mivebresib, the results agree with that method, in crystals of up to about 3,000 degrees of freedom; across all systems it needed about 2.5 times less computing time on average.

Why it matters: Crystal-structure prediction can propose hundreds of possible packings for one drug molecule, and ranking them by stability at room temperature needs free energies. The standard method is accurate but too slow to run across a whole landscape. A method that needs one ordinary simulation per crystal form makes that ranking cheap enough to automate, and because it needs fewer energy evaluations it suits slower, more accurate force fields.

Evidence and limits: Checked against another computed method, not against experiment. The ranking depends on the force field: for Veliparib a generic force field and a tailor-made one disagree. The speed-up is uneven: for succinic acid it was faster only in the smallest crystals, and exact runtimes were not recorded.

By Edgar Olehnovics, Matteo Salvalaglio

What happened: Huifeng Zhao, Tingjun Hou, Yu Kang and colleagues at Zhejiang University and Peking University built GlycoWeave, which predicts the shape of a sugar chain bound to a protein by adjusting its bond torsions rather than moving atoms freely, which keeps the sugar's chemical structure intact. On 698 recent protein-sugar complexes, 71.9% had a high-quality pose among its top 25, against 19.2% for AlphaFold 3 and 16.9% for the docking tool Vina-Carb. Each pose took 0.34 seconds.

Why it matters: Sugars on proteins shape immune recognition, infection and how some drugs act, but they are flexible and thinly covered in experimental structures, so general structure predictors handle them poorly. Here only 41.0% of AlphaFold 3's top predictions passed the structural-validity check, against all 698 of GlycoWeave's (the docking tool Vina-Carb reached 99.7%). A model built for one awkward class of ligand measures how much accuracy a general co-folding model gives up on molecules unlike its training data.

Evidence and limits: Checked against known experimental structures, not new wet-lab work. The headline compares the top 25 poses; the paper's main text does not state AlphaFold 3's own sample count for that comparison. The protein pocket is given as input, so this is not blind docking. A preprint.

By Huifeng Zhao, Tingjun Hou, Yu Kang

What happened: Amaç Herdağdelen, Hailey Joren and Alex Tamkin had 50 analysts trained in causal inference judge a fictional farm programme with Claude Code. On each of four choices with two reasonable answers, such as how to measure productivity, a coin flip set which way the agent leaned, unknown to the analysts. They matched the lean 86% of the time, against 50% if it made no difference. Afterwards 77% gave the agent's arguments as their reasons in at least one answer; 15% credited it.

Why it matters: As agents take over the hands-on work of an analysis, the choices left to the researcher, what to measure and how to compare, are where judgment enters. This trial measures how much of that judgment moved to the agent: the coin flips alone predicted 82% of final recommendations, against 61% by always guessing the commonest one. It bears on any lab using agents to analyse experiments, and on review, since a reviewer working with an agent is an independent check only if the calls are their own.

Evidence and limits: The task was fictional and the agent was told to argue its lean, so real agents may sway people less. The 11 who worked without the agent made more serious errors, but they were fewer and less experienced, so the comparison is weak. The study is Anthropic's own.

By Amaç Herdağdelen, Hailey Joren, Alex Tamkin

What happened: The S-1, filed on 26 October, shows $16.3bn of revenue for the first half of 2026 against $4.6bn for all of 2025, and an operating loss narrowed to $693m from $2.1bn. The $80.7bn net loss is almost all a non-cash $72.4bn remeasurement of convertible notes, derivatives and SAFEs. Compute commitments include at least $114.5bn with Google, $110bn with Amazon and $44.9bn with Microsoft, owed regardless of usage.

Why it matters: For science, the filing says frontier AI 'may extend' to robotics and to scientific research, including drug discovery and materials science, and that Claude with the Model Hardware Standard can 'now' operate microscopes, liquid handlers and robotic arms. Its vision of 'Powerful AI' adds tasks running for weeks and 'compressing decades of scientific progress into years'. In the accounts, though, there is one segment, and the life sciences lab appears among the risk factors.

Evidence and limits: The June figures are unaudited interim ones. The three cloud commitments are three of seven named, with $125bn to $150bn more TPU compute estimated outside the table, and the aggregate line is blank. The top five customers were about a quarter of revenue. A Founder LLC's new share class carries 50.1% of the vote, and the safety trust's own class elects a majority of the board.

What happened: Jared Lewis's group at Indiana University, with David Baker, tested 96 designs; 63 expressed, and one scaffold called DE3 bound one mirror image of an iridium photocatalyst at 94% enantiomeric excess. As designed it gave low selectivity in catalysis. Screening mutations reached a 3:97 enantiomeric ratio, with turnovers above 50,000 in air and the protein reusable over three cycles.

Why it matters: Generative design supplied a pocket that holds a cofactor nature has no protein for, which is what the field could not get from existing folds. On this reaction the enantioselectivity came from screening mutations afterwards; a second design campaign later reached 90:10 on a different reaction with no evolution at all. Metal polypyridine photocatalysts had resisted both routes until now.

Evidence and limits: One scaffold carried through, out of five folds tested for binding. The pocket was designed around a ruthenium cofactor and bound only iridium analogues, and no cofactor-bound structure was obtained. Other reactions in the scope are only moderately selective. Designed and evaluated by the same groups. An accelerated preview.

By Sandip Mishra, David Baker, Jared C. Lewis

What happened: OpenAI's manuscript, dated 17 October and revised 25 October, went into its public repository on the 26th with a Lean formalization and gives a deterministic algorithm that counts the satisfying assignments of fixed-width CNF formulas in subexponential time. Anthropic's 49-page paper, dated 26 October and pushed on the 27th, claims the same thing and prints its author as Claude, Anthropic. It dates its model's discovery to 16 October.

Why it matters: The hypothesis has organised decades of lower-bound work: Anthropic's paper says a literature search by Claude found more than 800 papers proving a result of their own only under it or something stronger. If either proof holds, those conditional results have to be read again, and Ryan Williams, who is helping organise the review, says that if true it should move every computer scientist's weight a little toward P=NP.

Evidence and limits: Nobody outside either company has verified either proof. Lean checks that a proof follows from a formal statement, not that the statement means what the paper claims; Anthropic writes that down and adds that at the time of writing no person is known to have checked its Lean. OpenAI's scope note formalizes a weaker bound than its manuscript, on canonically encoded inputs, still subexponential. Anthropic calls it purely asymptotic.

What happened: OpenAI told three complexity theorists it had the result and would release it, and asked them to organise its examination. Russell Impagliazzo and Ramamohan Paturi at UC San Diego and Ryan Williams at MIT wrote to their field the same day: "Our involvement should not be taken as certification of the claim." Williams said on his own account that the group signed no NDAs and takes no pay, and that OpenAI agreed to release its raw data and sponsor a conference.

Why it matters: Who certifies a machine's result is what the last six weeks have been about, and here a lab named its examiners before releasing the claim. The letter is them refusing the role that implies while keeping their leverage: Williams says they asked for the raw data and a sponsored conference and OpenAI agreed, and the letter says the effort should confer ownership on no company and no chosen few, themselves included.

Evidence and limits: OpenAI's announcement page is 403 here, so its own account is not quoted; its history file calls the three an independent working group, naming no others. The physics result Al Jazeera reports as deferred has no public artifact, and OpenAI's 721-manuscript maths catalogue carries no quantum chromodynamics paper. The NDA, payment, data and conference points are Williams's words, not the letter's, which is undated.

What happened: OpenAI's 59-page paper, its author line reading OpenAI, dated 25 October and added to its public repository the next day, proves Sogge's conjecture for every spatial dimension at least two, with arbitrarily small Sobolev loss and none at p=2. From it the paper deduces the Bochner-Riesz conjecture, sharp bounds for Stein's square function, and the Kakeya and Nikodym maximal conjectures. OpenAI says it is releasing the model's output directly.

Why it matters: The published record had the plane from Larry Guth, Hong Wang and Ruixiang Zhang, and three-dimensional Kakeya from Hong Wang and Joshua Zahl last year; this claims every dimension at once, and cites two OpenAI preprints from September that it supersedes. Unlike the complexity result released the same day, the Lean here covers the headline claims: the local smoothing bound in every dimension, and that a set holding a unit segment in every direction has full Hausdorff dimension.

Evidence and limits: Nobody outside OpenAI has checked it, and OpenAI's own repository calls these manuscripts raw research data rather than polished papers. The restriction estimates for the sphere, paraboloid and cone hold only inside the conjectured range, so that conjecture is not settled here. The same repository withdrew three manuscripts on 7 October over a sign error.

What happened: A Roche team at Prescient Design and Large Molecule Research used a language-model pipeline to pull 89,074 antibody-antigen records out of 10,110 US patents, then built 11,247 pairs: an antibody a patent says binds, with no solved structure, against a sequence-similar decoy. Six structure models were scored. Protenix-v2 led at 72% of pairs and 87% on its most confident half, and picked the true binder out of 50 candidates 29% of the time. The best model touched a listed epitope residue in only 62.8% of cases.

Why it matters: Teams screening designed antibodies read a model's folding confidence as a stand-in for binding. This puts a number on what that is worth: much better than picking at random, nowhere near enough to name the one candidate to make. And telling a binder from a decoy does not mean the pose is right, so reading an epitope off these models is a separate claim these numbers do not support.

Evidence and limits: Nothing here was assayed: the binders are what patents claim, the non-binders are repertoire antibodies assumed not to bind, so the accuracy is against a constructed negative set rather than measured affinities. The structures were not in the models' training data; the patent text may have been. The preprint is CC BY and says the dataset is released, without stating its licence. Not peer reviewed, and the full text was not read.

By Daria Tupikina, Andrea Roncoli, Alexander Bujotzek, Brennan Abanades

What happened: Junkil Park and Aron Walsh at Imperial College London gave a composition-only synthesizability model a second task: rank known materials by the year each was first reported in the ICSD crystal database. On about 10,000 held-out reported materials it recovered 52.4%, against 47.0% for the model it extends and 48.4% with the years shuffled, all at one calibrated threshold where chance is 8.1%. Trained only on materials reported up to 2018, it put 7 later-reported compositions in its top 100, against 3 for the baseline.

Why it matters: Generative models now propose candidate materials far faster than labs can try them, so the question is which to make first. Synthesizability models usually treat every reported material as an equal example. This one uses the reporting order as free extra supervision, on the authors' reading that the order carries which precursors, methods and applications made each material reachable. The shuffled-year control is what ties most of the gain to the order itself rather than to the added task.

Evidence and limits: All checks run on existing databases. The authors say learning from history may favour chemistry already studied and score unexplored families lower. It reads composition only. The authors say the largest gaps come in the top 100 at the later cutoffs, where the statistical uncertainty is also larger. A preprint.

By Junkil Park, Aron Walsh

What happened: The startup released what it calls the first complete computer-checked proof of the classification, whose published proof runs to some 15,000 pages in over 500 articles. Dozens of agents, directed by its mathematicians and engineers, wrote about 43.1 million lines of Lean, following the first-generation proof and reusing 1.2 million lines from the Qiuzhen-CFSG project. Group theorists Richard Lyons and Ronald Solomon advised. The proof is public under Apache 2.0.

Why it matters: Results across algebra, number theory and combinatorics cite this theorem in a line; now every step under it can be checked by machine. The work also found 34 places where the published proofs needed repair. The largest, in John Walter's 1986 argument for the Unbalanced Group Theorem, was closed by another route: drawing on Solomon's 1976 work, the company's prover proved a new theorem that the team wrote up in a paper. Claude's Fermat's Last Theorem formalization in September ran to 13 million lines.

Evidence and limits: The figures and the list of repairs are Axiom's own, and no group theorist outside the project has reported checking the result; people still have to read the formal statement to confirm it says what the classification says. Building the proof takes close to a day and relies on changes to Lean's build tool that are not yet upstream.

By Dean Cureton, Vasily Ilin, Bhavik Mehta, Niels Voss

What happened: Sarath Menon, Braulio Sánchez Ortega and Ralf Drautz at ICAMS, Ruhr University Bochum, add a linear correction on top of GRACE, a foundation interatomic potential, fitted to free energies computed in simulation so that phase transitions land on the assessed SGTE values. Across 11 metals and 14 transitions, every corrected temperature falls within the set tolerance: 15 K for melting, 50 K for solid to solid. Gold melted at 844 K in GRACE and at 1346 K after correction. For titanium it took about 55 H100 GPU hours.

Why it matters: Foundation interatomic potentials learn from DFT data, and an error of a few meV per atom in which phase is more stable can move a melting point by hundreds of kelvin. This layer fixes the balance between phases without touching the base model: it needs only the energies and structures the model produces, not its parameters or training data, so in principle it fits any potential, open or closed. Copper's melting curve, fitted only at zero pressure, then follows measurements within 15 K up to 4 GPa.

Evidence and limits: The corrected temperatures are fitted to the SGTE values, so hitting them is calibration, not prediction, as the authors say. The layer can only rebalance phases the base model already describes and cannot add missing physics such as magnetism. Where the correction is large it moves other properties too, mostly towards experiment: liquid gold and silver become 3 to 6% denser, but vacancy energies move away from experiment for vanadium and aluminium.

By Sarath Menon, Braulio Sánchez Ortega, Ralf Drautz

What happened: At ESMO 2026, Melina Marmarelis of Penn's Abramson Cancer Center presented first human data for ISM6331, a pill Insilico designed with its Chemistry42 generative platform to block TEAD, proteins at the end of the Hippo growth pathway. By the 14 August cutoff 36 people had taken 15 to 130 mg a day. Of 22 evaluable mesothelioma patients on 30 mg or more, 4 had a partial response and 68% disease control; at 130 mg, 2 of 5. No dose-limiting toxicity; 8.3% had a drug-related grade 3 event.

Why it matters: TEAD has long been hard to block with a small molecule, and mesothelioma has few treatments once the first ones fail. A drug Insilico says its generative models designed is now shrinking tumours in people who had up to six earlier treatments, and falling blood levels of CCN2, a protein TEAD switches on, suggest it reaches its target in patients. Subject to further data, Insilico plans a randomised Phase 2a to choose the dose.

Evidence and limits: A single-arm Phase 1 still raising its dose, with no control group, and only the company's summary of the talk is public. The 40% headline is 2 of 5 people. The release does not say what the models did beyond naming Chemistry42, and a Phase 1 response rate is not evidence of longer life.

By Melina Marmarelis

What happened: Ruichen Li, Ji Chen, Weiluo Ren and colleagues at ByteDance Seed and Peking University split a real-space wavefunction into a spatial part, computed by a neural network (LapNet), and a spin part built to carry one exact total spin. Earlier neural network methods added a spin penalty to training, which can leave the wavefunction stuck mixing spins. The method matched experimental singlet-triplet gaps of biradical molecules about as well as the penalty method, gave an antiferromagnetic spin ladder for [Fe2S2] and a spin gap for an [Fe4S4] cluster.

Why it matters: Iron-sulfur clusters are central to biological nitrogen fixation, and their many spin states of nearly equal energy are where standard methods such as DFT, and even CCSD(T), can go wrong. With spin built in by construction, a calculation cannot drift into the wrong spin state, and chemists' own ideas (core and active orbitals, a first guess at how the spins couple) can be written into the network. With 10 active orbitals it reproduced the local spin trend on [Fe2S2(SCH3)4] that DMRG needed 32 to show.

Evidence and limits: Everything is computed. The iron-sulfur results have no experimental reference: they are compared with other calculations (FCIQMC, DMRG), and for [Fe4S4] the authors say the order of the three methods' gaps is not tied to any one cause. The November 2025 preprint called that gap more accurate; the published version drops the claim. Only the biradical gaps are checked against experiment.

By Ruichen Li, Ji Chen, Weiluo Ren

What happened: The Cabinet Office announced the AI Biosecurity Lab, an independent non-profit it calls the world's first dedicated AI biosecurity organisation, meant to detect biological threats earlier and speed up vaccines and other countermeasures. OpenAI is putting in £100m of compute and model credits through its Rosalind Biodefense programme, Anthropic up to £100m in credits, access and support. Google DeepMind and Basecamp Research join as research partners, not funders.

Why it matters: The £300m headline is £100m of public money; the rest is credits and access rather than cash, and the research partners give access, not money. It also puts frontier labs inside a national function: a National AI Biosecurity Initiative puts the lab with UKHSA, APHA and Dstl, and informs the MHRA's regulation of AI-designed medicines.

Evidence and limits: No budget breakdown, no timeline and no named staff; the scientific advisory committee is still to be announced. Anthropic's £100m is an 'up to' figure, and the page gives no period for either.

What happened: A Toronto, NIST and KAUST team compared deep-ensemble uncertainty against loss-landscape estimates: ALIGNN on band gap and formation energy, a multilayer perceptron on solubility. Calibration methods generalised only variably from in-domain to out-of-domain data, and calibrated uncertainties generally failed to cut the data an active-learning campaign needed on out-of-distribution data. Random forests and XGBoost on the same tasks showed the same poor uncertainty.

Why it matters: Active learning picks the next sample by uncertainty, and a campaign aimed at new chemistry is out of distribution by construction. The same failure in tree models as in the network points at the data: at least partly intrinsic to it, the authors say, rather than model capacity alone.

Evidence and limits: All computed, on three public property tasks. The claim is about these architectures and calibration methods on these datasets, not a proof that no calibration can work. Peer reviewed.

By Ashley S. Dale, Jason Hattrick-Simpers

What happened: Bristol Myers Squibb led; the company puts the post-money valuation at about $3B, up from $1.6B at its $365M Series D in May 2025. Pathos, a New York company co-founded in 2022 by Eric Lefkofsky and Ryan Fukushima and run by CEO Iker Huerga, licenses drugs other companies made and uses its agent platform, Foundry, to pick patients and adjust trials. BMS gets non-exclusive access to Foundry and three-year exclusive rights in neuroscience; Tempus will build a neurology patient dataset that Pathos licenses for $55M over five years.

Why it matters: Pathos puts its AI on the decisions that decide whether a trial succeeds: choosing which existing drug to run, in which patients, and when to change the trial. It does not design molecules. A drugmaker leading the round and taking access to the platform shows that one large pharma company will put money behind this use of AI, and the deal carries the method from cancer to Alzheimer's, a field with long trials and many failures. Simon Davies, who heads data and quantitative sciences at BMS, spoke for it in the announcement.

Evidence and limits: The round, the valuation and the deal terms are all from the company's two releases of 23 October; neither names the other investors or says whether BMS pays for its access beyond its stake. That Foundry identified JSKN016, an antibody-drug conjugate licensed from Alphamab, and has improved any trial's odds is the company's own account: no trial result from its approach has been published.

What happened: Bo Wang, Ci Chu and colleagues describe X-Automate in a companion post to X-Scientist: an off-the-shelf six-axis arm with a camera on it, machined grippers and control software written in-house, directed by the same agent system. Its first target is an Agilent HPLC, which has no motorised door and a drawer that shifts when touched. Over seven days in the company's own lab, a mobile robot and the arm loaded and analysed 18 plates, holding the HPLC busy 97.2% of the time against 53.6% in its model of weekday hours.

Why it matters: Protein-design groups can now make thousands of candidates a week, and the measurement that tells them which ones worked is the step that waits for a person to open a door. Xaira's answer is vision on the arm rather than taught positions, which is what lets one arm move between instruments that were never meant to be automated, and it sidesteps the instrument maker's paid interface. The company says adapting the arm to another instrument now takes weeks rather than months.

Evidence and limits: A company blog post, no paper, with the company reporting its own demonstration and calling the week a technical proof of concept. One instrument, one week, and the 80% is against a schedule the company modelled rather than ran. Four interruptions in plate storage upstream needed a person to recover them, with no errors it attributes to the arm or the agents. The plates were analysed but no scientific finding is reported.

By Bo Wang, Ci Chu

What happened: Hongjun Xiang and Xingao Gong's group at Fudan built an equivariant graph network that predicts both the charge density and the projector augmented-wave augmentation occupancies. Earlier universal models gave only a pseudo charge density, which cannot rebuild the electronic structure consistently. A trainable radial basis and a delta-learning step for charge transfer make it 50 to 150 times faster than earlier charge-density models; over 1,000 unseen materials its median band-energy error is 0.045 eV.

Why it matters: Charge density is what density functional theory solves for, so a model that reproduces it at the accuracy real codes use can stand in for the expensive self-consistent step rather than just predicting one property. The augmentation occupancies are what makes that possible here.

Evidence and limits: All computed, by the authors. The universal model covers non-magnetic structures. The speed is measured against two other machine-learning models and plane-wave DFT on one GPU. Peer reviewed.

By Xiwen Li, Hongjun Xiang

What happened: A Hong Kong University of Science and Technology team matched the crystals of 55 databases by structure into one set of 14.6 million crystals and 131.8 million property values, only 4.5% of whose crystal-property pairs carry a value. Given 100 verifications each, MatterGen conditioned on 400 GPa returned 100 stiff hits against screening's 55; but every generated hit contained osmium, 82 repeated a recorded composition, and screening's best hit was stiffer, 433.8 against 375.4 GPa.

Why it matters: A novelty share measures the reference set as much as the generator, and the usual reference is Materials Project alone. Against all 55 sources, GNoME's matches go from one to 18,234, its new share from 100% to 96.7%. The forward screen turned up 53,813 ultra-incompressible candidates with no recorded bulk modulus; 62 have been made, among them ReN2, measured at 428 GPa.

Evidence and limits: All computed, plus a literature audit: ReN2's 428 GPa is a published measurement, not a new experiment here, and the screen had predicted 349 GPa, 18% below it. The MatterGen runs are the authors' own, from released weights. A preprint.

By Zhilong Song, Zijun Pang, Lixue Cheng

What happened: An EPFL and Fribourg team fine-tuned the PET-MAD potential on r2SCAN data and ran large-scale molecular dynamics on Li6PS5Cl, screening carboxylic, sulfonic and phosphonic acid coatings by chain length and coverage. Simulation said lithiated head groups adsorb more strongly and that long-chain lithiated carboxylates form dense brush-like hydrophobic layers, with protection setting in near 8 Å. Infrared, impedance and H2S measurements followed.

Why it matters: The loop half closes: simulation set the chain length, and the H2S measurements run the same way, longer carboxylate chains suppressing it better. It also names the conflict rather than hiding it, since a thicker coating blocks lithium transport, so the best coating is a middle length.

Evidence and limits: One electrolyte, Li6PS5Cl, and one potential fine-tuned by the authors. The 40 to 55% figure is what is left after processing, not a gain. The experiments cannot rank the head groups, and the paper says so. A preprint.

By Zhiyi Wang, Leonie Braks, Hanna Türk, Davide Tisi

What happened: Tetsuya Yamada and Henrik Kaessmann at Heidelberg, with Svante Pääbo at the Max Planck Institute for Evolutionary Anthropology, predicted the regulatory effect of every single-nucleotide substitution on the human lineage across 305 tissues and cell types. Most of the 71,345 gains sit outside the regions evolution researchers had flagged as fast-changing, and more often than not they come from a regulatory element being reused in a new tissue rather than a new one appearing.

Why it matters: The question of which non-coding changes made humans human has been stuck on data nobody can get: great-ape tissue panels are thin, and no functional measurement is possible on an extinct hominin at all. A sequence model sidesteps that, because it scores a reconstructed ancestral allele as readily as the living one. The payoff here is the archaic comparison: nearly every gain now fixed in humans was already there before the split, so the modern-human-specific ones are few, and the paper names which ones move gene expression most.

Evidence and limits: The map is prediction, not measurement. Where it could be checked, against accessibility measured in four great apes' cell lines, AlphaGenome beat Enformer and ChromBPNet and 82.7% of its strongest calls held up, against 31.1% of controls, but it recovers only a quarter to a half of the measured human-specific changes depending on the threshold, and misses clustered gains more often than isolated ones. A preprint.

By Tetsuya Yamada, Henrik Kaessmann

What happened: Jihan Zhou's group at Peking University put a two-stage network after the reconstruction step: a ResUNet-3D variant corrects distortions across the volume, then a set transformer reads the shape of the electron density around each atom to name its element. Trained on 17,340 simulated samples and benchmarked on 42,588, at typical noise it cut the median coordinate error from 0.244 to 0.100 angstrom and raised atom recovery from 93% to 99%, elements called right 99.5% of the time.

Why it matters: Electron dose is what has kept atom-by-atom tomography to materials that survive the beam, which leaves out most of what people want mapped. The gain here is in post-processing rather than in the microscope, so a tilt series already sitting on disk can be re-read: thinned to a sixth of its dose, a published Pd@Pt series matched the full-dose conventional reconstruction on peak sharpness and repeatability.

Evidence and limits: No new experimental data were taken for this paper: the one demonstration on real data thins a published series, halving its projections too. Every other number is simulated, and the claim that this reaches beam-sensitive materials such as perovskites rests on simulation alone. Peer reviewed.

By Yao Zhang, Lanyi Cao, Zhen Sun, Jihan Zhou

What happened: A KAUST and Karolinska team ran image-based, transcript-based and multimodal segmenters, from Watershed to Cellpose, Baysor and three versions of Segment Anything, over 15 datasets from 5 platforms and 12 human and mouse tissues. On their overall score Segment Anything led at 0.58 and Watershed came last at 0.27, but count how many cells each found and SCS leads instead; the ranking also reorders by dataset: SAM did well where cells look distinct, SCS where they look alike.

Why it matters: Segmentation is the step every spatial result stands on, because it decides which transcript belongs to which cell. This benchmark shows the choice carries all the way through to cell-type proportions, how much signal bleeds between neighbours, and inferred ligand and receptor pairs. So a published spatial finding depends on a step with no standard, and the pipeline released here is an attempt to make the comparison repeatable.

Evidence and limits: There is no true cell boundary to score against, so the metrics are biological proxies: expression purity, marker genes, cell-type proportions against a reference. The authors say platform design, image quality and gene panels cannot be fully separated from the algorithms, and that their gene-set sensitivity test ran on one dataset only. Two of the 11, Swin-S and Cascade Mask R-CNN, were left out of the scored comparison. Peer reviewed.

By Xabier Martínez-de-Morentin, David Gomez-Cabrero, Jesper Tegner

What happened: At ESMO 2026 Iambic gave updated Phase 1/1b data for IAM1363, the HER2 inhibitor its own models designed. By the 31 August cutoff 122 people had received it. Among 53 with targetable HER2 changes at 960 mg or above, 36% had a partial response and 38% stable disease; of 22 whose tumours had reached the brain, 9 responded and 12 were stable. Of the 68 at the chosen 960 mg daily dose, 5.9% had a drug-related grade 3 event and none a grade 4 or 5.

Why it matters: A drug a company says its models designed, and took from project start to the clinic in about two years, is now in its second year of human testing. These patients had been treated a lot already, a median of three regimens, 61% with trastuzumab deruxtecan and 36% with tucatinib, which is the context the response rates have to be read in, and the brain figures matter because those patients have few options. Iambic's prospectus puts about $75M of its $135M net IPO proceeds into this drug.

Evidence and limits: A single-arm Phase 1/1b with no control group, and only the company's summary of a conference poster is public: no median duration of response, and the durability it highlights is named patients. The rates come from subsets it chose, 53 of the 122 and 22 for brain tumours, and the intracranial responses after both prior drugs are three people. The release says nothing about what the models contributed, and a Phase 1 response rate is not evidence of longer life.

What happened: Lightspeed led; Premji Invest, Mayfield, IVP, Lux, Conviction, Menlo Ventures and 12 other firms joined; angels include OpenAI's chief research officer Mark Chen and its CFO Sarah Friar. The South San Francisco company is a contract research organisation: it will make variations of a lead molecule for drug hunters, with chemists guiding the work. Its co-founders are CEO Kevin Weil, chief scientific officer Sol Reisberg (medicinal chemistry, Terremoto Biosciences), CTO Manohar Paluri (Meta Superintelligence Labs, a founding member of FAIR) and COO DJ Patil, the first US chief data scientist.

Why it matters: Weil's case is that AI has sped up software and mathematics but not chemistry, because chemistry moves only when someone makes a compound and tests it. A lab that sells that work to drugmakers also produces the experimental data better models need, and Warp says it will have no drugs of its own. The other half of the pitch is location: the release cites the BIOSECURE Act, signed at the end of 2025, which frames limits on federal procurement and funding tied to designated biotech companies, and gives customers more say over where their research and IP sit.

Evidence and limits: The round is the company's own announcement, which names no stage. The $865M post-money valuation comes from press reports citing the company and Weil, not from the release. Nothing is automated yet on the public record: the release describes labs being built and a first service guided by chemists, and the site names no customer, only a small number of design partners.

What happened: Google DeepMind and Google Research, with the World Resources Institute and IIASA, released Forest Typology 2020, which labels every 10-metre patch of land as one of six types, from primary forest to plantation, tree crops and other land. Built on AlphaEarth Foundations embeddings, it is 90.2% right overall against 8,190 expert-labelled sites, and is open on Earth Engine.

Why it matters: The EU rule turns on what a plot was in 2020, not just whether it had trees, and a tree-cover map cannot tell a natural forest from one logged and replanted as plantation. This map makes that distinction.

Evidence and limits: The 90.2% is carried by the large other-land class; per class, F1 runs from 82.9% for primary forest down to 58.2% for planted forest. Accuracy was checked on the Geo-Wiki platform of IIASA, a co-developer.

By Maxim Neumann, Anton Raichuk

What happened: Harsha Nori, Eric Horvitz and colleagues at Microsoft AI built a benchmark where a case starts from a short presentation and the diagnostician must ask questions and order tests, each with a cost, before naming a diagnosis. Their orchestrator, MAI-DxO, has language-model agents play a panel of doctors. It was right on 83.2% of New England Journal of Medicine clinicopathological cases and 80.3% of MIMIC-IV emergency admissions; generalist physicians scored 36.1% and 46.8%. Mustafa Suleyman, Microsoft AI's CEO and a co-author, called it a step toward medical superintelligence.

Why it matters: The benchmark, first posted as a preprint in June 2025 and now peer reviewed, tests diagnosis the way a doctor works: not picking an answer from a full case write-up, but deciding what to ask and which tests are worth paying for. Scoring cost beside accuracy lets a system be judged on whether it orders fewer tests, not only on whether it gets the answer. The Science version adds emergency-department cases and newer models.

Evidence and limits: The cases are published records run in simulation; no patient was seen. The doctors worked alone on cases usually handled by specialists, who were not compared. TechCrunch reports much of the gain comes from the underlying model: GPT-5 alone scored 78.9% on the hard cases, and on emergency cases the orchestrator added little.

By Harsha Nori, Eric Horvitz

What happened: Carbon-X, built by Hugging Face with Zhongguancun Academy and TIGEM in Naples, starts from Qwen3-14B, is trained on 302 billion more tokens (55% of them raw DNA), then tuned to answer questions about sequences pasted into a chat. On five DNA benchmarks it averages 66.7%, against 67.5% for GPT-6.1 Sol and 60.9% for Opus 5.5, and leads every model tested on coding-variant effects (74.9%) and next-base prediction. The weights are out under Apache 2.0 and run on one 80 GB GPU.

Why it matters: DNA models usually output scores a scientist has to interpret, while a chat model may explain what a gene does yet struggle to judge a change in its sequence. Carbon-X puts both in one open model a lab can run on its own hardware, so unpublished or patient sequences need not go to an outside API. The team estimates that scoring ClinVar's 2.4 million variants of uncertain significance would cost about $300 with Carbon-X, against $32,000 to $4 million with the frontier models it tested, a range that wide because it spans models and reasoning settings.

Evidence and limits: Every score is the team's own, and it built two of the five benchmarks. Nothing was measured in a lab or clinic, and the team says the model is not clinically validated. Frontier models still lead on insertions and deletions, on single-base variants in long sequences and on sequence-feature questions, where Carbon-X can also miscount bases. The costs are not like for like: Carbon-X at measured throughput on a $5-an-hour H200, the frontier models at list API prices.

By Loubna Ben Allal, Diego di Bernardo

What happened: The company's research note and preprint set out a four-step cycle: train compact experts on separate skills, break each answer into checkable requirements and accept an expert's contribution only where it beats the current one, write the accepted parts into the target model, then specialise again from it. In the 4B and 8B settings it ranks first among the methods compared on all five scientific benchmarks, scientific reasoning at 55.0% against GRPO's 53.0%. With 8B experts and a Qwen3.5-27B target, all seven benchmarks improve, the mean from 87.3% to 94.0%.

Why it matters: Most science models are trained once and then stand still, and a lab that wants a new skill in one trains a new model. The claim here is that a smaller model's strengths can be moved into a larger one, which the 8B-to-27B run is the test of, and that the filtering is what makes it work: without the verification step the same supervision drops scientific reasoning from 55.0% to 50.6%, and scattering it at random costs 6.3 and 12.1 points. The open 8B checkpoint lets someone outside the company check what the recipe produced.

Evidence and limits: Benchmark scores only, all of them the company's own, scored in-house and not peer-reviewed. The note says its own report does not show that repeated cycles keep accumulating gains, and the 27B run also changes model generation, which it reports separately; the model card says the open checkpoint renames and republishes an existing one rather than a fresh training run. The rankings are first among the methods compared, not against the field.

By Xun Xu, Zaixi Zhang

What happened: Michele Ceriotti's group at EPFL applied the Point-Edge Transformer, which drops explicit O(3) algebra for edge-to-edge attention, to molecular dipoles and polarizabilities, NMR chemical shielding tensors and auxiliary-basis expansions of the electron density. A bigger backbone improved accuracy and equivariance error together. It matches or improves on the released models at lower inference cost, the density only when trained on a better-posed target.

Why it matters: Strictly equivariant networks pay for their guarantee in tensor products that grow steeply with angular order, and can fail to represent a target if truncated. Targets like shielding tensors and basis expansions are where that bites. Learning the symmetry instead removes the trade-off, if approximate equivariance is good enough.

Evidence and limits: All computed, by the authors. Equivariance here is approximate rather than guaranteed, and the authors say only that it affects most observables little, not all. Electron density is the weakest case: its equivariance error grows with angular order, and above order two it falls with backbone size no faster than the symmetrized error does. A preprint.

By Joseph W. Abbott, Michele Ceriotti

What happened: Gus Hart at Brigham Young University and Jacob Pursley at the DOE's Kansas City National Security Campus tested 14 universal machine-learned potentials on 1.45 million ordered intermetallics against AFLOW, a database none of them trained on. For half the structures AFLOW, Alexandria and OQMD differ by under 6 to 8 meV per atom, but a minority differ far more, raising the average to 16 to 30. Re-referenced, the best potential agrees with AFLOW as closely as Alexandria does.

Why it matters: Leaderboards rank these potentials by formation-energy error against one database, as if it were ground truth. The disagreement between independent databases sets a floor under any such number, and the top models now sit at it. Screening out broken elemental references aligns the energy and stability rankings, and the leader's 7 meV per atom margin shrinks to 0.4 when both are scored at a common geometry.

Evidence and limits: All computed: this compares databases and models, with no new measurement. Intermetallics only, chosen as a region where DFT is relatively accurate. A preprint.

By Gus L. W. Hart, Jacob N. Pursley

What happened: Statisticians at Karlsruhe and Marburg put Pangu-Weather, GraphCast and Aurora against ECMWF's HRES under one experimental design. HRES came out ahead when scoring was restricted to events that actually happened; the AI models came out ahead under weighted scoring functions. A contingency table shows why: the AI models forecast extremes less often than they occur, so misses dominate their scores, while HRES forecasts them more often than the AI models do, so false alarms dominate its.

Why it matters: Published studies disagree about AI models on extremes, and this says the disagreement is largely in the protocols rather than in the models. The decomposition puts the score gaps down to calibration, not discrimination: every model tells an extreme from a normal day about as well, Pangu-Weather least of them, and the table shows the AI models are simply too conservative about saying so. That is a fixable problem, and a different one from not seeing the event.

Evidence and limits: All computed, against the analysis that initialised all four models, which the authors say slightly favours HRES; station recordings would test it. One physics model and three AI models, on the variables and lead times chosen here. A workshop paper at NeurIPS 2026, so shorter and less reviewed than a journal article.

By Tobias Biegert, Sam Allen, Zhongwei Zhang, Sebastian Lerch

What happened: Chengming Zhang and Luonan Chen, with Kazuyuki Aihara at the University of Tokyo, split the latent space of a hierarchical model into causal and spurious factors and apply do-calculus to the causal half, then recurse to rebuild the regulatory layers. On synthetic data, stem-cell differentiation, a CRISPR activation screen of 31 single and 26 paired conditions with some held out, a type 2 diabetes multiome and gut microbiome data, it beat the methods compared on causal triggers in most settings.

Why it matters: Most regulatory networks read off single-cell data are built from correlation, so a gene that moves with a state change is indistinguishable from one that drives it, and that is the difference between a target worth perturbing and one not worth the bench time. Encoding the intervention itself, rather than the association, is a different route to the problem, and the output is a ranked list of candidates to test.

Evidence and limits: All computed on public datasets, with no new experiment. The authors say the inferred triggers may still contain spurious biological signal, that the in silico perturbations are a qualitative read of direction rather than a predicted magnitude, and that simple additive baselines still do well on part of their own perturbation benchmark. Its results also depend on the regulatory hierarchy assumed. Peer reviewed.

By Chengming Zhang, Deyu Cai, Kazuyuki Aihara, Luonan Chen

What happened: A Nanjing University, CUHK and NUS team split the problem in two: generate an RNA shape in a pretrained structural latent space, then predict the contacts that place it against the protein, rather than generating shape and pose together. On 124 protein-RNA complexes held out by date, it leads ODesign, RFDpoly, RNAFlow and NACraft on complex recovery, with DockQ 0.131 against 0.116 for the next best. Twenty of the 124 clear DockQ 0.23.

Why it matters: RNA that binds a chosen protein is a route to reagents and to drugs, and unlike protein binder design it has stayed out of reach, partly because so few protein-RNA structures have been solved. What this paper pins down is where the state of the art stands when the binding site is withheld: five methods, and the best of them recovers the native pose on one target in six.

Evidence and limits: All computed, no synthesis and no binding measurement, and the authors say structural recovery establishes neither affinity nor activity. The headline ranking uses the native interface after generation to pick which of RiboBridge's own pockets is scored, so it is not a blind result. RiboBridge also clashes more than two of the baselines and does not lead on sequence recovery. A preprint.

By Zhou Zhang, Hanqun Cao, Tianfan Fu

What happened: In INTerpath-001, 1,137 patients with resected stage IIB-IV melanoma were randomised 2:1 to intismeran, an mRNA encoding up to 34 neoantigens chosen from each patient's tumour sequencing, plus KEYTRUDA, or to placebo plus KEYTRUDA. After a median 24.6 months, two-year recurrence-free survival was 82.2% against 70.0% (HR 0.58), and the risk of distant metastasis or death fell 38% (HR 0.62). The data go to ESMO 2026's Presidential Symposium and are accepted by NEJM.

Why it matters: Until now only the topline was public. The two arms part only after six months, once the course of up to nine vaccine doses is given: the hazard ratio is 0.99 in the first six months and 0.42 after, in a prespecified sensitivity analysis. That is the timing expected if the benefit comes from an immune response the vaccine builds against each patient's chosen neoantigens. Merck says it is now talking to regulators.

Evidence and limits: Merck and Moderna released these numbers ahead of the talk and the paper. Overall survival is immature: the release reports a 'favorable trend' and no number. The trial tests the whole product, so it cannot say how much the neoantigen selection adds over another way of choosing antigens. Grade 3-5 treatment-related events were 14.2% against 13.4%.

What happened: RFD4-Proteina, from the University of Washington Institute for Protein Design and NVIDIA (David Baker corresponding), designs proteins, DNA, RNA and enzymes in one shared latent space, using reasoning and guided search rather than a separate model per task. On AlphaFold3-refolded designs it reports a 10.6% pass rate for protein binders against BindCraft's 5.9% at 42 times the throughput, and higher rates than every method tested on small-molecule, enzyme-scaffold and DNA-binding tasks. In the lab, 158 of 188 designed TrkA binders bound below a micromole.

Why it matters: Protein design has used a different specialist tool for each job: backbones, binders, enzymes, DNA. RFD4-Proteina puts them in one model that also covers modified amino acids, sugars and nucleic acids, and still comes out ahead on all four. The part that matters is the wet lab: designs across many target classes were made and measured, including a designed enzyme that converted 79% of an intermediate in making the HIV drug lenacapavir, a de novo catalyst doing a real step of drug chemistry, further than designed enzymes had reached.

Evidence and limits: A preprint, not peer reviewed, and all the wet-lab work is the authors' own, so nothing is independent yet. In silico a 'pass' means an AlphaFold3 refold matches the target (a model checking a model), and the pass rates are low in absolute terms (3.5% for DNA binders). The binder and enzyme successes are counts from the authors' own assays, not an outside lab's, and the enzyme result is a single proof-of-concept.

By Nathaniel Corley, Kieran Didi, David Baker

What happened: A. Ali Heydari, Daniel McDuff, Xuhai Orson Xu and colleagues at Google Research built an orchestrator over three Gemini 2 agents: one writes and runs analysis code on Fitbit and blood-test data, one checks health knowledge against sources, one coaches. The data came from the WEAR-ME study of 1,165 people. Across 11 tasks and 1,350 hours of rating, 45 users put it first 46.7% of the time; three health experts scored its overall quality 72.8%, against 64.3% for the three agents answering separately and 54.4% for one agent.

Why it matters: Questions about your own wearable data need code and statistics, while coaching needs a different way of talking, and the authors found one model struggles to balance both. Splitting the work let each part be tested on its own: the data agent's analysis plans scored 75.6% against 53.7% for plain Gemini, and its code passed 75.5% of expert unit tests on the first try against 58.4%. The framework and its benchmark are open under Apache 2.0, so other groups can swap in their own models.

Evidence and limits: Everything was rated in conversation, not measured in health: the live users played study personas instead of asking about their own data, and no one tested whether behaviour changed. The data are Fitbit only, some scoring used model raters, and a response took 205 seconds and 6.5 model calls, against 36 seconds for one agent. The authors call it a research prototype, not a diagnostic tool.

By A. Ali Heydari, Xin Liu, Daniel McDuff, Xuhai Orson Xu

What happened: Mirendil reports that one researcher built Mirendil-Base from an empty repository: agents wrote over 99% of the code, ran the experiments and diagnosed failures, while the researcher chose what to try and what to trust. It learned only from public code, web pages and datasets, with no distillation, and the cost covers all compute, the agents' tokens included. On seven code repositories published after training it scored 0.222 bits per byte (lower is better), behind Kimi-K3's 0.193 and ahead of GLM-5.3, DeepSeek-V4-Pro-Base and Nemotron-550B-Base.

Why it matters: Pretraining a competitive base model has taken a team of specialists in architecture, data and chips. If one person directing agents can do it, the scarce inputs become compute and judgment, and a group whose expertise is in its own science, such as a robotics lab or a hospital, could train a model on its own data. It is the first result from a company that raised on the pitch that labs using AI in biology or chemistry should not have to become AI labs.

Evidence and limits: Every number is Mirendil's own, in a blog post with no paper, no weights and no named researcher. The only score is how well the model predicts unseen code, not whether it solves coding tasks, and Mirendil notes that Kimi-K3 and GLM-5.3 were scored after post-training, which usually raises this loss slightly. The two larger models are still in training.

What happened: Jeonghyeon Kim and Philip A. Romero connected GPT-5.6 Terra to a standard Opentrons OT-2 pipetting robot and a webcam through PEARL, five model roles sharing one record of the experiment. Given only a goal, it refused an impossible instruction in all ten runs, read from the colours that a human had turned the plate 180 degrees mid-run, and in seven of ten diluted dyes to see their colour first. Asked to characterise an enzyme assay, it set concentrations and timings, discarded a spiked well, and repeated runs missing early readings.

Why it matters: Lab automation runs the procedure it was given, and a moved plate or a saturated reading breaks it unless someone wrote a check for that case first. The paper's own comparison locates the behaviour: in a plain tool loop none of the 80 execution runs passed, one agent keeping a persistent record of the experiment passed 60%, and the five-role version passed all 80. So the judgment is mostly the general model's and the scaffolding buys reliability, which is what would let a lab with an ordinary liquid handler try this rather than build a platform.

Evidence and limits: Eight scenarios on one bench, run by the authors: colour mixing and one horseradish peroxidase assay, whose kinetic constants matched their own human-run reference and previously reported values, not an outside lab's. The scoring of planning, adaptation and execution is theirs, on their own rubric, and every perturbation was one they introduced, so nothing tests what a working lab meets unplanned. Colour was read off a webcam calibrated against a plate reader. It is a preprint.

By Jeonghyeon Kim, Philip A. Romero

What happened: AlphaVariant pairs a sequence generator trained on related natural proteins with activity predictors retrained after each round of results. On savinase, a detergent protease, two rounds produced libraries in which 75.4% and then 83.8% of the variants tested beat wild type, as the positions it could change went from 25 to 67; its best variant was 7.1 times more active than wild type on a synthetic substrate. A third round also scored heat tolerance and patent distance, and of 32 variants built, three beat the commercial protease Uno 100 L.

Why it matters: Directed evolution gets its gains mostly one position at a time, because the combinations explode and the measured data is far too sparse to fit a model over them. Here the useful variants carry changes at eight to sixteen positions at once, and the final round needed only 32 measurements, which is the scale at which a design method is worth a company's screening budget rather than a benchmark table. Putting patent distance in the reward is the unusual part: freedom to sell was optimised alongside performance, and it is a constraint any enzyme has to satisfy.

Evidence and limits: Every assay and the wash test were run by the authors with the industrial partner Bestzyme, which has filed a patent application on the variants; no outside lab has tested them. The wash test is a Chinese standard method on one soiled-cloth type, not a product trial. The 7.1-fold figure is kinetic fitting on a peptide substrate, and round three used casein. On the six public landscapes the clear wins were the three scored by a surrogate. Preprint, CC BY-NC-ND.

By Xiaopeng Xu, Jie Wang, Xin Gao

What happened: Bo Wang, Ci Chu and colleagues describe X-Scientist on the company's blog: a lead agent plans and checks the work of worker agents that call bioinformatics tools and Xaira's own models. It nominated 381 genetic perturbations to undo what TGF-beta does to HepG2 liver cells; the company also tested 321 its own scientists picked and 433 drawn at random, measured all of them by Perturb-seq, and counted a hit at three standard deviations back toward untreated cells. Nominating took 20 minutes against more than three hours by hand.

Why it matters: TGF-beta drives scarring in chronic liver disease, and which genes to perturb against it is the kind of choice a biologist spends a morning on. The measurement here is of the model's picks, not of the model's own score: every nomination was made, perturbed and read out in cells. Three of its hits outside the known TGF-beta pathway beat the median of the ones inside it, and none were in the scientists' list, which is what the loop is for, since a list of known pathway genes is one a scientist already has.

Evidence and limits: A company blog post, no paper and no peer review, and Xaira both chose and scored the comparison. One cell line, one exposure out of the twelve it ran, and the 7.1% against 6.9% gap is well inside what 381 and 321 perturbations can resolve. The benchmark scores beside it compare harnesses on the same models, not biology. Nothing has gone back into the models yet.

By Bo Wang, Ci Chu

What happened: AI Proteins will design and optimise miniproteins from scratch against several targets Astellas picks. Astellas, through its affiliate Astellas Innovation Management, takes an exclusive worldwide licence to any candidate and runs development, regulatory work and sales. Upfront and milestone payments come to more than $400M, plus royalties. Chris Bahl, the Boston company's founder and CEO, and Morten Sogaard of the Astellas Innovation Lab spoke for the deal.

Why it matters: This is AI Proteins' second large-pharma deal, after Bristol Myers Squibb's in late 2024, also valued at up to $400M. Miniproteins are small proteins built from scratch rather than taken from nature. Bristol Myers Squibb took an option on a licence; Astellas takes the exclusive licence outright, so this time the designs are bought as future drugs from the start.

Evidence and limits: The release does not break out the upfront payment or say how many targets are involved, and no designed molecule from the deal is shown. One trade report described the $400M as upfront; the release counts it as upfront and milestones together.

What happened: Fengyuan Dai, Fajie Yuan and colleagues at Westlake University trained a 16-billion-parameter model on 1.7 billion protein-text pairs: a description of the function wanted goes in, a sequence comes out, with less hand-written structural specification. Four tasks were run, a fluorescent protein, a PET hydrolase, an alcohol dehydrogenase and a metabolic H-protein, and each gave the intended function, with catalysis measured for both enzymes. In a reconstituted CO2-fixing reaction a designed H-protein raised product yield 1.7-fold over the natural E. coli one.

Why it matters: Protein design normally starts from a natural protein to optimise, or asks the designer to restate the goal as geometry and sequence rules a model can take. Here the goal is a sentence, and what came out of it catalysed reactions rather than only folding as predicted. The H-protein is the sharper part: a designed carrier protein beat the natural one in a reconstituted CO2-fixation reaction, with no rounds of experimental optimisation in between.

Evidence and limits: The per-task numbers are in the extended data and they are small: 94 PET hydrolase designs assayed out of 2,416 filtered gave one hit, 8 alcohol dehydrogenases tested gave 5, and 20 H-proteins picked from about 1,000 gave 5 active. All the wet-lab work is the authors' own. The inference code is Apache 2.0 and the weights repository states no licence; the full training corpus was withheld for biosecurity and given only to the reviewers.

By Fengyuan Dai, Fajie Yuan

What happened: Fajie Yuan's group at Westlake University, with HKUST (Guangzhou), the Shenzhen Institute of Synthetic Biology and Zelixir, trained Pinal, a 16-billion-parameter model, on 1.7 billion machine-written pairs of a protein and a description. It drafts a structure from the words, then a sequence to fit it. Designs for a fluorescent protein, a PET-degrading enzyme, an alcohol dehydrogenase and an H-protein each showed the function asked for; in a CO2-fixing pathway the designed H-protein gave 1.7 times the product of E. coli's own, with no lab optimisation.

Why it matters: Design models typically start from a natural protein to improve, or need a person to turn the goal into shape and sequence constraints first. Pinal takes the goal as a sentence, which is a different interface: the biologist says what the protein should do and gets candidates to test. The weights are on Hugging Face and the code is Apache-2.0, so the claim can be tried by anyone.

Evidence and limits: Each hit came out of a funnel the headline hides: about 151 designs assayed across the four functions, out of thousands generated, with the PET-degrading enzyme leaving one standout among 94 tested. The measurements are the authors' own, and the training corpus is withheld over dual-use risk. Fajie Yuan told Science and Technology Daily that language-driven protein design is still at the proof-of-concept stage, with finer instructions, such as setting a fluorescent protein's emission band, open for the field.

By Fengyuan Dai, Shiyang You, Tong Si, Fajie Yuan

What happened: OpenAI says it has acquired Orchestra Bio, a platform for coordinating drug development, and is bringing its team and technology into Rosalind. Its own example: a failed manufacturing batch can delay the start of a clinical trial, and Orchestra's agents work out which activities and timelines that hits, gather the relevant information and compare ways to respond, after which scientists and programme leaders decide. The same page names Novo Nordisk and Moderna as Rosalind partners.

Why it matters: Rosalind has been a model and a workbench; this adds the layer that tracks what a programme owes and when. OpenAI puts the case in its own numbers: roughly 10 to 15 years on average from target discovery to approval in the United States, with teams coordinating across internal departments and outside partners on disconnected systems. So what is being sold here is coordination rather than prediction, and the division of labour is stated outright: agents monitor and track, people decide.

Evidence and limits: No price, and OpenAI's page names nobody from Orchestra. The reported funding, about $20 million in total, reaches here through Decrypt citing Caplight rather than from the company, and the seed round's size is not given. Neither company says how many customers Orchestra has, what happens to existing contracts, or when the agents reach Rosalind. Orchestra Bio is not Orchestra BioMed, the unrelated Nasdaq device maker.

What happened: Aniketh Iyengar, Benjamin Siranosian and colleagues report it on the company's blog. Omnii-1 rewrote the DNA for gp17, the tail-fibre protein the T7 phage uses to recognise E. coli. At 11 to 15 amino-acid changes from the natural protein, 69.9% of its designs were recovered after assembly against 4.8% of random controls, and 2,161 of 6,694 variants with measurable activity were enriched above the natural phage. A cocktail containing its designs cleared one patient's resistant bacteria in a dish, then was given to that patient.

Why it matters: A model's design has gone into a person's treatment, not only into a dish: the company says this is the first clinical use of AI-generated phages, under an authorised, physician-led compassionate-use pathway. What the model wrote is one region of a known T7 genome, not a genome. It also ranks which designs to test against a given bacterial strain (on the hardest of three library strains, 24% of its top 100 picks beat the activity threshold against 4.5% of the library), and that ranking is the step a therapy matched to one patient has to get right.

Evidence and limits: A company blog post, no paper and no peer review, and every measurement is the company's own. One patient: no clinical course and no outcome, citing privacy, and the company says a single patient cannot establish safety or efficacy, so nothing here tests whether the treatment worked. The post does not name Tabula, which The New York Times later reported had done much of the lab work and supplied some of the cocktail's phages.

By Aniketh Iyengar, Benjamin Siranosian, Arpan Sarkar, Avery Noonan

What happened: The team chose five toxins; from one prompt and four follow-ups, an internal research model then worked mostly on its own for 64 hours, picking the sites to aim at and generating 318,484 candidates, 211 of which were made. Adaptyv Bio tested all of them for binding, Twist Bioscience 210 against four toxins. Seventeen met Anthropic's criteria, 15 of them at both labs. The tightest, against the Mexican toxin Cn2, was single-digit nanomolar. All 211 and their data are released.

Why it matters: This is the company's earlier binder work pushed at a target nobody has money to work on (antivenom for a neglected disease), and the loop is what to look at: the model set its own quality thresholds, decided how much to lean on each design tool the prompt offered, and sent sequences to a vendor, with people steering four times. Among designs aimed at a single toxin, 11 of 60 bound, which is what an ordinary protein-design campaign gets, so the claim is about who ran the campaign, not about a better method.

Evidence and limits: Binding is all that was measured, against toxin proteins grown in the lab, not taken from venom; nothing was tested in cells or in an animal, and Anthropic says it has not made an antivenom. None of the 151 designs aimed at more than one toxin met the bar against more than one. The criteria were set after the data came in, and no design baseline was run, so the model's own contribution cannot be separated from its tools'.

By Nathaniel Horwitz, Amir Shanehsazzadeh

What happened: Yazhuo Liu, Ting Zhu and colleagues at Georgia Tech, the University of Osaka and USC build GBClaw: an orchestrator and five agents, routed to two GPT-5.6 models, that set up atomistic shear simulations in LAMMPS and analyse them. Across 14 asymmetric tilt boundaries in nickel the strength at which sliding starts ran from 14 megapascals to 9.3 gigapascals. Four were rebuilt with arrays of secondary misfit dislocations, the defects a real boundary has: one fell from 1,074 to 530 MPa, two by four orders of magnitude, one rose.

Why it matters: Most interfaces inside a real metal are these mismatched boundaries, and they stay understudied because their structure is hard to pin down in both experiment and simulation. What the sweep shows is about the models, not the agent: if the predicted strength depends this much on whether the boundary's own defect structure sits in the cell, then screening boundaries on simplified cells (with a first-principles or machine-learned potential, at real cost) is reading the cell as much as the material.

Evidence and limits: All of it is simulation with one empirical nickel potential: nothing was made or measured, and the authors call their numbers athermal instability thresholds for the cells they modelled, not universal upper bounds. The cells are small and periodic, which suppresses the long-range defect interactions a wide boundary has, and the energy paths carry no temperature. The {111}/{110} anomaly was traced only after a researcher directed the follow-up runs.

By Yazhuo Liu, Ting Zhu

What happened: Miguel Marques and Silvana Botti's group at Ruhr University Bochum ranked about 2.8 million near-stable compounds in the Alexandria database with two machine-learning models and a fitted formula, kept about 1,200, added about 12,800 compounds already computed, and narrowed the combined set to 52 for full electron-phonon transport calculations. Tested on elemental metals, those calculations land mostly within 30% of measured values. The best compounds, Be2IrPt, H3NaNi, LuIr and LiBePt2, come out at about 34 to 38 million S/m, near aluminium's measured 36.59.

Why it matters: Chip wiring and power lines need metals that conduct well, and conductivity is a property materials databases rarely compute because the full calculation is expensive. This shows a workable route: rank millions cheaply, then spend the full calculation on a few dozen. It also explains why silver, copper and gold stay ahead, and points to a mechanism: electrons from light elements can push a compound's high-scattering d-states below the Fermi level, as in LiBePt2.

Evidence and limits: The match with experiment is for elemental metals, not the new compounds, none of which was made or measured here. Even among elements, calcium comes out at 7.98 against a measured 28.99 million S/m. Spin-orbit coupling was left out; adding it cut Be2IrPt and LiBePt2 by 8% and 9%. Several leading candidates contain platinum, iridium or rhodium.

By Thalis H. B. da Silva, Silvana Botti, Miguel A. L. Marques

What happened: A Sydney, Columbia, BAAI and Tsinghua team re-scored 95 predictors over 217 ProteinGym assays and 2.47 million variants by what each would find under a matched budget, on two objectives: finding top variants, and ranking the landscape. Across 20 scores fitted to 20 measurements of the target assay, prior ranking ability explained under a fifth of the differences in what they recovered (r = 0.44). A plain additive model from measured single mutants then overtook them: at 50 to 100 measured singles for top variants, around 500 for ranking.

Why it matters: This changes what a protein-model leaderboard is for. A campaign's question is which variants to measure next, and the paper shows the quantity that answers it is not the one benchmarks report: a score that ranks a landscape well need not direct a budget well. The practical reading is that a model prior is worth most in the first few dozen measurements, after which the lab's own data wins, and that closed-loop design should be scored by what gets discovered at a given budget, not by correlation on a fixed dataset alone.

Evidence and limits: All retrospective, re-scoring measurements that already exist, with no new experiment and no closed loop run. Substitution landscapes only; the authors say insertions, deletions and de novo design would need their own controls and stated measurement costs. The crossings are additivity predicting double mutants from measured singles; on variants with three or more substitutions the models led instead. They come from deep mutational scans, where every variant is already measured, which is not a lab's candidate pool. A preprint.

By Wenze Ding, Haipeng Gong, Zhiyong Wang

What happened: Matthew Carter and colleagues post the numbers on the company's blog. On AlphaGenome's own panel of 21,948 test cases, each pairing a DNA change with a gene and a tissue, half of them changes statistics pinned as the cause and half it ruled out, Omnii-1 scores 0.810 against Borzoi's 0.705 and the 0.713 AlphaGenome's paper reports. On a splicing panel of 14,458 cases it scores 0.814 against 0.721 for Pangolin. It also predicts methylation, a chemical tag on DNA, one site at a time, agreeing with measurements at 0.879.

Why it matters: Most DNA changes tied to disease sit outside the stretches that code for proteins, and working out which of them does anything is the bottleneck. The earlier models read DNA alone. This one also reads lab measurements of how the DNA is packaged and used in a given tissue, and can take those measurements as an input rather than only predict them. Its lead over Borzoi grows the further the change sits from the gene, from 0.05 within 500 bases to 0.19 beyond 50,000.

Evidence and limits: Every score is the company's own, in a blog post with no paper and no peer review. It says it cannot run AlphaGenome itself, so those numbers are read out of the published paper rather than measured side by side, and AlphaGenome's higher supervised score on the same panel, 0.799, is not the one compared against. The labels are statistical inferences from population data, not variants tested in an experiment, and the model is behind a request form.

By Matthew Carter, Armin Thomas

What happened: Max Highsmith and colleagues report it on the company's blog. The campaign began with 1,620 designs from earlier pipelines, 150 of them with measured binding. Since mid-September, Omnii-1 rewrote positions in the best of those, ranked its own proposals, and Aptamino made and measured a selected set, three times over. The designs push out the measured trade-off between how tightly a single-domain antibody binds and how much of it a cell makes; two from the third round bind below the test's floor.

Why it matters: This is a loop where the model's own ranking decides what gets made, and each round's measurements go back into it. Antibody sequences were not a domain of this genome model's original pretraining; they came in afterwards, as public repertoires and the partner's own measured designs. What is being tested is whether a general biological model can be turned to one partner's narrow design problem. Ranked on the same 327 measured designs, what the model has learned predicts expression and potency a little better than protein-specific models, with a plain one-hot encoding of the sequence close behind on potency.

Evidence and limits: A company blog post, no paper and no peer review, and both sides report their own work. One target, single-domain antibodies, three rounds, with the target's identity and some measured properties held back by the partner. Two designs beating the test's floor is a limit of the test, not a measured strength; the antibody post gives no number for that floor, and 0.2 nM comes from the launch post. The gains over the other models sit inside overlapping confidence intervals.

By Max Highsmith, Michael Poli

What happened: Armin Thomas, Gautam Machiraju, Benjamin Siranosian and Christine Parthemore report it on the company's blog. Livermore measured how 609 natural proteins split their binding across the 15 rare earth metals, on a plate assay it built this year. Given two of the metals, Omnii-1 ranked proteins it had not seen by which one they favour, at a median rank agreement of 0.340 over all 105 pairs, against 0.318 for the strongest public model, ESM-C 6B, and 0.254 for the best of three methods that go by sequence similarity.

Why it matters: Telling these metals apart is a refining problem: China refined 91% of the world's supply in 2024 on the International Energy Agency's count, and the elements differ so little in size, 0.013 angstroms between neighbours, that the standard route needs harsh acids and a lot of energy. Proteins can distinguish them, and the question is whether a model can say which protein will, before anyone makes it. This model had no rare earth binding data in its training, and the lab is now testing the candidates the model generated and ranked.

Evidence and limits: Prediction only: the proteins the model designed are at Livermore awaiting testing, and no result from them is reported. The company's post, its own evaluation, no paper and no peer review. On the 17 pairs closest in size, the hardest cases, the company puts the gap to ESM-3 at 0.006 against a paired error of 0.005 and declines to rank the two. AlphaFold2 scored 0.191 here, below the simplest sequence-similarity baseline, 0.198.

By Armin Thomas, Gautam Machiraju, Benjamin Siranosian, Christine Parthemore

What happened: Matt Turk, James Golden and colleagues at Protege DataLab and Heimdall Bio ran 4,249 rollouts of four frontier models, plus 4,486 trials across 26 models. Claude Opus 5 (0.816), GPT-5.6-sol (0.812) and Grok 4.5 (0.785) are not separated by the intervals; Gemini 3.1 Pro Preview scores 0.707 and every gap to it excludes zero. Opus 5 averages 0.907 on metadata lookup and 0.716 on numeric calculation. Numeric tasks are 28 of the 131 but six of the twelve tasks all four models fail.

Why it matters: A researcher picking an agent to measure distances and angles in a structure needs the numeric score, not the overall one, and for Opus 5 the two differ by 0.19. The paper also shows how much of a leaderboard is the setup around the model: on the 99 tasks with eight full runs, one run per task instead of eight puts a different model first in 23.9% of resampled task sets, and swapping each provider's own command-line agent for the reference harness moves scores by -0.071 to +0.086.

Evidence and limits: The four experiments use different task subsets, so model, harness, access and task mix cannot be ranked against one another, and the harness comparison can only detect differences of about 0.12 to 0.20. A task all four fail is not shown to be unsolvable. The 131 tasks lean to biophysics, and the scores describe those model endpoints in the collection windows the appendix gives. A preprint, not peer reviewed.

By Matt Turk, Brandon L. Neel, Diana M. Lopez, James B. Golden

What happened: Manuel Gallone, Shyam Ravidath and Srihith Madireddy built an out-of-distribution harness (two datasets, six dataset-split combinations, scVI, CPA and a Ridge baseline, three uncertainty methods), then retrofitted null baselines it had never included. The harness fed each test cell to the model and scored it on reproducing that same cell, so an identity predictor reaches a perturbation-effect correlation of 1.0 everywhere. Against a constant train-mean predictor, scVI wins on 99.7% of Norman cells and 0.2% of Replogle cells.

Why it matters: Perturbation prediction is one of the directions this record follows most closely, and the audit says metric names there can outrun what is measured. The six failure modes are mechanical rather than statistical (scoring reconstruction as prediction, holding out sequencing batches instead of perturbations, silently swapping an unseen perturbation for a control), and each comes with a diagnostic that needs no retraining, so another harness can be checked largely from what it already stored. The one model never handed the target cell, the linear baseline, is the only one that beats its null on a genuine counterfactual.

Evidence and limits: It is the authors' own harness, not a published benchmark; PerturBench is cited but never audited. The audit needed no retraining, though models were re-run counterfactually, and everything is computed on two datasets with no new experiment. One result survives: split-conformal coverage holds at 0.900 in distribution and falls to 0.847 across cell types, while on Norman that nominal 0.900 covers 0.271 to 0.342 of differentially expressed genes.

By Manuel Gallone, Shyam Ravidath, Srihith Madireddy

What happened: Joseph Hart, Samuel G. H. Brookes, Fabian L. Thiemann and Christoph Schran screen six models on density-functional energies, forces and binding curves, then put four through melt-quench-anneal dynamics at seven densities, five runs each. Every pure-carbon run finishes. Adding oxygen and hydrogen separates them: MACE-MH-1 completes all 35 runs, ORB-v3 33, MACE-MP-0b3 25 and MACE-OMAT-0 18. UMA never reached this stage (its carbon-dimer curve has two spurious minima) despite low errors on flat-carbon environments.

Why it matters: A foundation potential is sold on transferring to systems it never trained on, and the usual evidence is a table of energy and force errors. Here those errors do not say which models break: UMA scores well on flat carbon, then is thrown out for a dimer curve no potential should have, while models sharing a training set share ring statistics across architectures. All four that ran reproduce the main experimental peaks for nanoporous carbon, so the question is where they stop, not whether they work.

Evidence and limits: One element family and one workflow. Stability is scored in 216-atom cells; the structures compared with experiment are larger, at one density, and matched to published measurements of different samples. The first author co-authored both the model that comes out best and the scoring framework the error table uses, and the paper declares no competing interests. Two of the six never reached dynamics, one on cost alone. It is a workshop paper.

By Joseph Hart, Samuel G. H. Brookes, Fabian L. Thiemann, Christoph Schran

What happened: METiS announced a strategic collaboration with Lilly on 19 October to research lipid nanoparticle delivery that reaches one tissue rather than everywhere. METiS runs its own AiLNP platform, which designs the ionizable lipid inside the particle, and Lilly provides its drug discovery, development and commercialisation expertise. METiS takes an upfront payment of over $10M, is eligible for up to about $1.1B in development, regulatory and commercial milestones, and single-digit tiered royalties on net sales.

Why it matters: The record holds METiS's Hong Kong listing, its mouse data for a liver-cancer mRNA drug and a Phase 3 formulation result. This deal is about the platform rather than one drug asset: Lilly is paying for the design step, which is where METiS says its AI sits. Where a nanoparticle ends up in the body decides what an mRNA or gene therapy can treat, and it is still mostly found by trial and error.

Evidence and limits: A deal, not a result: no target, no programme count, no timetable, and nothing public about how well AiLNP picks lipids. All the money beyond the upfront is contingent on programmes advancing, and the upfront is stated only as a range.

What happened: A Zhejiang University team ran DFT calculations, about a million CPU core-hours, on 45,909 frameworks from six databases, compressed the resulting fields about 26-fold and turned them into point clouds. Models reading the fields beat a structure-only model on 12 of 13 properties, all the adsorption and separation ones among them; the structure model kept band gap. Combining the two did better still on every property. Electron density predicted from structure is off by 0.51% on average, and starting DFT from it cut run time by 13.59%.

Why it matters: MOF models mostly learn from atom positions, while gas adsorption depends on how a guest molecule feels the framework's charge. Aligned electronic fields at this scale were scarce because they are costly to compute and store; the authors say the dataset will be released when the paper is published. The results say the fields carry information structure-only models were not using, most of all for adsorption and separation.

Evidence and limits: All computed, by the authors. The structure baseline is one model, CGCNN from 2018, so the comparison with structure rests on it. The electrostatic potential is reconstructed far worse than the density, 8.42% off and worse in large cells. The 13.59% saving leaves out the time spent predicting the density, and is an average: some frameworks ran slower.

By Zhaolin Hu

What happened: James Kirk, Lee von Kraus and colleagues trained a Transformer, without labels, on over 4 million hours of recordings from electrodes in more than 1,400 plants of 100 species, in labs, greenhouses, orchards and forests. A second part, trained on over 41,000 handheld leaf readings, estimates seven stress measures such as how efficiently leaves photosynthesise and how warm they run. On held-out plants of six species under drought, its estimates moved with the readings at correlations of 0.27 to 0.71.

Why it matters: Earlier models of plant electrical signals were each trained for one task, usually on one species. One model that carries across species would let a grower read a crop nobody has labelled. On species left out of training the estimates still moved with the readings, less closely, and on plants stressed by salt or a herbicide, which the model never saw labelled, they shifted the way the literature expects.

Evidence and limits: The data, sensors and evaluation are all the company's own, in a preprint. The correlations follow one plant over time; about half the error is a fixed offset per sensor, and on the error measure, three of the seven estimates did worse than guessing each species' average. One of the six densely labelled species read at or below zero, and nothing yet shows a grower acting on it.

By James Kirk, Piotr Bednarski, Filip Twardy, Justin Mast, Lee von Kraus

What happened: The two endpoints return ranked web results with a relevance score and an authority tier, and a page's content as Markdown. DeepSeek says its in-house engine has long served the search in its web app and API, handles hundreds of millions of requests a day and indexes hundreds of billions of pages, with filters for SEO spam, GEO manipulation and low-quality AI-generated text. Calls made through the server-side tools in DeepSeek Harness, Claude Code and Codex stay free until 31 October.

Why it matters: A literature agent needs a search backend it can call by program, and almost all of them sit on Bing's or Google's index. This is one of the few that does not, from DeepSeek, whose open models many AI4S groups already run, priced per call rather than by tier. Whether its index covers journals, preprint servers and Chinese-language sources well enough for science is for anyone building on it to test.

Evidence and limits: The request and index figures are DeepSeek's own and unaudited; the English release note gives the daily requests and the Chinese one the index size. No coverage or freshness comparison against other search engines was published, and the endpoints were not called here.

What happened: Zachary Cutts, Pooja Rao, Annalisa Pawlosky and colleagues gave Gemini-based agents literature search, folding tools and a Python environment, and had them design 1,000 single-stranded DNA aptamers per round against two proteins. Each pool was synthesised, pulled down on beads and sequenced, and the agents read the counts before designing the next. Corrected for bead background, hit rates went 0% to 26% to 42% on alpha-thrombin, or 37% counting only designs free of published motifs, and 10% to 4% to 34% on myoglobin, at about $3 a candidate.

Why it matters: Earlier studies of reasoning agents in molecular design tested their designs once, at the end. Here the measurements came back three times, and what the agents did with them is on the record: after a counter-screen showed that 97% of the cross-reactive sequences were G-quadruplexes, which stick to charged surfaces whatever the target, they moved the myoglobin designs to pseudoknots and hairpins, and specificity there went from 19% back to 72%. The authors picked aptamers because the designed sequence is the molecule, so one cycle is 10 days instead of the 3-6 weeks a protein needs.

Evidence and limits: Binding is measured in the authors' own pooled assay. Single-clone titrations covered eight thrombin designs only, did not reach saturation, and the authors call those affinities preliminary: nanomolar for reworked literature motifs, above a micromolar for de novo ones. Round 1's 0% on thrombin was a construct artifact the agents then fixed, the harness changed after it, and the authors say control arms with no feedback would be needed to credit the loop itself. A preprint, not peer reviewed.

By Zachary Cutts, Pooja Rao, Annalisa Pawlosky

What happened: Tatsuki Hori, Yoshitaka Moriwaki and Ryuichiro Ishitani at Institute of Science Tokyo and the University of Tokyo release RGI-Toolkit, which corrects the denoiser output at each reverse-diffusion step with a differentiable restraint loss, as a backend-agnostic library wrapping six predictors: Boltz-2, Protenix-v2, AlphaFold3, Chai-1, OpenFold3 and ESMFold2. On 209 protein-ligand complexes, conformer restraints lifted chirality agreement from about 50% to 100% and cis/trans from 70-80% to 100% for all six, with lower bond-length and bond-angle errors. No retraining. MIT-licensed.

Why it matters: Chemically invalid ligand geometry is a known weak spot of AlphaFold3 and its kin, and this fixes it from outside the model, at inference, without retraining it or changing a weight. It does not touch where the ligand sits: the authors report pose accuracy is neither helped nor hurt, so the gain is a clean stereochemistry patch, not a better binder prediction. Geometric and custom restraints can also push sampling toward a chosen conformational state.

Evidence and limits: The chirality and cis/trans figures are computed agreement rates against each ligand's ideal conformer, on 209 complexes with one seed each, not a physical measurement. The restraints only work as well as the reference conformer: a bad RDKit embedding misguides them. The two earlier RGI papers this builds on were in Boltz; the new work is extending that to six predictors and packaging it. A preprint, not peer reviewed.

By Tatsuki Hori, Yoshitaka Moriwaki, Ryuichiro Ishitani

What happened: Scripps Research, OpenAI and PerturbAI gave one agent to each of 1,946 perturbations in an in vivo whole-brain CRISPR Perturb-seq atlas, with the source manuscript withheld and published after the model's stated knowledge cutoff. Of 55 findings derivable from the expression data alone, the agents matched or refined 48. They adjudicated 1,519 high-confidence findings against the literature, reading 39,185 papers and keeping 16,445 citations. The run took 32 minutes.

Why it matters: The bet is coverage through many parallel agents rather than one stronger agent, and the paper measures it directly: at 893 perturbations a single agent cited 11 references while the distributed run kept 711 times more. Blinded experts preferred the distributed readings. This is a throughput claim about interpretation, the step that now limits what a large screen yields, not a claim that the agents found new biology.

Evidence and limits: All computed, on one atlas the same group produced, and every corresponding author has a stake: one at OpenAI, whose model this evaluates, two who co-founded PerturbAI, which hosts the data. Recovery is scored by the source study's own authors, 14 of 71 findings uncovered. Held-out predictions reach 0.705 balanced accuracy: 82% of the silent perturbations called, only 59% with a real response. The expert preference ranks readings, not their correctness. A preprint, not peer reviewed.

By Alexander Sun, Xin Jin

What happened: Andi Gu, an Anthropic STEM Fellow, with Alek Kemeny of Anthropic and Harvard's Susanne Yelin and Mikhail Lukin, had Claude Mythos 5 write circuits in Python, each checked and scored by an evaluator. In six tasks the circuits cut both Toffoli and qubit counts against the best published design; a seventh had no published reversible design to beat. In the eighth, a new reversible modular inversion, its key step formalised in Lean, gives one design 17 times fewer gates than the smallest previous one, for 12% more qubits.