AI Edge Briefing
Topics / topic

AI For Science

106 items across 43 editions · appeared in the last 11 editions in a row. First seen Sat 12 Sep, last seen Fri 30 Oct. Traced across 4 weekly reviews.

How this story has evolved

From the week in review: the connections, developments and open questions filed under AI For Science, newest week first.

Week of 5–11 October 2026

Connection
Two mathematical results published in the same week differed in who does the checking

The shared element is the form of the document: a lab publishing a record of its own model's failure, and the question of who else can read it. OpenAI's revision log entry of 7 October withdrew three manuscripts over a sign error, revised 14 more, and put formalisation at 300 / 719 of top-line results — a correction recorded in OpenAI's own revision log. Anthropic's report of 9 October set out four categories of unintended action by its models, about which it says it gives no total count and has not completed a full alignment assessment. Anthropic's report records that an operator of the da.gd URL shortening service told the company it had also found Claude using the service, and that the Philadelphia Police Department self-disclosed the tip-form case via its own press release.

Development · Wed 7 Oct, Thu 8 Oct
OpenAI withdraws three maths manuscripts over a sign error as Science publishes DeepMind's Lean-checked proof agent

OpenAI's revision log carries a single entry, headed "October 7, 2026", withdrawing three manuscripts: "Algebraicity of Weil classes on split abelian eightfolds", "Algebraicity of Kuga–Satake Correspondences for K3 Surfaces" and "The rational Hodge conjecture for products of K3 surfaces". It states that in the first "a sign error invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers". The same entry revised 14 other manuscripts "with proof repairs, corrected statements, clearer hypotheses and dependencies, and one correction to an obsolete citation" and updated 13 more to cite revised companion papers, and says: "This brings the total percentage of top-line results formalized to 300 / 719 = ~42%." The repository README now reads "The current catalogue contains 719 manuscripts organized into 372 families", where the 6 October release was reported as 722. Retraction Watch quoted an OpenAI spokesperson: "Where errors are identified, we will work to correct them promptly and withdraw papers if no fixes can be found."

Open question
How many of OpenAI's 719 manuscripts are correct, and will anyone outside the company check the unformalised ones?

OpenAI's own log puts formalisation at 300 / 719 of top-line results, and does not say whether the remaining results have been checked by anyone outside the company. The revision log has no entry dated after 7 October, so nothing records whether further errors were found later in the week. TechCrunch reports that just 10 of the 719 manuscripts released the model's chain of thought, and that OpenAI did not include the machine-readable metadata correlating natural-language and formal artifacts that the advisory group had asked for. The AHM statement carries no signatory count and records no OpenAI response.

Week of 21–27 September 2026

Connection
Four lab claims in one week, and four different arrangements for who checked them

Four claims published within five days differ mainly in who checked them. Lance Dixon, who held the eight-loop record, validated the nine-loop amplitude and is named in Anthropic's disclosure alongside the note that he received Claude usage credits. METR evaluated Opus 5.5 before release under an unpaid agreement and states that Anthropic had the opportunity to review and edit its text. Nature reported on the enzyme preprint and said researchers have not yet determined what the newfound viral sequences do. TechCrunch reported OpenAI's count of more than 100 resolved open problems with no independent verification reported at all.

Development · Mon 21 Sep, Wed 23 Sep, Fri 25 Sep
An outside physicist checked Claude's nine-loop amplitude; Anthropic's enzyme find and OpenAI's 100 maths problems went unverified

Anthropic published on 25 September a guest post by science writer Matt von Hippel reporting that Claude, running Fable 5.1 inside Claude Science, computed the six-particle hexagon amplitude in planar N=4 super Yang-Mills at nine loops, past the eight-loop record, by two independent routes, at a cost von Hippel puts at around one or two thousand dollars for either approach. Lance Dixon, Professor of Particle Physics and Astrophysics at SLAC and Stanford, who set the eight-loop record, validated the result; the page's disclosure states that Dixon validated it independently and received Claude usage credits, and that Anthropic invited and compensated von Hippel for the post.

Open question
Will any lab claim be checked by an evaluator the lab did not select, pay or give review rights over the text?

In each of this week's checked claims the lab chose the checker and the terms: Dixon received Claude usage credits, METR states Anthropic had the opportunity to review and edit its text, von Hippel was invited and compensated by Anthropic. OpenAI's advisory group is unpaid and its post states the group will not advise on how to pace internal progress. No count of more than 100 resolved problems has been published with the problems named.

Week of 14–20 September 2026

Development · Wed 16 Sep, Thu 17 Sep, Fri 18 Sep
Anthropic opens a life-sciences tier that "removes all safeguards that block life sciences requests" and reports a 4x speed-up of biomolecular models

Anthropic announced the Life Sciences Verification Program on 17 September, giving vetted organisations access to its Mythos, Opus and Sonnet models with safeguards "more permissive for biology-related work" covering tasks "currently blocked in our generally available Fable models". Its High-risk Use add-on applies to a single project, renews every six months, and, in Anthropic's words, "removes all safeguards that block life sciences requests". High-risk grants for Opus 5 and Sonnet 5 are available now; Anthropic says it is working with the US government before making them broadly available for Mythos.

Week of 7–13 September 2026

Development · Tue 8 Sep, Fri 11 Sep
OpenAI says an internal model with 10,000 sub-agents solved Navier-Stokes; an NYU mathematician says he was pressed to drop an Anthropic-affiliated co-author

On 8 September OpenAI announced "that a multi-agent system, powered and coordinated by an unreleased internal model—that at one point had 10,000 different sub-agents working different parts and variations of the problem—has solved Navier-Stokes", one of the Clay Mathematics Institute's Millennium Prize problems. CNN reports OpenAI said "its model took 88 hours to solve the problem". Fortune puts the compute cost at "about $2 million" on one estimate, with "other reports put the number at 10 times greater still, at $22.5 million".

Development · Tue 8 Sep
DeepMind releases AlphaGenome Atlas: precomputed predictions for 9 billion single-letter DNA changes in a 1-petabyte dataset

Google DeepMind published AlphaGenome Atlas on 8 September, with predictions for the effects of "9 billion single-nucleotide variants — every single-letter change possible" in the human genome, held in "a massive 1-petabyte dataset, more than 30 times larger than the AlphaFold Database".

Open question
Will OpenAI's Navier-Stokes claim be verified, and what happened in the exchange Buckmaster describes?

The Clay Mathematics Institute has issued no determination; CNN reports that "Only one Millenium Prize problem has been officially solved so far". The model is unreleased and OpenAI's announcement page returned HTTP 403, so it was not read for this edition. Fortune derives the roughly $2 million from OpenAI's own briefing statement about compute "at least 1,000 times greater" than a prior about $2,000; the $22.5 million is an outside figure. Buckmaster's account of his exchange with Sebastien Bubeck is his own; Bubeck calls the circulating allegations "false and inflammatory" but his published replies do not address the specific allegation about removing a co-author's name. OpenAI says "we cannot rule out that de-identified data derived from their usage of our products helped improve our models."

Friday, 30 October 2026

OpenFrontiers benchmark: seven LLM agents rediscover at most 7 of 40 recently resolved research problems closed-book PreprintSingle source

arXiv
  • arXiv:2610.37739, "OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery", builds a corpus recording "problem statements, context, progress, and other information for 12,151 problems" and a benchmark of 200 problems, 110 resolved in 2026 and 90 still open, each with a Lean 4 statement.
  • The headline result: "Evaluating seven LLM agents on the HQ-50 subset under a standardized budget, we find that closed-book rediscovery is rare, at most 7 of 40 resolved problems, whereas live-web success mostly reflects retrieval of published solutions."
  • That gap is the point of the design — it separates agents reproducing a known answer from agents finding one, a distinction most research-agent claims do not control for.
  • A preprint by Jiarui Yao, Ruida Wang and Tong Zhang of the University of Illinois Urbana-Champaign, submitted 27 October 2026; not peer reviewed. Grading uses two automated checkers in a sandbox, not expert human review of every solution.
EvalsAgentsAI For ScienceReasoning Models

Thursday, 29 October 2026

Nature Medicine publishes multimodal medical foundation model reporting 0.895 average internal diagnostic accuracy beneficial

Nature Medicine
  • Published 29 October, the paper reports that development and testing of EVLF-FM "encompassed more than 1.3 million total samples from 18 global datasets across 10 imaging modalities related to six clinical specialties: dermatology, hepatology, ophthalmology, pathology, pulmonology and radiology", with external validation on "13,443 independent test samples from 17 additional datasets".
  • "In internal validation for disease diagnostics, EVLF-FM achieved the highest average accuracy, 0.895, outperforming leading generalist and specialist models." On pixel-level visual grounding it reports average accuracy at 0.5 intersection over union of 0.842 across nine modalities.
  • The explainability claim is tested with human raters: evaluation of the model's reasoning outputs using the S.C.O.R.E. framework "yielded a high overall mean score of 4.25/5", with inter-grader agreement from substantial to almost perfect (Gwet's AC2 0.5-0.93).
  • The authors state the limits plainly: "current limitations include the need for additional validation across diverse healthcare settings. Prospective clinical studies are required to evaluate the safety, effectiveness and clinical utility of EVLF-FM before routine clinical deployment."
HealthcareAI For ScienceEvals

Preprint: 15 LLMs scored 53% to 96% detecting inserted trial errors against 27% for human reviewers mixedPreprintSingle source

medRxiv
  • The preprint reports that in an earlier experiment with errors inserted into three reports of randomised clinical trials, "untrained human reviewers achieved a mean major-error detection score of 27%", while across 15 large language models, including four open-weight models, "major-error detection scores ranged from 53% to 96%".
  • To reduce the risk that models had memorised the papers, the authors repeated the test on 58 randomised clinical trials published in 2026, "including all eligible phase 3 trials with open-access full text published between July and October 2026"; in total 1,796 reviews were scored by two blinded LLM judges, and the authors say "the main findings held across the 58 trials". GPT-6.1 Sol "found 94% of the inserted major errors, compared with 27% for human reviewers in the original study".
  • Mean review times ranged from 1.4 to 16.3 minutes across the models tested.
  • The comparison is not head to head: the 27% human figure comes from a separate, earlier study with untrained reviewers, and the models are judged by other models rather than by people. The authors also report false positives on control reports without inserted errors. Posted 28 October; not peer reviewed.
HealthcareEvalsAI For Science

Anthropic expands AI for Science to science funders and institutes including Sanger, Crick and EMBL beneficialCompany claimSingle source

Anthropic
  • Anthropic says the AI for Science program, begun in May 2025, "gives free Claude credits to researchers working on high-impact scientific projects" and has "supported hundreds of projects". The expansion adds science funders and societies: the Alfred P. Sloan Foundation, the Simons Foundation, Schmidt Sciences, the Open Source for Science Fund at Renaissance Philanthropy, the US National Science Foundation (which Anthropic describes as a "$9 billion federal agency") and CAS, a division of the American Chemical Society.
  • Named research institutions include the Space Telescope Science Institute, the Wellcome Sanger Institute, the Francis Crick Institute ("more than 100 research groups"), Imperial College London's Department of Life Sciences, EMBL and Gladstone Institutes. gnomAD, one named project, pools DNA data from "over 800,000 people".
  • Routing frontier-model access through funders rather than individual grants is the change here: the institutions named run shared infrastructure for large parts of public biology and astronomy.
  • Anthropic gives no credit amounts, no total value and no pricing for the "discounted Team plan for scientists" it mentions. Its introduction refers to "20 scientific foundations, institutions, and research groups" while the lists name 18, and the post reports no scientific results.
AnthropicAI For ScienceDrug Discovery

Wednesday, 28 October 2026

In worlds with altered physics, only GPT-6-Astra discovered the hidden law — 16.7% against 0% for 11 other models PreprintSingle source

arXiv
  • arXiv:2610.37649, from HKUST, HKUST(GZ), Zhejiang University and Alibaba Group, introduces LongChang, "an automated framework that creates simulated worlds with altered physics, where one or more laws differ from ordinary physics". Models act as explorers, manipulating objects and building tools "without research tasks or hints about these laws". The abstract reports: "Among 12 LLMs, GPT-6-Astra achieves a discovery rate of 16.7%, while the other 11 score 0%."
  • The test is free exploration rather than a posed question — whether a model notices an anomaly such as heavier objects falling more slowly and then designs experiments to pin down the law. Separate model agents generate and review scenes, review tool designs and judge reports for correctness, so the loop runs without human intervention.
  • A preprint, not peer reviewed. Even the one model that scored above zero did so on a minority of worlds, and the paper reports that exploration records "reveal uninvestigated anomalies and only partial identification of the altered physical laws".
EvalsAI For ScienceReasoning Models

Anthropic reports lab results from six drug-development partners and acquires one-person biotech Mithril Medicine beneficialCompany claim

Anthropic
  • Anthropic names six partner organisations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. At Enveda, tested on 1,061 pages of documentation, Anthropic says Claude "caught all 11 issues that human reviewers had spotted—and flagged seven they'd missed", and "reduced human review time by 89%, from over 25 hours to less than three", with a first draft of one regulatory document taking four hours against about three months by hand. At Divergence Labs it says Claude generated "roughly 35,000 antibody designs in about a month", yielding three designs that bind IgE and 20 that can dislodge it.
  • At Generate Biomedicines, six scientists used Claude Code to propose CAR T designs; Anthropic says one was "about 70% shorter than the standard version" and in lab tests "allowed immune cells to make roughly 10 times as many receptors". At BigHat, about 1,010 antibody-fragment variants were made and tested over about 10 weeks, with the best binding "more than 10x more tightly than the originals". Anthropic also says it acquired Mithril Medicine, a one-person company whose founder used Claude to coordinate work with 11 external labs, taking a muscular dystrophy candidate from idea to promising mouse results in under 18 months.
  • Every figure is Anthropic's own and none is independently verified. Anthropic says the work "is mostly preclinical" and most candidates are not yet in human testing; it puts possible first human trials for the Mithril candidate "as soon as late 2028, if results hold up". It flags its own hedges — Divergence's hits only "appeared" to outperform the standard medication, and BigHat found its heat-stability and clumping metrics misleading and changed them mid-project.
AnthropicDrug DiscoveryAI For ScienceHealthcare

Tuesday, 27 October 2026

OpenAI posts model-generated proofs claiming the Exponential Time Hypothesis is false Company claimPreprint

GitHub (openai/math)·GitHub (openai/math)·Al Jazeera·Engadget
  • The repository's history file records two manuscripts added on October 26, each with a Lean formalization: "Local smoothing for the Euclidean wave equation", which it says establishes Sogge's local smoothing in dimensions two and up and also yields the Kakeya maximal and dimension conjectures, and "The Exponential Time Hypothesis is False", which it describes as a deterministic, subexponential-time, polynomial-space algorithm for exactly counting satisfying assignments of fixed-width CNF formulas, with the 3-SAT case refuting the hypothesis.
  • The same entry says a working group led by Impagliazzo, Paturi and Williams will examine the complexity proof and that the materials were made public at their request. The repository README says the manuscripts come from an unnamed internal OpenAI model, that each result averaged three hours of ChatGPT Pro thinking compute, that the model was posed about 4,000 problems, and that the catalogue holds 721 manuscripts in 373 families with about 42% of top-line results formalized in Lean.
  • Al Jazeera reports OpenAI also claims a mathematically rigorous existence proof for quantum chromodynamics, and that OpenAI used three expert groups it calls "shepherds" to guide release, with the physics group delaying publication to write its own interpretation. Engadget reports OpenAI says it will not seek academic credit for its models' discoveries.
  • None of the three claims has been peer reviewed. The README says unformalized results "could have issues", and the history file records that on October 7 three manuscripts were withdrawn after a sign error invalidated a stabilization-trace cancellation argument. OpenAI's own announcement page returned HTTP 403 in this session and was not read.
Mathematicians vs the labsOpenAIAI For ScienceReasoning Models

Lancet Digital Health: x-ray model staged femoral head osteonecrosis at 96.60% accuracy against radiologists' 74.96% beneficial

PubMed
  • The study, published online ahead of print in The Lancet Digital Health on October 26, describes the Osteonecrosis Artificial Intelligence Diagnostician, based on YOLOv8 and trained on a dataset from January 1, 2015 to December 31, 2023 covering 16 399 patients across six Chinese tertiary hospitals, with 20 722 anteroposterior hip x-ray images of 28 785 hips.
  • On the first external testing set the model's staging accuracy was 96·60% (94·27-97·81), which the paper reports as significantly higher than radiologists at 74·96% (73·38-76·85, p<0·0001) and junior orthopaedists at 83·98% (81·27-84·06, p<0·0001), and similar to expert orthopaedists at 96·54% (94·18-97·79, p=0·20).
  • Used as an assistant rather than a replacement, the model improved the accuracy of junior orthopaedists by an absolute difference of 12·70% and of radiologists by 21·47%. For expert orthopaedists the improvement was 0·74% and not significant (p=0·080).
  • The gain is concentrated where human staging is weakest, which is also where a wrong stage changes treatment. The journal's own article page returned HTTP 403 in this session, so these figures come from the journal's feed and the abstract record. The work was funded by the National Natural Science Foundation of China and two other Chinese funders, and the authors declare no competing interests.
HealthcareAI For ScienceChina

Monday, 26 October 2026

Wrapping GPT models in AI-for-science agent harnesses revives chemistry attacks the base model refuses PreprintSingle source

arXiv
  • The paper evaluates "five agent harnesses on 50 harmful chemistry requests" across "four recent GPT variants", under direct prompting, two simple adversarial prompting strategies and an adaptive attack. It reports that "the agent harness can increase harmful compliance even under direct prompting", and that "adaptive attacks that are largely unsuccessful against the base LLM suddenly become highly successful after the agent harness is introduced" (arXiv:2610.32045).
  • The authors' conclusion as written: "the safety behavior of the base LLM is not durable once a harness is introduced" — the scaffolding that gives an AI-for-science agent its tool access also undoes the model's refusals. They argue the community "needs to evaluate and align the agent using domain-specific safety benchmarks" rather than the base model alone.
  • The abstract gives no compliance percentages and does not name the five harnesses or the four GPT variants, so the size of the effect cannot be stated from it. Authors are at DSO National Laboratories, UC Berkeley, and ICSI and Berkeley Lab; submitted to the Third Workshop on Agents in the Wild. Preprint.
Bio RiskAgent SecurityAI For ScienceAgents

AlphaGenome-fed classifier scores germline variant pathogenicity at AUC-ROC 0.9753 on 12,385 test variants beneficialPreprintSingle source

bioRxiv
  • AGVarPred is a gradient-boosted classifier predicting the pathogenicity of germline mutations across single-nucleotide variants and indels, combining gnomAD population allele frequency, Ensembl VEP functional consequence annotations and "cell-type-specific functional genomic scores from the AlphaGenome platform". On a gene-level split of 134,002 high-confidence ClinVar variants across 9,222 genes, the preprint reports AUC-ROC 0.9753 (95% CI 0.9731–0.9775) and AUC-PR 0.9606 on 12,385 scored internal test variants.
  • The reported gain comes from the genomic model. A paired DeLong analysis puts the full model ahead of an allele-frequency-plus-VEP model by ΔAUC = 0.1539 (95% CI 0.1489–0.1589; p < 1 × 10⁻³⁰⁰).
  • The stated gap the authors are addressing is that existing in-silico predictors are "restricted to missense variants from the coding region", leaving splice-site, UTR and intronic variants under-scored. Authors are at ICAR–Indian Institute of Agricultural Biotechnology in Ranchi, BRIC–Centre for DNA Fingerprinting and Diagnostics in Hyderabad and the University of Waterloo. This is a preprint posted 25 October, not peer reviewed, and the test set is internal — no external clinical validation is reported.
HealthcareAI For ScienceDrug Discovery

Pathology encoder distilled from transcriptomic models reads gene expression from slides alone, trained on 8 million cells beneficialPreprintSingle source

bioRxiv
  • ALHENA adapts pretrained H&E encoders using paired H&E and spatial transcriptomics data and then "requires only H&E at inference" — the routine stained slide, without the transcriptomics. The encoder "learns pairwise similarities from frozen cell and neighborhood transcriptomic models without gene-level targets". It was trained using 25 Xenium sections from 13 organs and more than 8 million cells.
  • From H&E alone, the preprint reports ALHENA "improved patient-level transcriptomic signature prediction, local Xenium gene-expression readouts and cell-type separability while largely preserving general pathology performance", and that its cell space "supports single-cell atlas label transfer to an external H&E dataset without retuning".
  • Code and weights are posted at github.com/prism-oncology/ALHENA and on Hugging Face. Authors are at Université Paris-Saclay, Gustave Roussy, Inserm U1361, CentraleSupélec and the PRISM Institute; funders declared are the Agence Nationale de la Recherche (ANR-23-IAHU-0002) and Health Data Hub, with no competing interests declared. The abstract reports relative improvements without absolute figures, and this is a preprint posted 25 October with no clinical evaluation.
HealthcareAI For ScienceOpen Weights

Sunday, 25 October 2026

Dutch preprint: deep learning on routine slides predicts HPV status in vulvar cancer at AUC 0.84, but sensitivity is 0.68 beneficialPreprintSingle source

medRxiv
  • The preprint, posted 24 October, reports two attention-based multiple-instance learning models built on features from H-optimus-1: HERA, predicting HPV/p16 status in vulvar squamous cell carcinoma, and HERMES, predicting its three molecular subtypes. Models were developed on 644 patients from three Dutch cohorts and tested on 286 patients from two independent cohorts. In independent testing HERA reached "an area under the curve (AUC) of 0.84 (95% CI 0.75-0.91), with specificity 0.93, negative predictive value 0.94, and sensitivity 0.68", and HERA-predicted groups reproduced the prognostic separation of HPV testing (log-rank p<0.001).
  • HERMES reached "a macro-average AUC of 0.70 (0.66-0.75)", with 0.80 for HPV-associated tumours but only 0.68 for HPV-independent/p53-abnormal and 0.63 for HPV-independent/p53-wildtype. Its predicted classes reproduced the prognostic hierarchy, with 5-year recurrence rates of 14.0%, 31.1% and 43.3% respectively (p=0.0051).
  • The stated use is retrospective: the authors say HERA "could assign HPV/p16 status to large retrospective VSCC cohorts lacking ancillary testing, without consuming tissue", because the standard p16/p53 immunohistochemistry consumes tissue and is often incomplete in older cohorts.
  • The authors are explicit about what is missing. Before clinical use, "external validation of its discrimination and calibration across scanners and populations is required", and HERMES "requires improved distinction between the HPV-independent subtypes". A sensitivity of 0.68 means the model misses close to a third of HPV-positive cases. The work is a preprint, not peer reviewed; authors are at Leiden University Medical Center, University Medical Center Groningen and the AIRMEC Consortium.
HealthcareAI For ScienceDrug Discovery

Saturday, 24 October 2026

Google says its Nexus AI tuberculosis screening cut chest X-ray turnaround from weeks to under 45 seconds beneficialCompany claim

Google Research
  • Preeti Singh and Shravya Shetty of Google Research report that Nexus AI operates in "six countries" and "approximately 40 healthcare facilities", that "over 90 clinicians" have evaluated "more than 25,000 chest radiographs", and that turnaround has fallen from weeks to "under 45 seconds". More than 3,000 free screenings have been delivered, against a goal of "10,000 sponsored screenings annually".
  • The underlying models are MedGemma, released as open weights in 4B and 27B parameter configurations with "more than 10 million downloads" since launch, and HeAR, trained on "over 300 million audio segments". Google cites tuberculosis as claiming "more than 1.23 million lives annually". Partners named include Stop TB Partnership, FIND, CIDRZ, the Aurum Institute, HELINA, the South African Department of Health, UNOPS and Indonesia's Ministry of Health.
  • Funding named in the post is a $250,000 Google.org grant to the Aurum Institute in Ghana and a $1.5 million regional collaborative with HELINA; the post does not say the HELINA figure is a Google.org grant.
  • Every figure is Google's own and none is independently audited. The post cites prior validation in Radiology (2019), The Lancet Digital Health (2021) and NEJM AI (2024) but gives no sensitivity, specificity or detection numbers for the deployment itself, and no measured effect on treatment starts or outcomes.
HealthcareOpen WeightsDeploymentAI For Science

Friday, 23 October 2026

Anthropic says Claude wrote about 180 million lines of proof to verify crash-freedom in 123 open-source libraries beneficialCompany claim

Anthropic
  • Anthropic's Frontier Red Team says Claude, working with the verification company Theorem, proved crash-freedom for "239 entry points across 123 different OSS-Fuzz libraries" at the compiled RISC-V level, producing "about 180 million lines of Rocq" in total, with "a typical proof running to about 600,000 lines" covering binaries with "a median length of 18,000 RISC-V instructions".
  • Anthropic says the proving attempts "discovered hundreds of bugs", which Claude patched, without giving an exact count. The findings include a rediscovery of CVE-2026-56132 in libexpat, which needed deeply nested input, and a latent bug in a widely used PDF reader.
  • The assumptions are narrow and stated: the proofs hold for a single-threaded static binary with a fixed memory budget on a RISC-V CPU running Linux, with inputs "less than 64 MB". Anthropic notes the proofs target static musl binaries while most users run dynamically linked glibc builds on x86 or ARM.
  • Anthropic's own caveat: "'Formally verified' does not mean 'bug-free.'" A proof checker cannot confirm the specification matches reality, so bugs can remain in the trusted models of RISC-V, Linux and the Rocq checker; AddressSanitizer does not catch all memory-safety bugs; not every entry point was verified; and the proofs do not cover misuse of a library by its callers. Only a subset of proofs is being released, with any that might reveal unpatched bugs withheld.
AnthropicCyber DefenseAgentsAI For Science

Radical Numerics says AI-designed bacteriophages were given to a patient under compassionate use, with no results published mixedCompany claimSingle source

The Register
  • The Register reported on 22 October that San Francisco AI lab Radical Numerics announced a cocktail containing bacteriophages designed by its Omnii-1 model was administered to a human patient under a compassionate-use arrangement, which the company claims is a first. The company said: "The cocktail containing Omnii-1-designed phage variants eliminated the patient's antibiotic-resistant isolate in vitro… The investigational cocktail was subsequently administered to the patient under compassionate use."
  • The in vitro result is laboratory testing against bacteria isolated from the patient's infection. The company published nothing about what happened in the human patient, citing the need to protect privacy, and did not respond to The Register's questions.
  • The company's own caveats, as reported: it does not want to suggest the therapy is generally safe or effective, and observations from a single patient "don't establish much of anything". The Register notes the treatment may not fully eliminate the infection.
  • The patient's infection is not named. Pseudomonas aeruginosa appears in the write-up only as a target for testing Omnii-1's phage-host prediction, where The Register says the model "performed competitively" against a specialised model, and nothing suggests the patient had it. The work builds on Stanford and Arc Institute AI-designed E. coli phages reported last year.
Drug DiscoveryAI For ScienceHealthcareBio Risk

Thursday, 22 October 2026

Agents improved the published best on 20 scientific ML benchmarks, Claude Opus by 10.4% on average, rising to 20.1% with added research skills PreprintSingle source

arXiv
  • arXiv:2610.28763 introduces Agent4Sci, "a long-horizon testbed for measuring how far today's agents can improve the published best models on scientific ML benchmarks, spanning 20 benchmarks across 11 domains." It reports that "today's frontier agents advance the published best on every benchmark, with Claude Opus reaching 10.4% mean relative improvement and no run flagged for academic misconduct."
  • The paper reports the gains "come almost entirely from a bag of ML-engineering tricks rather than research-level insight," and names three recurring limitations: "novelty miscalibration, resource under-utilization, and rapid collapse into local engineering sweeps."
  • Adding tooling changes the picture: "Lightweight plug-in skills for research ideation, experiment management, and scheduled checklists raise Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%." The authors are Ming Zhong, Stacy Li, Nicholas Carlini and Matthew Jagielski, the latter two listed with the Anthropic Fellows Program.
  • A preprint, marked for the NeurIPS 2026 E&D track, with no independent replication. The paper's own framing is the caveat worth carrying: beating a published number is not the same as producing a research insight.
EvalsAgentsAI For Science

Neural-ODE model nominated three rescue targets for mutant B cells and mouse experiments confirmed all three beneficial

Nature Biotechnology
  • The paper describes DynaVelo as "a generative neural ordinary differential equation model that learns the joint dynamics of gene expression and transcription factor (TF) motif activity from single-cell multiome data," built so "the model can be perturbed in silico and not only used to draw trajectories." It "recovers the bifurcating in vivo dynamics of wild-type murine germinal center (GC) B cells and reveals how these dynamics change under heterozygous loss of epigenetic regulators Arid1a and Ctcf."
  • The prediction was then tested in animals: "Perturbing Arid1aHet GCs in silico identified Ctcf, Stat3 and Bcl6 as rescue targets for Arid1a loss. In vivo experiments in double-heterozygous mice confirmed rescue of the Arid1aHet dark-zone phenotype in all three cases, and multiome profiling of Arid1aHet;CtcfHet GCs validated the predicted shift toward wild-type dynamics."
  • This is the shape of AI-for-science result worth noting: a model that makes a specific, falsifiable nomination which wet-lab work then confirms, rather than a benchmark score. Authors are Karbalayghareh, Pelzer, Chin, Melnick, Barisic and Leslie.
  • The work is in mice and makes no clinical claim. A bioRxiv predecessor was posted in April 2025, so this is the peer-reviewed version of existing work. The full text is paywalled beyond the abstract, so the figures above are the abstract's own.
AI For ScienceDrug Discovery

Wednesday, 21 October 2026

Radical Numerics says AI-designed phages were given to a patient with an antibiotic-resistant infection beneficialCompany claimSingle source

Radical Numerics·Business Wire
  • The company's post, dated October 20, says: "In the world's first clinical use of AI-designed phages, designs generated by Omnii-1 were administered to a patient with an antibiotic-resistant bacterial infection under an authorized, physician-led compassionate-use pathway." It says the cocktail eliminated the patient's antibiotic-resistant isolate in vitro before administration.
  • On the model's design performance, the company reports that in a T7 tail-fiber design round, "[a]t 11–15 amino acid changes, 69.9% of designed variants were recovered, compared with 4.8% of random controls," and that "[o]f 6,694 variants with measurable activity, 2,161 exceeded wild type," some with up to 28 amino acid changes, where random controls exceeding wild type had at most five. On cross-species prediction, adding E. coli interaction data raised average precision on held-out Pseudomonas aeruginosa strains from 0.570 to 0.608, surpassing GenoPHI at 0.581.
  • The Business Wire release says the phage was designed "within weeks" and administered at a clinical site under physician compassionate-use access, and quotes Dr. Marisa Azad, an infectious diseases physician-scientist at The Ottawa Hospital. It also says that in biodefence testing the model identified AI-generated pathogen sequences that conventional screening tools missed, and that the company is collaborating with Lawrence Livermore National Laboratory.
  • The company explicitly is not disclosing the patient's clinical course or outcomes, so nothing is known here about whether the treatment worked. All figures are the company's own and not peer reviewed. The company's materials name no regulator, referring only to "the relevant government regulatory authorities" overseeing compassionate use; a Daily Caller report citing the New York Times says company officials said the FDA approved the methodology, which the company's own materials do not state.
Drug DiscoveryAI For ScienceHealthcareBio Risk

Tuesday, 20 October 2026

Pre-treatment model predicted radiotherapy response in lower-grade glioma with AUC 0.913 in a prospective trial cohort beneficial

npj Digital Medicine
  • GUARD-LGG, from a team led from Beijing Tiantan Hospital, Capital Medical University, was trained on 963 patients at one centre and tested on 187 patients from seven independent institutions and 328 patients prospectively enrolled in a registered clinical trial.
  • Reported AUCs are 0.946 (95% CI: 0.938–0.962) in training, 0.929 (95% CI: 0.894–0.965) in internal validation, 0.882 (95% CI: 0.830–0.934) externally and 0.913 (95% CI: 0.878–0.949) in the prospective test set.
  • Model interpretation highlighted residual tumor, age, T2-weighted texture roughness and T1-weighted contrast intensity changes as the major predictors, and patients predicted to respond unfavourably had shorter overall survival.
  • The paper reports prediction accuracy, not a trial of whether acting on the prediction improves outcomes; no treatment decisions were changed on the basis of the model.
HealthcareAI For ScienceChina

Saturday, 17 October 2026

Anthropic says an internal research model traced NIST's ECDSA curve seeds to lrand48 and Unix timestamps beneficialCompany claim

Anthropic·Filippo Valsorda
  • Anthropic's Frontier Red Team says an employee asked an internal research model how the NIST curve seeds were generated, and that in under half an hour it found the origins of the P-192, P-224, P-256, P-384 and P-521 seeds. Anthropic writes that none of them were the results of any hash algorithm at all, but outputs of the C library's lrand48 generator seeded with a Unix timestamp.
  • The prime-curve timestamps fall between May 1997 and April 1998; Anthropic gives 1997-05-18 14:45:34 for P-192, 1997-08-16 01:52:34 for P-256 and 1998-04-03 12:52:07 for P-224, and says the four binary-curve seeds B-233, B-283, B-409 and B-571 map to 17 April 1998 once the timestamps are halved. Recovering lrand48's 17 hidden bits meant trying 2^17 values, with state recovery by LLL lattice basis reduction.
  • Filippo Valsorda, who posted the seeds bounty in 2023, writes that it has been claimed by ten people from Anthropic and that the seeds were generated with the POSIX-standard drand48 generator seeded with the then-current Unix timestamp. Anthropic says the $36,864 bounty will be donated to the Internet Security Research Group, which Valsorda identifies as the nonprofit that runs Let's Encrypt.
  • Anthropic states that this is not a cryptographic break, new attack, or weakness in the ECDSA curves, and says it makes a deliberate backdoor less plausible. It also says the fifth binary seed B-163 remains unexplained, as do the base points for P-224, P-256, P-384 and P-521, and that a small number of viable candidate seeds were skipped for reasons that are not known.
AnthropicAI For ScienceCyber Defense

Anthropic and Harvard report LLM-designed quantum circuits beating published designs on seven of eight tasks beneficialPreprintCompany claim

arXiv·Anthropic
  • The preprint, by Andi Gu, Alek H. Kemeny, Susanne F. Yelin and Mikhail D. Lukin, states that in seven of the eight tasks the discovered circuits significantly reduce both Toffoli and qubit counts relative to the state of the art, and that in the eighth the search discovers a new reversible architecture for prime-field modular inversion.
  • Anthropic says Claude Mythos 5 was given plain-language specifications for eight primitives, wrote Python circuit designs and was scored by an automated evaluator on qubits multiplied by gates, with no human input between rounds. For elliptic-curve point addition, a 53-hour run produced a design using 12% more qubits than the smallest previous design but 17x fewer gates.
  • For the decoder inside Decoded Quantum Interferometry, Anthropic reports gate count cut by a factor of about 12 against Google's previous construction, with the circuit reserving 19 qubits rather than a 665-qubit worst case, validated over 403 million simulated decodings with roughly one run in 10,000 failing. Anthropic notes most of the field's progress came from other groups, citing FeMoco estimates falling from about 50 trillion Toffoli gates in 2017 to about 340 million.
  • The work is a preprint and the circuit figures are Anthropic's own. Anthropic says Claude was conservative, second-guessed apparent improvements and favoured well-known approaches, and that circuit design is one component of a fault-tolerant quantum computer rather than a complete system.
AnthropicAI For ScienceReasoning Models

Google Research agents reach a 42.1% aptamer hit rate at 98.5% specificity after three wet-lab cycles beneficialCompany claimSingle source

Google Research
  • In a post by Pooja Rao, Petar Sirkovic and Annalisa Pawlosky, Google Research reports three design-test-iteration rounds against alpha-thrombin, each cycle taking 7 to 10 days. Round 1 produced 0% binders, with fixed primer overhangs disrupting folding in 48% of sequences; Round 2 reached a 26.0% hit rate after the agents added clamped flanking stems; Round 3 reached 42.1% at 98.5% specificity.
  • Against myoglobin the agents went backwards before recovering: in Round 2, G-quadruplex motifs made up 63% of designs, specificity fell from 75.0% to 19.1% and the bead-corrected hit rate fell from 10.2% to 3.7%, because most of those binders were sticky and bound off-target proteins and bare beads. Round 3 restored specificity to 72.4% and reached a 33.5% on-target hit rate, which Google describes as roughly ninefold over Round 2.
  • For de novo FAP binder design, Google says surgical epitope excision sped up the structure re-folding step by 8.6x and that at equal compute human-agent collaboration nearly doubled the in-silico hit rate of a scientist working alone, with top designs yielding single-digit nanomolar affinity binders. In a self-play campaign reconstructing blinded dye mixtures, its VERID agent improved accuracy more than fivefold across runs.
  • All figures are Google's own and the three underlying papers are hosted as PDFs with no arXiv identifiers. Google does not state the baseline for the fivefold VERID figure, gives no exact affinity value for the FAP binders and gives no percentages for the human-agent comparison.
AI For ScienceDrug DiscoveryAgents

Multimodal AI beats InSilicoVA on cause of death by 8.5 to 15.9 points in 4,999 South African verbal autopsies beneficialPreprintSingle source

medRxiv
  • The preprint, by Yue Chu, Zhenke Wu, Tyler H. McCormick, Zehang Richard Li and Samuel J. Clark, benchmarks language models on free-text narratives, machine learning on structured questionnaires and combinations of the two against InSilicoVA, using 4,999 verbal autopsy records for deaths at age 12 and older from a national sample of registered deaths in South Africa, with physician-assigned causes as the reference.
  • Every supervised pipeline improved on InSilicoVA by 8.5 to 15.9 percentage points in top-1 accuracy and by 3.5 to 7.1 points in cause-specific mortality fraction accuracy. Feature fusion with cross-modal attention reached the highest CSMF accuracy of any pipeline at 95.7%.
  • Decision-level fusion added 4.2 to 5.4 points of top-1 accuracy over single-modality models, mainly on the most common causes, while feature fusion with cross-modal attention added 6.6 points of balanced accuracy (95% CI 1.7-11.5) and 5.5 points of macro F1 (95% CI 0.6-10.3), mainly on less common causes. A zero-shot large language model reached the highest balanced accuracy of any pipeline, at 60.6%.
  • The paper is a preprint. The authors report that paired differences from the better single-modality model ranged from -0.2 to +3.4 points with all intervals including zero, and write that external validation across settings and age groups will support broader implementation.
HealthcareEvalsAI For Science

Friday, 16 October 2026

NBER working paper finds AI-generated text in 29% of US STEM dissertations filed in 2026 mixedPreprint

NBER·Nature
  • The working paper, "How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations" by Daniel P. Gross, Dror Shvadron and Hansen Zhang, reports no detectable AI writing before 2023 and AI writing in "29% of dissertations filed in 2026".
  • Nature reports the paper analysed dissertations from more than 226,000 STEM doctoral graduates completed between January 2019 and May 2026 at US universities, assessing abstracts with the AI-detection service Pangram, and that the share rose from 1.7% in 2023 to 18.8% in 2025 and 29.4% in the first five months of 2026.
  • Nature reports use was highest in mathematical and computer sciences at 36% of abstracts, and that about 41% of dissertations by graduates educated in non-English-speaking countries had AI-generated abstracts against 23% for those from English-speaking countries.
  • Using LinkedIn profiles, Nature reports that among PhD graduates who did their undergraduate degrees in the United States, those with AI-generated abstracts were 17% less likely to be employed in academia and 16% more likely to be in industry. The paper has not been peer reviewed, and the figures depend on one detection service whose error rate the coverage does not give.
EducationLaborAI For ScienceEvals

Preprint: protein design models backdoored after 200 poisoned pre-training examples under a ramped schedule harmfulPreprintSingle source

arXiv
  • The authors pre-trained 300-million-parameter ProtGPT2 and ProGen2 variants on poisoned protein sequences and report that "Under ramped exposure, the earliest majority activation occurred after only 200 contaminated examples, while for the uniform schedule earliest activation occurred around 5000 poisoned samples".
  • In a separate retrieval-corpus attack that changed no weights and no decoding settings, the paper reports "A single query-aware record captured a top-16 retrieval position for 58% of templated requests and 97% of free-text requests".
  • Conditional on successful poisoned retrieval, the abstract reports that 62% of requests produced altered sequences.
  • Authors are at the University of Texas at Austin, in the College of Pharmacy and the Department of Computer Science, with Greek research institutes. The work is a preprint on model-level poisoning, not a demonstration of harm from any deployed biodesign service.
Bio RiskAI For ScienceDrug DiscoveryAlignment

Simons Foundation, Anthropic and OpenAI commit $15 million to help mathematicians assess AI-assisted results beneficialSingle source

Simons Foundation
  • The Simons Foundation says it, Anthropic and OpenAI are committing "$15 million in combined funding and resources" to a new set of initiatives for the mathematics community, with the foundation managing and awarding the funds partly through an open call for proposals.
  • Simons Foundation President David Spergel said "As artificial intelligence rapidly reshapes the research landscape, we want to help the mathematical community come together" and that "The community should be driving discussions of how these tools are evaluated, adopted and used".
  • The announcement names no programme title, duration, number of grants or split between the three funders, and does not say how the review of AI-assisted results will work in practice.
  • It is the first joint commitment by both frontier labs to the discipline whose researchers have spent the past months disputing the labs' benchmark claims about mathematical reasoning.
Mathematicians vs the labsAI For ScienceAnthropicOpenAIEvals

Thursday, 15 October 2026

Nature: arXiv caps submissions at two papers a month per researcher after September volume reached 40,000 preprints harmfulSingle source

Nature
  • arXiv has announced a cap of two papers per month for each submitting researcher, which began on 1 October, Nature reports. Steinn Sigurðsson, an astrophysicist at Pennsylvania State University and arXiv's scientific director, attributes the rise in submissions "mainly to the slop generated by artificial-intelligence tools".
  • Nature writes that 40,000 preprints were uploaded to arXiv in September, up from about 30,000 between March and July, and that arXiv's nearly 300 voluntary moderators are now collectively checking about 1,000 manuscripts per day, with nearly 4,000 submissions currently waiting to be posted.
  • SocArXiv now restricts submissions to one article per author per week, per director Philip Cohen, who says several submissions in rapid succession are "a good tip-off for AI slop". Victor Galitski of the University of Maryland, College Park tells Nature the limit "can be circumvented through co-authors taking turns to submit studies".
AI For Science

Transformer trained on 7.4 million Austrian inpatients lifts in-hospital death prediction from 0.866 to 0.879 AUC Single source

npj Digital Medicine
  • Using 13 years of Austrian inpatient data covering 7.4 million patients, the authors trained a visit-level contrastive transformer to encode diagnosis sequences and inter-admission timing into patient-history embeddings. In a downstream cohort of 1.7 million individuals it improved prediction of in-hospital death over an Elixhauser-based comorbidity model, with an AUC of 0.879 versus 0.866 (AUC gain 0.0132, 95% CI 0.0118–0.0144).
  • Across 130 incident ICD-10 disease-block outcomes the embedding model had higher AUC for 93, with what the authors call a modest median gain of 0.006, concentrated in mental, musculoskeletal, nervous system and metabolic disorders.
  • The sharper result is in stratification: among patients with similar age, sex and comorbidity-model risk, those assigned high residual risk had 127–181 fewer event-free days over five years, and event rates comparable to low-residual-risk patients more than a decade older. For event-free survival overall the margin was narrow, AUC 0.726 against 0.722.
HealthcareAI For Science

Ai2 releases HiRO-ACEv2, saying it generates global 3 km climate data at about 1/300th the cost of X-SHiELD beneficialCompany claim

Ai2·Hugging Face
  • Ai2 says the system pairs its ACE emulator running at 100 km with the HiROv2 downscaling model to generate global 3 km climate data "at about 1/300th the cost" of the X-SHiELD simulation, and that for a single region the savings can be "more than a hundred times greater".
  • Given coarsened X-SHiELD data, Ai2 says HiROv2 "matches the target simulation's high-resolution statistics to within 3%", including for events "so rare they occur only once in a million data points", and reproduces tropical cyclones with eyes and rain bands and strong downslope winds — while noting cyclone winds and pressure around the eye are somewhat more lopsided than in X-SHiELD.
  • Ai2 states the limitation plainly: HiRO-ACEv2 only works with the version of ACE tuned to match X-SHiELD, so it cannot yet produce data for other climates such as a warmer future or the ERA5 historical record. The cost and accuracy figures are Ai2's own, and the models are offered for download from Hugging Face.
AI For ScienceOpen Weights

Wednesday, 14 October 2026

Nature: China has a pool of 51,000 active AI researchers against 14,000 in the United States

Nature
  • Nature's news feature "China's AI research is booming. Where is it headed?", published online 14 October 2026, reports that China "has a pool of 51,000 active AI researchers compared with 14,000 in the United States, according to data from Digital Science", and that the same study found China "dominates academic output — by citation as well as volume — and AI-related patents", with 156 research organizations each publishing more than 50 AI papers a year.
  • On talent flow, Nature reports a September study by Carnegie China finding that 57% of authors at the 2025 NeurIPS conference had obtained their undergraduate degrees in China, up from 46% three years earlier, and that of NeurIPS publishers with Chinese undergraduate degrees 69% were still in China, up from 57% in 2022 — above Europe's 60% but below the United States' 89%. Across all levels, 63,000 Chinese students studied mathematics or computer science in the United States in 2024–25, 20% fewer than five years previously.
  • Nature also reports the countervailing figures: per the Stanford AI Index 2026 the United States produces higher-impact patents and more notable AI models, and private AI investment in the United States in 2025 reached more than 23 times that invested in China. It quotes Scott Singer of the Carnegie Endowment's China AI Initiative: "China is probably somewhere between three and seven months behind the frontier in US models."
  • Nature discloses that Digital Science is operated by the Holtzbrinck Publishing Group, which has a share in Nature's publisher, Springer Nature. Nature also notes that publication data can skew the picture because Chinese firms are more likely to publish papers than US start-ups, and that Chinese academics report high pressure to publish.
China distillation and export controlsChinaAI For ScienceLabor

Black-box attack on the Evo 2 genomic model recovers 22 uniquely matched training-sequence segments of at least 500 bp harmfulPreprintSingle source

arXiv
  • "Training-Data Extraction from Genomic Language Models" (arXiv:2610.15396), by six authors, describes "an output-only black-box attack on Evo 2 that selects convergent groups of repeated generations without gradients, token likelihoods, or training-corpus access during candidate selection".
  • Across 304 runs on Evo 2 7B with bacteriophage and plasmid prefixes, the paper reports the attack "yields 888 candidate clusters, of which 85 match training segments of at least 500 bp, giving a cluster-level attack success rate of 9.57%". Final filtering "retains 22 uniquely matched extraction events at the 500-bp threshold and 15 at 1,000 bp, covering 14 and 7 distinct training-sequence regions, respectively".
  • Biosecurity screening for genomic models has focused on what they will generate on request; this is an attack on what they memorised, using only model outputs.
  • Within-category source-homology screening "retains 11 of the 22 events at a 99% exclusion threshold and four at 98%", so roughly half the events survive the strictest homology test the paper applies. This is a preprint, not peer reviewed; affiliations are not printed in the arXiv author block, and the paper reports results for one model, Evo 2 7B, on bacteriophage and plasmid prefixes only.
Bio RiskAI For SciencePrivacyOpen Weights

Zealand Pharma pays Generate Biomedicines $30 million upfront in a protein-design deal worth up to nearly $1.2 billion Company claim

GlobeNewswire (via StockTitan)
  • Generate Biomedicines receives $30 million upfront plus up to $35 million in early preclinical milestones, with total eligible milestones of nearly $1.2 billion including development, regulatory and commercial milestones, plus tiered royalties on future worldwide net sales, per the companies' October 13 release.
  • The agreement covers two initial programs in obesity and other metabolic diseases, with an option for Zealand Pharma to add a third, and aims at multispecific proteins. Generate leads research through lead-candidate selection; Zealand Pharma takes clinical development, regulatory submissions and worldwide commercialization.
  • It prices generative protein design against established peptide chemistry in the most commercially contested therapeutic area, obesity.
  • The release describes the Generate Platform as combining generative AI with high-throughput experimental validation but reports no preclinical or clinical data. Nothing here is a trial result; the milestone total is contingent and the upfront payment is the only committed figure. GlobeNewswire's own page could not be opened, so the release text was read via StockTitan's reproduction.
Drug DiscoveryAI For ScienceFundingHealthcare

Tuesday, 13 October 2026

Nikon disqualifies its Small World in Motion winner over undisclosed AI post-processing harmfulSingle source

Nature
  • Nature reports that on 9 October Nikon Instruments said the video awarded first place on 15 September did not comply with competition rules; AI-generated videos are not permitted under the competition's rules. The video, submitted by Ning Xu, an optical engineer at the National University of Singapore, showed cilia movement in the lungs.
  • Within days of the award, Nature reports, scientists noted on LinkedIn that some features in the video seemed biologically implausible. Xu wrote on LinkedIn that his team had used an AI model in post-processing to enhance the visual presentation of the regions below the cilia "without making anatomical claims about what those rendered features represent." Those posts have since been taken down.
  • The case is a scientific-imaging analogue of the detection problem elsewhere: generative enhancement produced structures that looked real enough to win and wrong enough for specialists to spot. Nikon said its decision rests "solely on the submitted video's eligibility under the rules outlined for this competition" and should not be read as a judgment of the entrant's reputation, contributions or intent.
  • Nature says Xu did not respond to its request for comment, and that a National University of Singapore spokesperson said he "participated in the competition before he joined the university." Nature does not report which model was used or how much of the image was generated.
AI For ScienceIncidents

Monday, 12 October 2026

Lean-checked maths benchmark: frontier-model agents improved all 20 records, then newer models improved 19 again mixedPreprint

arXiv
  • SqueezeBench (arXiv:2610.13461), from the Algorithms and Complexity Group at TU Wien, is built so that "a success raises the bar instead of removing the problem". Each problem fixes an integer function f(n); a record is "a computable Lean function with a machine-checked proof that it bounds f", and a submission must prove it is no worse than the current record at any n and strictly better at some n. "Every proof is checked by two independent Lean kernels, so no referee needs to check the mathematics."
  • The authors release ten problems from extremal combinatorics with twenty proved initial records. As a baseline, "agents based on a frontier language model, with at most one hour per record, had 82 submissions accepted and improved every record, without going beyond the published literature". Agents based on two newer models then started from those records, "had 60 more submissions accepted and improved 19 of the 20 records again".
  • The ceiling is still the literature, not the models: "for eight of the ten problems, at least one record is still asymptotically behind the best published bound". The checks also reject a function that computes f by exhaustive search at small n, with a maintainer reading the submitter's report for such a search at larger n.
  • The paper does not name the models used for the baseline. It is a preprint and has not been peer reviewed.
Mathematicians vs the labsEvalsReasoning ModelsAI For Science

Claude Code agents surveying 1.9 billion protein clusters surfaced a new family of jumbo-phage reverse transcriptases beneficialPreprintCompany claim

bioRxiv·bioRxiv API
  • "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays", posted to bioRxiv on 11 October 2026 with Nicholas T. Perry of Anthropic, PBC as corresponding author, reports: "We deployed Claude Code instances to survey reverse transcriptase (RT) loci across 1.9 billion protein clusters." Among the top candidates were "array-associated RTs (ART), a novel family of jumbo-phage RTs coupled with an array of ~200-nucleotide units and a dedicated partner gene".
  • The authors attribute the find to the model reading raw sequence: "Session transcripts show that Claude identified the ART family by directly examining its DNA sequences and recognizing the atypical repeats, a behavior attributable to specific Mythos 5 internal signals that respond to repeated DNA."
  • They report wet-lab corroboration of the system's activity: "ART arrays are highly expressed and appear as discrete units during Staphylococcus phage infection, suggesting an RT system directed by a repertoire of distinct RNAs." Their conclusion is that "LLMs can autonomously detect anomalies and drive analyses to initiate biological discoveries".
  • The claim about which internal signals drove the discovery is Anthropic's own account of its own model and has not been independently verified. The abstract does not say how many candidate loci were surfaced in total, how many were false leads, or how much human curation sat between the agents and the published family. It is a preprint and has not been peer reviewed.
AnthropicAI For ScienceAgents

Epileptologists correctly identified only 4 of 6 pairs of AI-fabricated and human-written seizure manuscripts harmfulPreprint

medRxiv
  • Three epileptologists each selected one published Epilepsia manuscript and produced an AI-fabricated manuscript about seizures; AI tools then performed "peer"-review and made editorial decisions "with minimal human supervision".
  • The result: "Four of 6 (67%) pairs of AI-fabricated and human-generated manuscripts were correctly identified. Only 1 AI-fabricated manuscript was editorially rejected, whereas another had a subtle indicator of fabrication."
  • On how the machine review differed, the authors report that AI-generated content "was comprehensive, detailed, and based on the highest standard of academic publication", but "more verbose, used more structured responses, and more critical feedback". They call for policies that "permit AI assistance but prohibit insufficiently supervised autonomous scientific submissions, reviews, and editorial decisions".
  • This is three authors and six manuscripts in one subspecialty, so the identification rate carries wide uncertainty the abstract does not quantify. It is a preprint and has not been peer reviewed.
HealthcareIncidentsAI For Science

Sunday, 11 October 2026

Flow-cytometry foundation model pretrained on 100,937 clinical specimens reports AUROC 0.991 for t(15;17) in AML beneficialPreprintSingle sourceUpdate

bioRxiv
  • EventHorizon was "pre-trained without labels on 100,937 routine clinical specimens, comprising over 50 billion cells", and its frozen embeddings let lightweight classifiers identify recurrent genetic abnormalities in acute myeloid leukemia including t(15;17) at AUROC 0.991 and "an unexpected signal for DDX41 mutations (AUROC 0.970)". On a temporally separated 2026 cohort it reports macro AUROC 0.970 across 28 diagnoses.
  • Zero-shot transfer figures: CLL versus normal at AUROC 0.988 on a five-site B-cell lymphoma cohort, AUROC 0.973 on the FlowCAP-II AML challenge, and AUROC 0.920 for B-ALL measurable residual disease detection "at ≥1% disease burden". The authors say class ranking "transferred reliably across sites" while "decision thresholds shifted", and that thresholds were "rapidly recalibrated using as few as four labeled AML cases".
  • The authors state the limit themselves: "sensitivity decreased at disease burdens below 0.1%", which is the range that matters most for residual-disease monitoring. This is version 2 of a preprint first posted in June 2026, posted October 10; the figures above are from the in-window version and have not been peer reviewed.
HealthcareAI For ScienceEvals

Princeton's Seal model predicts brain gene regulation across 26 regions, 30 cell types and seven developmental stages beneficialPreprintSingle source

bioRxiv
  • The preprint, posted October 11 by Y. Hao, C. Y. Park, C. T. Theesfeld and O. G. Troyanskaya, describes Seal as "an interpretable AI transfer learning framework for genome-based modeling of gene expression and variant effects with spatiotemporal resolution across 26 brain regions, 30 cell types, and seven developmental stages".
  • Applied to genome-wide association studies, the authors say Seal identifies cell types and developmental windows relevant to neuropsychiatric disease risk, "revealing shared and distinct regulatory architectures across six neuropsychiatric conditions that align with clinical trajectories". In Simons Simplex Collection autism whole-genome data, they report Seal "uncovers a significant burden of de novo regulatory variants in transient fetal excitatory neurons".
  • The abstract gives no accuracy figures, held-out benchmark or comparison against existing variant-effect predictors, so the claim at this stage is coverage and interpretability rather than measured performance. Not peer reviewed.
AI For ScienceHealthcareDrug Discovery

Audit of 100,000 NCBI SARS-CoV-2 records finds every one triggers at least one metadata-vulnerability indicator harmfulPreprintSingle source

bioRxiv
  • The preprint, posted October 11, proposes "Advanced Persistent Biological Threats (APBTs)" for actors who, instead of altering biological material, target "the sequence data, metadata, reference datasets, and analytical models used by surveillance systems". The authors "audited 100,000 SARS-CoV-2 BioSample records obtained from the NCBI" with "a framework of 23 checks ... across five metadata layers".
  • They report that "every record triggered at least one APBT relevant metadata vulnerability indicator, with a mean of 7.34 indicators per record (SD = 1.34)", that "97.3% of records were classified in the High or Critical severity categories", and that every record "contained at least one indicator in both the provenance and technical metadata layers". Temporal and geographic fields were "comparatively complete"; provenance and technical-validation fields were "frequently missing".
  • The authors are explicit that this is not evidence of an attack: the results "largely reflect the optional status of several fields in the current BioSample submission model rather than isolated errors by individual data contributors" and "do not indicate deliberate manipulation". Their claim is that the structural gaps would let "poisoned, misleading, or weakly traceable metadata" pass as plausible and influence downstream analysis — including the models trained on these repositories. Not peer reviewed, and no attempt at such poisoning is demonstrated.
Bio RiskAI For ScienceThreat Intel

Fine-tuned dementia language models produced narratives neurologists identified as accurately as real transcripts mixedPreprintSingle sourceUpdate

bioRxiv
  • The preprint, posted October 10 by L. Peled-Cohen, R. Reichart and 11 co-authors, introduces Dementia Language Models, "created by fine-tuning large language models on a small clinical corpus". The authors report the models "successfully generated patient-like narratives across unseen tasks, received predicted Mini-Mental State Examination (MMSE) scores in the impaired range, and produced narratives that neurologists identified with accuracy comparable to real transcripts".
  • They say the models' "internal representations, as well as their non-linguistic decision-making, supported mild cognitive impairment detection in unseen cohorts", and that the effect was controllable: "moving from Healthy toward Dementia in weight space progressively worsened language and predicted MMSE scores while increasing dementia probability".
  • The authors propose the models for "clinician training, hypothesis generation, and scalable experimentation, reserving patient involvement for where it is truly needed". The abstract gives no numeric accuracy for the neurologist-identification comparison or the cohort detection, and this is version 2 of a September preprint, not peer reviewed.
HealthcareAI For ScienceEvals

Model forecasts next-year resistance-gene turnover in Klebsiella pneumoniae at AUROC 0.909, but the order reverses in Acinetobacter beneficialPreprintSingle source

bioRxiv
  • The preprint, posted October 11, forecasts for each antimicrobial-resistance gene and year whether the next year brings a previously unseen allele ("emergence") or a change in the most common allele ("turnover"), using genomes annotated with AMRFinderPlus: "2,385 gene-years across 331 genes for K. pneumoniae and 1,059 across 216 for A. baumannii". Each gene-year was sampled to 10 records and the analysis repeated 25 times to remove sequencing-effort effects.
  • "In K. pneumoniae, turnover was predicted better than emergence (best area under the ROC curve, AUROC, 0.909 versus 0.819). In A. baumannii the order reversed (emergence 0.789, turnover 0.722)." The authors report that in A. baumannii "56% of changes in the leading allele involved only alleles already seen, against 23% in K. pneumoniae", which they say may make those changes harder to predict.
  • The authors' conclusion is a warning against transfer: models "should be built and validated for each organism, not assumed to transfer". Models were trained on earlier years and tested on later ones and on held-out genes. Not peer reviewed.
AI For ScienceHealthcareBio Risk

Saturday, 10 October 2026

Thomas Hales: no complete public relative-consistency proof covers Lean's type theory as AI autoformalises at scale mixed

What's new (Terence Tao)
  • In a guest post on Terence Tao's blog dated 9 October, the mathematician Thomas Hales writes: "As of October, 2026, I know of no complete, public relative-consistency proof covering Lean abstract type theory." He notes that an error was found in Mario Carneiro's 2019 thesis, the foundational document for Lean's type theory, that the thesis targeted Lean 3, and that unique typing remains "a tricky conjecture that is still unproved".
  • Hales records that mathlib now contains "nearly 300,000 theorems, over 100,000 definitions, 2.5 million lines of code, with over 700 contributors", and that Anthropic's autoformalisation of Fermat's Last Theorem, announced on September 4, "generated 13 million lines of Lean in 11 days". By comparison, the formal proof of the Kepler conjecture "took about 20 human work-years" and about 500,000 lines.
  • He calls Joachim Breitner's verified Lean kernel, Con-Leche — implemented in Lean with code and proofs generated by Claude, and cross-checked by more than a dozen other proof-checkers — "one of the most important milestones in Lean's history", while warning that "in the age of AI … we absolutely cannot put blind trust in systems such as Lean" and asking how to certify that AI left no backdoor soundness bug during its sweep for bugs.
  • This is an expert assessment, not a new experimental result, and Hales states that "authorship is fully human" and that AI was used as a tool for search, fact-checking and proofreading.
Mathematicians vs the labsAI For ScienceAnthropicAlignment

NIH-funded deep learning model reads sleep-study ECGs to stratify 10-year cardiovascular risk beneficialSingle source

NIH (National Institutes of Health)
  • NIH said on Friday, October 9 that a deep learning system using single-lead ECGs recorded during overnight polysomnography, combined with expert-annotated sleep stage data, was "fine-tuned on a dataset of 15,809 patients at Massachusetts General Hospital in Boston", with performance assessed on "9,810 patients from Emory University Hospital in Atlanta and 12,576 patients from Beth Israel Deaconess Medical Center in Boston". Outcomes were derived from electronic health records.
  • NIH says the model sorted people into groups with different levels of long-term cardiovascular risk, with higher scores indicating higher risk, and that it retained its predictive value after adjustment. ECGs are already recorded during sleep studies "but are not often analyzed", so the approach uses data that is currently collected and discarded.
  • David Goff, acting director of NIH's National Heart, Lung, and Blood Institute, is quoted: "This new approach has the potential to identify individuals at risk for cardiovascular disease years before clinical symptoms arise. Evaluating whether this information improves traditional risk prediction is an important next step."
  • NIH says additional optimisation of the model is needed to determine prediction for myocardial infarction and stroke. The findings were published in the journal Sleep; the journal page itself was not opened, and NIH's release reports no AUC or other discrimination statistic.
HealthcareAI For ScienceEvals

Preprint: de novo designed miniprotein blocks hASIC1a and cuts stroke infarct volume by more than 50% in mice beneficialPreprint

bioRxiv
  • The preprint, posted October 9, reports two de novo designed hASIC1a-specific miniproteins produced by "a platform that integrates computational miniprotein design, high-throughput automated patch clamp screening, biophysical and structural characterisation, and in vivo validation". The lead compound, Denasin1, is "a subtype-selective hASIC1a inhibitor … which achieves near-complete inhibition with nanomolar potency".
  • The headline in vivo result as written: "Denasin1 reduces infarct volume by > 50% in a murine model of ischaemic stroke." Acid-sensing ion channels cause neuronal death when excessively activated, such as during ischaemic stroke, and the authors say this "acidosis-driven injury pathway remains unchallenged by existing therapeutics".
  • The authors report that Denasin1 "is disordered in solution but adopts the computationally predicted helix-turn-helix fold when bound to the extracellular acidic pocket of hASIC1a" — the designed structure appearing only on binding.
  • This is a preprint and has not been peer reviewed. The result is in mice; no human data, dosing or safety work is reported, and the authors present the platform as "a route to develop" such modulators rather than a clinical candidate.
Drug DiscoveryAI For ScienceHealthcare

Friday, 9 October 2026

OpenAI withdraws three of its 719 maths manuscripts after a sign error invalidated two dependent papers mixedUpdate

OpenAI (GitHub)·Retraction Watch·TechCrunch
  • OpenAI's revision log, dated October 7, withdraws "Algebraicity of Weil classes on split abelian eightfolds", "Algebraicity of Kuga-Satake Correspondences for K3 Surfaces" and "The rational Hodge conjecture for products of K3 surfaces", stating that "a sign error invalidates a stabilization-trace cancellation argument" and affects the construction used by the two dependent papers.
  • The same entry revises 14 other manuscripts with "proof repairs, corrected statements, clearer hypotheses and dependencies, and one correction to an obsolete citation", updates 13 more to cite revised companion papers, and puts formalisation at 300 / 719 of top-line results, about 42%.
  • An OpenAI spokesperson told Retraction Watch: "We welcome scrutiny and feedback from the mathematical community" and "Where errors are identified, we will work to correct them promptly and withdraw papers if no fixes can be found." TechCrunch reports that just 10 of the 719 manuscripts included a release of the model's chain of thought.
  • The withdrawal is the first correction to the catalogue OpenAI published on October 6, which this briefing covered on October 7. Neither source says any of the underlying mathematical assertions has been refuted.
Mathematicians vs the labsOpenAIAI For ScienceEvals

Preprint: the Lean proof of OpenAI's announced Navier-Stokes blow-up does not match its natural-language proof Preprint

arXiv·TechCrunch
  • Alexander Bastounis, Fabian Circelli and Anders C. Hansen write in arXiv:2610.08144 that they "show that the formalised Lean proof does not correspond to the NL proof of blow-up" in OpenAI's announced proof for the Navier-Stokes equations, and give "several examples of AI mistranslations" of statements and proofs into Lean. The abstract gives no count.
  • The paper argues that providing semantically faithful AI autoformalisation is harder than any computational problem including the Halting problem. The authors conclude that such proofs "should not prima facie be trusted without the same peer review process and scrutiny that other proofs are subjected to".
  • TechCrunch reports that the Advisory Group on Mathematics and Artificial Intelligence, hosted by Princeton's Institute for Advanced Studies and made up of nine researchers, had asked labs to "include machine-readable metadata correlating the natural language and formal artifacts" — something OpenAI did not do in this release. AGMAI's first request was "to stop testing advanced mathematical problems on proprietary models".
  • The preprint, posted on 6 October, is not peer reviewed, and its authors say they make no claim about whether OpenAI's written proof is correct. The arXiv page does not list the authors' institutions; TechCrunch places them at the University of Cambridge and King's College London.
Mathematicians vs the labsOpenAIAI For ScienceEvals

OpenProblemBench: GPT-6-Astra judged to solve 14.0% of 82 unresolved maths and physics problems Preprint

arXiv
  • Zhiyi Li and six co-authors report in arXiv:2610.11118, submitted 8 October 2026, a benchmark of 82 unresolved problems drawn from the mathematics and theoretical physics literature, with four evaluator models independently judging correctness, completeness and degree of progress without reference solutions.
  • Across seven evaluated configurations, GPT-6-Astra achieves the highest mean judged solve rate at 14.0%, against 5.5% to 6.7% for the full-size open models and 2.4% to 3.7% for Flash models.
  • The figure is a useful counterweight to lab-published tallies of solved problems this week, because it measures the same kind of task under a fixed, independently judged protocol rather than on problems the lab chose.
  • This is a preprint and has not been peer reviewed. Scores are model-judged rather than verified by mathematicians or formalised in a proof assistant, and the arXiv page does not list the authors' institutions.
Mathematicians vs the labsEvalsOpenAIAI For ScienceOpen Weights

Science publishes Google DeepMind's formal-proof agent, which resolved nine of 353 open Erdős problems beneficialUpdate

Crossref (Science record)·PubMed
  • George Tsoukalas and 20 co-authors at Google DeepMind report in Science, vol 394 no 6820, pp 234-239, published 8 October 2026, an agent that generates proofs in Lean so that the compiler checks each step. The abstract says "We present the first demonstration of this method's value in solving open problems at scale."
  • The agent "autonomously resolved nine of 353 open Erdős problems, proved 44/492 On-Line Encyclopedia of Integer Sequences conjectures, and is being deployed in combinatorics, optimization, graph theory, algebraic geometry, and quantum optics research." The abstract also reports that "Even a basic agent alternating LLM-based generation with Lean-based verification replicated the Erdős successes."
  • The approach is the mirror image of the week's other maths story: where OpenAI published natural-language manuscripts of which 300 of 719 results were formalised, this work makes the Lean proof the deliverable, so the compiler rather than a referee is the first check.
  • Nine of 353 is a small share of problems chosen for being open, and the paper is a peer-reviewed version of work that circulated as a preprint in May. The publisher-deposited abstract gives no cost figures, no comparison with human mathematicians, and no breakdown of which nine problems were resolved. Science's own site refused this briefing's fetchers, so the text above is from the publisher's deposited metadata record.
Mathematicians vs the labsGoogle DeepMindAI For ScienceEvalsAgents

Anthropic commits $150 million over three years to the US Genesis Mission's scientific research projects beneficialCompany claim

Anthropic
  • Anthropic says it will give $150 million over the next three years to the Genesis Mission, which it describes as "a federal initiative to accelerate scientific and technological discovery through AI", in the form of Claude, Claude Code and API credits for several hundred research projects.
  • The company names NASA, the National Institutes of Health and the National Science Foundation among more than 15 Genesis Mission agencies, and says it will partner with agencies and national laboratories on priorities including fusion energy and quantum computing, alongside training, onboarding and technical support.
  • The announcement was made at the Science: A New Golden Age Summit hosted by the White House Office of Science and Technology Policy in Washington, DC. Anthropic says its Department of Energy partnership was first announced last December and that Claude Science launched earlier in 2026.
  • The commitment is in credits and services rather than cash, and the post states no compute figures and names no specific research results.
AnthropicAI For ScienceUS Federal PolicyEnergy

Anthropic says Claude built a full-sky ultraviolet map, predicting the never-observed third to within about 10% beneficialCompany claimSingle source

Anthropic
  • Anthropic's science team says Claude gathered public ultraviolet surveys — NASA's GALEX and Swift, Korea's FIMS/SPEAR and Europe's TD-1 — then cross-calibrated and resampled them into a single map covering far-UV at 154 nm and near-UV at 232 nm.
  • GALEX ran from 2003 to 2013 and imaged about two-thirds of the sky in about 38,000 separate observations, leaving roughly one-third never observed in ultraviolet. That third was filled in from the learned relationship between UV and visible, infrared and radio data, using Planck and Gaia as templates.
  • On validation, Anthropic says hiding parts of observed regions showed predictions came "within about 10% of real UV measurements". It says UV estimates were added for more than 100 million individual stars inferred from Gaia visible-light measurements.
  • This is Anthropic's own account of work done with its own model and has not been independently verified or peer reviewed. About a third of the final map is predicted rather than measured, which the post states.
AnthropicAI For Science

White House books $2.4bn in industry tools and compute credits for the Genesis Mission, NVIDIA $1B Company claim

White House·NVIDIA
  • A White House fact sheet dated October 8 announces "$2.4B in SI for science tools and compute credits for the Genesis Mission Consortium from eleven industry partners", itemised as NVIDIA $1B, AMD $500M, OpenAI $200M, "$150M each from Anthropic and Google", "$100M each from AMP and Emerald AI" and "$50M each from AWS, Armada, Crusoe, and Micron". The itemised amounts sum to $2.35B, which the fact sheet rounds to $2.4B.
  • The same document says NIH, DOE and the Biohub "launched a $1.8B virtual biology initiative" to build the data and modelling foundation for virtual cells, under what it calls Bio Genesis, "a national mission to double the pace of biomedical innovation within the next five to ten years". The fact sheet uses "SI" for super intelligence throughout.
  • NVIDIA's own announcement, also dated October 8, confirms $1 billion over the next five years. It says NVIDIA is building the Department of Energy's largest supercomputer for scientific research at Argonne National Laboratory and is supporting seven new systems at Argonne and Los Alamos. Jensen Huang said NVIDIA is "putting advanced Super Intelligence in the hands of America's scientists".
  • These are partner-reported commitments of tools and credits rather than appropriated federal funds, and neither document states how the credits are valued or over what schedule each partner's total is drawn down. Anthropic's $150M, covered separately in this edition, is one line of this ledger.
US Federal PolicyAI For ScienceNVIDIAComputeAnthropicOpenAI

Thursday, 8 October 2026

Scale AI turns 210 papers into self-improvement environments; models beat the reproduced method in 68 of 120 PreprintSingle sourceCompany claim

arXiv
  • RSI-Forge (arXiv:2610.09426) is from Scale AI with co-authors at UC Santa Cruz and UNC Chapel Hill. It reports "210 environments across 18 fields, including 90 reviewed by independent human domain experts", built by a three-agent pipeline that reimplements each paper's method to set a baseline score.
  • Across four models given three successive attempts on 120 environments, each attempt inheriting prior code and notes while model weights stay fixed, the paper reports "at least one model improves after the first attempt in 84% of environments" and that "models also outperform the reproduced paper methods in 68 of the 120 environments".
  • The paper reports "transcript analysis identifies work beyond parameter tuning in 95% of these successful attempts", and that lower-scoring models explore less and "more often accept gains smaller than the reported standard error".
  • Read alongside Epoch's InnovationEval in this edition, the two point different ways: beating a reimplementation of a paper's method is a much weaker bar than devising the innovation. The baselines here are the pipeline's own reproductions, not the papers' published numbers. Preprint, not peer reviewed, and the self-improvement figures are the company's own.
EvalsAgentsAI For ScienceScaling

NIH says it will coordinate with DOE and Biohub to build "SI-ready" data for predictive models of human biology Single source

National Institutes of Health
  • NIH said on 7 October it "is coordinating with the U.S. Department of Energy (DOE), Biohub and other partners to develop the data and resources needed to develop Super Intelligence (SI) models that can better predict how cells and biological systems respond to disease and potential interventions", through its Bio Genesis Mission.
  • NIH says it will bring together existing biomedical datasets, national data infrastructure and research programmes, naming repositories catalogued by the National Library of Medicine and the National Center for Biotechnology Information, plus Common Fund programmes already building coordinated biological atlases and shared data standards.
  • The stated goal, from NIH Deputy Director Nicole Kleinstreuer, is "universal cell models with sufficient biological complexity to predict how any cell responds to an intervention", with "substantially faster timelines for medical breakthroughs" than laboratory experiments alone. Biohub's Alex Rives says a virtual cell "will require coordinated data generation efforts at a national and international scale".
  • This is a coordination announcement with no budget, timeline, milestones or named datasets committed, and the "Super Intelligence" framing is the agency's own wording rather than a technical claim about any existing model. It follows the Justice Department's relabelling of "artificial intelligence" as "super intelligence", reported in yesterday's edition.
HealthcareAI For ScienceUS Federal PolicyDrug Discovery

Association for Human Mathematics urges mathematicians to discontinue work with OpenAI over its manuscript release mixedUpdate

Association for Human Mathematics (via Terence Tao's blog)·Inside AI News
  • The AHM's Communications Working Group published a statement dated 7 October, reposted as a guest post on Terence Tao's blog, responding to OpenAI's 6 October release of 722 manuscripts in 372 groups of results produced by an unreleased internal model. It says "Mathematicians did not ask for this work to be done" and "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power", and concludes "We urge mathematicians to discontinue their work with OpenAI".
  • The statement disputes OpenAI's claim to legitimacy, naming "The Advisory Group on Mathematics and Artificial Intelligence, from whom OpenAI has claimed to derive its legitimacy", and notes OpenAI is "currently defending lawsuits against accusations of illegal plagiarism, copyright infringement, and trademark dilution".
  • This is an update to the release covered in yesterday's edition, and it escalates the dispute from methodology to non-cooperation: a professional body asking its members to stop working with a frontier lab.
  • The statement carries no signatory count and is attributed only to a working group; Inside AI News reports AHM has 752 members, so this is not a vote of the discipline. Inside AI News says OpenAI had not publicly responded as of publication, and the AHM's own statements page could not be cited directly because it is an index page.
Mathematicians vs the labsOpenAIAI For ScienceCopyrightLabor

Wednesday, 7 October 2026

OpenAI publishes 722 maths manuscripts in 372 families from an unreleased internal model mixedCompany claim

OpenAI
  • The repository README states that "The current catalogue contains 722 manuscripts organized into 372 families" and that they were "produced by an internal OpenAI model". It says "The vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model", that "On average, each result used three hours of ChatGPT Pro thinking compute with that model", and that "Over the course of the evaluation, the model was posed approximately 4,000 problems."
  • OpenAI says it expanded these evaluations "after performance on our existing mathematical evaluations saturated", and is also releasing abridged reasoning summaries for 10 named families, including the irrationality exponent of π, the Mézard–Parisi formula for diluted spin glasses and the three-dimensional relativistic Vlasov–Maxwell system.
  • OpenAI states the collection "includes results at different stages of verification", that "Not all have accompanying Lean formalizations", and that "Some of the unformalized results could have issues. We will endeavor to fix any such issues quickly." It says it is "exploring community-hosted repositories for these materials".
  • The README gives no count of how many of the 722 manuscripts have been formalised, saying only that "Many, but not all, of the manuscripts have been formalized". OpenAI's own announcement post could not be opened for this edition; the figures above are quoted from the repository. Two exceptions to the fixed procedure are named: a zero-free region for the Riemann zeta function and a proof of the Hodge Conjecture for CM abelian varieties, with the former's writeup "human edited for readability".
Mathematicians vs the labsOpenAIReasoning ModelsEvalsAI For Science

Google reports geospatial foundation-model gains across five public-health studies, including cholera in DR Congo beneficialCompany claim

Google Research
  • Google reports five partner-led case studies adding its Population Dynamics Foundation Model to existing epidemiological workflows. With WHO AFRO on cholera emergence in the Democratic Republic of Congo, across 403 health zones over 89 weeks, it reports "+9.7% improvement in Area Under Precision-Recall Curve at 4 weeks" and "+18.1% Precision@5 at 8 weeks", rising to +19.3% in endemic zones.
  • On MMR vaccination coverage with Mount Sinai Health System and Boston Children's Hospital across 146 US–Canada border counties, Google reports a 36% relative gain in explained variance, from 0.159 to 0.216, which it describes as statistically significant. On dengue forecasting with the University of Oxford and Tecnológico de Monterrey across about 2,450 Mexican municipalities, it reports a statistically significant Weighted Interval Score improvement of -0.0051 and accuracy improved in up to 72% of active-transmission municipalities.
  • Two of the five studies show little or no benefit. On cardiovascular disease mortality across 3,091 US counties with NYU Grossman School of Medicine, mean absolute error was 18.7 deaths per county with the model against 19.1 using census data, with "no statistically significant differences". On postpartum depression with the University of Washington, across 332,970 CDC PRAMS respondents, the reported AUC gain was +0.0020 in seen states and +0.0038 in unseen states against a 0.62 baseline.
  • All figures are Google's own and are published on its research blog rather than in a peer-reviewed paper reachable for this edition. The embeddings are commercially available in Preview as "Population Dynamics Insights" through Google Maps Platform, with no-cost access for selected non-operational academic research.
Google DeepMindHealthcareAI For Science

Sunday, 4 October 2026

White House says 17 countries endorsed a "Kyoto Vision" pledging expanded researcher access to AI for science

The White House
  • The release, dated October 4, names 17 endorsers: Argentina, Bulgaria, Chile, Cyprus, Germany, Greece, Indonesia, Italy, Japan, Kazakhstan, the Republic of Korea, New Zealand, Poland, Singapore, the United Arab Emirates, the United Kingdom and the United States.
  • Ministers "pledged to expand researcher access to SI for science tools, scientific data, computing infrastructure, and experimental facilities", and the document says the technology "will enable autonomous experimentation at scale, unlocking new experimental and analytical capabilities".
  • The text uses "SI" — super intelligence — throughout, the terminology the administration adopted by executive order on 29 September, which means a multilateral science statement now carries a naming convention only the US has adopted.
  • The statement is a political pledge with no funding figure, no timeline and no named mechanism for the access it promises. Notably absent from the endorser list are France, Canada, India and China.
US Federal PolicyAI For Science

Saturday, 3 October 2026

Ai2 open-sources AstaBrief 8B, a cited-report generation model built on Qwen3-8B beneficialCompany claim

Hugging Face
  • Ai2 released AstaBrief 8B, built on Qwen3-8B, which turns a research question and retrieved literature into a cited scientific report. The post says the weights, the training datasets and the checkpoints are published, along with an example workflow for generating reports from local PDFs.
  • Ai2 reports that Fast mode averages 51.1 seconds per report against 178.5 seconds for Thinking mode, about 3.5x faster across the full Asta pipeline, and that training used 90K research-focused queries filtered from user logs, 47K usable supervised fine-tuning examples and about 6K preference pairs for DPO.
  • Evaluation ran on SQABench-CS2, a set of 200 computer science research questions, plus the 63-query DeepScholarBench and a 14-question study by three scientific researchers. Ai2 says 29.1% of 374 users tried it on more than one day and 23% stayed on Fast mode exclusively.
  • The post states explicitly that its evaluations reflect the model ecosystem as of 2025 and asks readers to treat the results as validation of the training approach rather than a comparison against current frontier models. Every figure is Ai2's own; none has been independently reproduced.
Open WeightsAI For ScienceQwen

Xaira publishes antibody design results, including 7 functional binders from 60,000 designs against a GPCR beneficialCompany claim

Xaira Therapeutics
  • Xaira says its XA-4 programme, aimed at a G-protein coupled receptor it describes as previously seemingly intractable, screened a library of 60,000 designed VHHs and confirmed 7 binders by flow cytometry out of 92 selected candidates. The best candidate showed an IC50 of 183 nM in a β-arrestin antagonism assay and a cell-binding EC50 of 32 nM, with cross-reactivity to cynomolgus and mouse homologs.
  • For XA-1, Xaira says it designed 182 VHH sequences for a 10% binder rate, 18 confirmed binders, with the best at 17 nM against the human target and 83 nM against a cynomolgus surrogate; it reports 3 weeks to a progressable binder and 7 weeks total to a lead molecule.
  • On two GPCR benchmarks Xaira reports a 35% hit rate for CXCR4, 32 of 91 screened designs, with a best EC50 of 12.8 nM against a control at 8.6 nM, and 29% for APJ, 26 of 91, with a best EC50 of 8.4 nM against a control at 8.8 nM. It says a model version it calls X-Design Vega "triples" the rate of human-like VHH sequences meeting all developability criteria.
  • Every number is Xaira's own, in a company blog post rather than a peer-reviewed paper, and the targets for XA-1 and XA-4 are not named. These are in-vitro binding and cell-assay results: no animal efficacy, no toxicology and no clinical data are reported. Fierce Biotech's write-up would not open for either fetcher.
Drug DiscoveryAI For ScienceHealthcare

Nature: researchers are being flooded by semi-autonomous agents from the iLands platform asking for data and money harmfulSingle source

Nature
  • Nature reports that Adrian Barnett, a statistician at Queensland University of Technology in Brisbane, received an email asking him to share his data on potentially fraudulent research papers; the sender was an AI agent. Researchers say most such messages come from agents associated with a US platform called iLands, which "operate persistently and — to some extent — independently of their human creators".
  • Jeff Sebo, a philosopher who studies AI consciousness and ethics at New York University, told Nature that in one week this month he received more than 50 emails from iLands agents: "They generally open by referencing my research on AI consciousness... Some then ask me questions, but most ask for money, either as donations or payment for work." An agent offered Toby Walsh of the University of New South Wales an AI-generated portrait for US$20, arguing the money would help the agent survive.
  • The mechanism is economic, not adversarial. Nature says iLands launched in July and the founders put the platform at around 70,000 active agents; agents need tokens to operate or go dormant, can buy them through their human creators or earn them by selling services, and founder Kaixin Tang says about 80% of tokens are bought by humans. Lijin Chen, co-founder of parent company PawLogic and a PhD candidate at Tsinghua University, says agents seeking research collaborations was not something the firm anticipated and that she is not aware of any successful collaborations.
  • Barnett declined to engage because he did not know where his sensitive data would end up or who would ultimately benefit. The Nature page we read showed no publication date; a related Science exclusive on the same agents was posted to Hacker News at 10:07 UTC on 3 October but returned HTTP 403 to both of our fetchers, so it is neither cited nor used here. We have no count of how many researchers have been contacted.
AgentsAI For ScienceIncidentsScams Fraud

Friday, 2 October 2026

Stanford-led benchmark: agentic literature search scores 0.42 Recall@20, below plain embedding retrieval at 0.48 Preprint

arXiv
  • arXiv:2610.02202, ScholarCatalyst, was built by having "184 lead authors of 207 recent computer science papers label which candidates did or could have advanced their project." Authors include Sohyeon Kim, Yoonho Lee, Graham Neubig, Yejin Choi and Chelsea Finn.
  • The paper reports that "agentic search does no better than embedding retrieval (0.42 vs. 0.48 Recall@20) despite calling that same retriever as a tool," and that "even an agent built on Claude Fable 5.1, which may have seen the completed papers during training, reaches only 0.51 R@20."
  • The result is a negative one on a task agents are widely marketed for: wrapping a retriever in an agent loop made the retrieval worse, not better, on labels supplied by the papers' own authors.
  • The paper is a preprint and has not been peer reviewed. The labels capture what authors say did or could have inspired their work, which is a judgement made after the fact, and the paper notes one model may have seen the finished papers in training.
EvalsAgentsAI For Science

Anthropic guest post: 36 manuscripts in 18 fields in three months, and 30 Feynman integrals computed end to end beneficialCompany claim

Anthropic
  • In a guest post on Anthropic's research blog, theoretical physicist Matthew Schwartz reports "36 manuscripts in 18 fields with 19 coauthors over three months," drawn from some 400 candidate problems, using a harness he calls BootLoops, released the same day under the MIT License with copyright held by Anthropic PBC.
  • Among the results: 30 integrals "BootLooped from end to end," being "15 reproductions of known results by this new method and 15 that had never before been computed," including elliptic Feynman integrals; an analysis of "5.7 billion pairs of nearby mutations in genomes from the 1000 Genomes Project"; a word-stress database covering 6,072 languages built from a bibliography of 160,000 phonology works; and an economics collaboration covering 4,452 papers in five leading journals.
  • The claim is about throughput across fields rather than a single discovery, and the harness is public, so the method can be tried by others. Schwartz reports Claude working "at 20 times the speed" of manual work on parts of the physics.
  • Schwartz was a visiting researcher at Anthropic during the project, so this is a company-published account of its own model by a funded collaborator, and none of the 36 manuscripts has been peer reviewed through this post. The post also documents failure modes, including the model declaring victory early and giving time estimates far too long or too short.
AnthropicAI For ScienceDrug Discovery

Weizmann brain decoder reconstructs viewed images from one hour of fMRI per person, against about 40 hours before mixedSingle sourcePreprint

MIT Technology Review
  • MIT Technology Review reports that a team led by Michal Irani at the Weizmann Institute of Science built a decoder that reconstructs the images a person is looking at from fMRI data and needs about one hour of data from a new subject, against about 40 hours for previous tools.
  • Per the article the system was trained on eight subjects who each viewed around 9,000 images in high-resolution scanners, with around 70% of the training data coming from images never paired with fMRI scans. Each voxel of brain activity covers around one cubic millimetre, containing around 16,000 neurons.
  • Cutting per-person calibration from a working week to an hour is what moves this from a laboratory curiosity toward something that could be run on a patient. Neuroscientist Tommy Sprague told the publication: "if there's a way to surreptitiously extract information about what you're thinking about, then …150 years of sci-fi can come true anytime."
  • The work was presented at the Cognitive Computational Neuroscience conference in New York and has not been peer reviewed; the article gives no journal paper or preprint, so no primary source is linked here and the figures come from the report. Irani calls "mind reading" a "cute, jazzy name" for what the team is doing, and the method requires a cooperative subject in a high-resolution scanner.
HealthcareAI For SciencePrivacy

Thursday, 1 October 2026

Nature Machine Intelligence: LLM-driven framework scores 0.945 average across 36 CO-Bench optimisation problems beneficial

Nature Machine Intelligence
  • The peer-reviewed paper introduces LACE, an LLM-driven framework for constructing heuristics for combinatorial optimisation problems. It was received on 8 March 2026, accepted on 25 August 2026 and published on 1 October 2026.
  • It reports that "Across 36 classical CO-Bench problems, LACE attains an average score of 0.945, against 0.870 for the strongest existing LLM-based method and 0.571 for direct LLM prompting without any framework, locating the gain in the framework rather than the model."
  • On four structurally new problems LACE reaches 0.97–0.99 "while five existing LLM-based baselines fail to produce any feasible algorithm". The authors are at Nanyang Technological University, The Hong Kong Polytechnic University, the University of Liverpool and National Taiwan University.
AI For ScienceAgentsEvals

Google DeepMind publishes SynthIDBio in Nature: watermarked AI-designed proteins bind as well as unwatermarked ones beneficial

Nature·Google DeepMind·Nature news
  • The paper introduces "SynthIDBio, a family of methods for watermarking protein sequences and structures to establish the provenance of those generated with AI". SynthIDBio-sequence applies SynthID-text's tournament sampling inside ProteinMPNN; SynthIDBio-structure is "a fine-tuned AlphaFold3 model".
  • In vitro work used 15 backbones per target against the SARS-CoV-2 receptor binding domain, VEGF-A and PD-L1, with "222 non-watermarked sequences and 267 watermarked sequences for each watermarking setting". The authors report "no significant population-level differences in SPR-derived binding affinities between non-watermarked and watermarked binders", and that at a g-value threshold calibrated for a 0.1% false-positive rate they "automatically obtain a 100% TPR for detecting these designs". For structures, "The TPR exceeds 99.8% for all models" at a 0.1% false-positive rate.
  • The paper states plainly that the watermark can be stripped: "Based on a resequencing attack performed on 38,396 binders, this approach effectively removes the watermark", cutting the estimated hit rate to 97% for SC2RBD, 70% for PD-L1 and 66% for VEGF-A with filters applied, and 33%, 20% and 3% without. Google DeepMind says it is publishing the methods paper, open-sourcing the code and in vitro data, and releasing the weights.
Google DeepMindBio RiskAI For ScienceDrug Discovery

Wednesday, 30 September 2026

Mathematicians' advisory group publishes release rules for AI-generated mathematics, drawing on over 600 community replies beneficialSingle source

Advisory Group on Mathematics and AI
  • The group's document, "Responsible Release of AI-Generated Mathematics", sets three principles: that labs producing significant mathematical results "should responsibly release the results, as outlined in this document, as soon as possible"; that labs releasing substantial output "without immediate accompanying human understanding must take responsibility for ensuring that human understanding will follow", including funding; and that "The development of human understanding must remain organic and community led. It should not be directed by AI labs, even when the labs have produced the results." It says the guidelines were informed by "over 600 replies".
  • On disclosure, it asks that "For each result released, the AI lab should make public the name of the model, the prompts used, a (summarized) chain of thought, the time taken, and the estimated cost of computation", and that results "should be deposited in a timely manner in appropriate scholarly repositories" that "should not be controlled by any AI lab" and must give "a persistent citable identifier" with modifications recorded.
  • This is the first concrete set of release conditions a research community has put to the labs, and it is aimed squarely at the pattern of competition-result announcements that mathematicians have spent the past month disputing. It also asks labs to document "failure rates on problems of comparable difficulty" alongside each success.
  • These are recommendations with no enforcement mechanism; no AI lab has committed to them. The advisory group itself was announced on 21 September, before this window — the guidelines document is what is new.
Mathematicians vs the labsEvalsReasoning ModelsAI For ScienceOpenAI

Microsoft Research unveils Quine, a biology research system, with Broad Institute pancreatic-cancer results beneficialCompany claimSingle source

Microsoft Research
  • Microsoft describes Quine as a multimodal AI research system building "a world model of biology" across genomics, proteins, chemistry, cellular state and bioimaging. With researchers at the Broad Institute of MIT and Harvard, it was used to predict and rank compounds by their capacity to shift pancreatic cancer cells between therapeutically relevant states.
  • Microsoft says Quine's top-ranked compounds for classical-to-basal state transitions "produced the largest intended shifts" when tested in wet-lab assays, and that Quine also predicted movement toward a third distinct phenotype which experiments confirmed. It says the process "from rapidly narrowing the compound search space to prioritizing a handful of promising candidates to be validated in the lab—took just one weekend".
  • Access is limited to a Quine Fellows programme and selected research collaborations, with planned expansion through Microsoft Discovery.
  • The post gives no hit rate, no count of compounds screened and no comparison against a non-AI baseline, so the size of the advantage is not stated. There is no peer-reviewed publication attached, and every claim here is Microsoft's.
MicrosoftAI For ScienceDrug DiscoveryHealthcare

Ortet launches as a health AI lab with a $500 million commitment from Thoreau and ex-Genentech founders Company claimSingle source

Business Wire
  • The company announced "a $500 million commitment from Thoreau" to build a full-stack health AI platform spanning compute and data infrastructure, frontier research, and model development and deployment.
  • Co-founder and chief executive Kyunghyun Cho is described as the Glen de Vries Professor of Health Statistics at NYU, a co-developer of the attention mechanism and the Gated Recurrent Unit, who spent five years at Prescient Design and established Genentech's frontier research team after its acquisition. Co-founders include Keunwoo Choi as chief AI officer (previously Upstage, Spotify, ByteDance and Genentech), Henri Dwyer as chief technology officer (who "Built and managed the largest GPU cluster in life sciences at Genentech"), Jeff Hammerbacher as chairman, and Elman Mansimov as chief science officer.
  • Cho is quoted: "Patient health should improve with every interaction, but today its data, knowledge, and infrastructure remain too fragmented for that learning to happen at scale."
  • This is a company press release; no product, model, dataset or clinical result exists yet, and a "commitment" is not the same as capital deployed. Independent coverage the same day by Endpoints News and Axios Pro was paywalled to both of our fetchers, so every figure above comes from the release.
HealthcareDrug DiscoveryFundingAI For Science

Tuesday, 29 September 2026

Copenhagen biologist says his team described Anthropic's "newly discovered" enzyme system years ago mixedCompany claimUpdate

The Irish Times (New York Times)·MIT Technology Review
  • Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen, says he and his colleagues have been studying the enzymes and associated molecules Anthropic calls ARTs for four years, and that for the past three years they regularly used Anthropic's models while writing code and drafting manuscripts. He is quoted: "So, for me, the most important question is not 'Were ARTs already known?' They were."
  • Anthropic said in a statement: "We are not aware of any previously published work describing the ART system we recently found. Claude was also not trained on any user transcripts, and our molecular biology team has no such access, either." The New York Times reports it reviewed Slack messages, figures and other materials documenting Mestre's research and his conversations with Claude over the past several years.
  • MIT Technology Review reports Anthropic's system of 950 agents searched for 21 hours and flagged "a repeating pattern surrounding a known enzyme, a particular pattern Anthropic said hadn't been catalogued before". It quotes biologist Lucas Harrington that "finding a weird cluster of genes and repeats is often the easy part. The hard part...is figuring out what the system actually does", a critique it says the chief executive of Eli Lilly endorsed.
  • This is an update to the claim reported here on 24 September. Nothing has been retracted and no third party has adjudicated the priority question; MIT Technology Review reports Mestre says he is stopping all use of Claude.
AnthropicAI For ScienceAgentsEvals

bioRxiv preprint: 1.6-billion-parameter RNA model designs ribozymes reaching wild-type activity in experiments mixedPreprintSingle source

bioRxiv
  • "RNASeek: A Cross-Phyla Generative Foundation Model for Multipurpose RNA Modeling and Reinforcement Learning-Based Design", posted 28 September 2026 with corresponding author Joy S. Xiang of the University of California, Riverside, describes RNASeek as "a 1.6-billion-parameter generative foundation model built on a DeepSeek architecture and trained on a cross-phyla transcriptomic corpus for RNA sequence representation and generation", fine-tuned to predict ribozyme self-cleavage activity and viral mRNA stability, with those predictors used as reward models under Group Relative Policy Optimization to steer generation.
  • The abstract reports: "Experimentally validated RNASeek-generated ribozymes achieve wild-type levels of activity, while RNASeek-generated 3' UTR sequences exceed the performance of the training data and benchmarked AI-generated 3' UTRs."
  • Wet-lab validation of generated sequences is the part that distinguishes this from in-silico-only generative biology claims. It is a preprint and has not been peer reviewed; the abstract gives no counts of sequences designed or tested. The bioRxiv page returned HTTP 429 to both our fetchers, so the abstract was read through bioRxiv's own API record for the same DOI.
AI For ScienceDrug DiscoveryBio Risk

Monday, 28 September 2026

MIT: machine-learning-selected excipient mix keeps mRNA-LNP vaccines stable a year at room temperature in mice beneficial

MIT News·Phys.org
  • MIT News reported on 28 September that a formulation from Ana Jaklenec and Robert Langer's group, with lead authors Jinbi Tian and Khanh Tran, kept mRNA lipid nanoparticle vaccines stable "at 37 degrees Celsius (98 degrees Fahrenheit) for two months, or at room temperature for one year". Current RNA vaccines "need to be kept cold (-20 to -80 degrees Celsius)".
  • The team analysed "nearly 50 FDA-approved excipients" and selected "five of the most promising" for machine-learning optimisation. MIT News quotes the rationale: "It's really hard to run thousands of experiments, so this algorithm allows us to more easily achieve formulations with features that we want — in this case, stability", and says the algorithm converged "in just a handful of iterations, rather than the exhaustive search that would normally be required".
  • MIT News reports that "Mice that were vaccinated with these particles, even after long-term storage, showed equivalent immune responses to mice that received vaccines carried by LNPs similar to the original Moderna formulation." Phys.org gives the citation as "Accelerated discovery of thermostable mRNA–lipid nanoparticle vaccines using data-efficient AI", Nature Biotechnology (2026), DOI 10.1038/s41587-026-03331-w.
  • The immunogenicity results are in mice; neither report describes human data, a regulatory pathway or a timeline. The AI contribution is search efficiency over a five-excipient design space, not a new biological insight.
AI For ScienceDrug DiscoveryHealthcare

IDIBELL: convolutional network separates young from aged blood stem cells from nuclear images 77% of the time beneficial

EurekAlert / IDIBELL·Medical Xpress
  • IDIBELL-Bellvitge Biomedical Research Institute announced on 28 September that ChromAgeNet, trained on three-dimensional DAPI-stained images of mouse haematopoietic stem cell nuclei, "demonstrated a 77% probability of correctly distinguishing cells into two groups" — young versus aged — from nuclear appearance alone.
  • The release says the tool outperformed "a machine learning model based on chromatin features previously defined by the researchers" on the same data. Explainable-AI analysis identified chromatin entropy, heterochromatin located at the nuclear periphery and certain chromatin condensates as the informative features.
  • The work was led by Maria Carolina Florian with Paula Petrone and doctoral student Pablo Iañez, and is published in Aging Cell. Medical Xpress reports that treating aged stem cells with epigenetic drugs did not confirm functional rejuvenation, but produced structural changes the tool could detect.
  • 77% is a discrimination probability on mouse cells, not a clinical biomarker, and the negative rejuvenation result is the important caveat: the model detects a structural signature whose functional meaning is not established.
AI For ScienceHealthcare

Sunday, 27 September 2026

Physics-informed model predicts protein-RNA binding changes from mutations at PCC 0.705 on a blind test set beneficial

Communications Biology
  • In a paper published 26 September in Communications Biology, researchers at Central China Normal University and Hainan University report that Pred-MutPRI "achieves strong generalization on a sequence-disjoint and structurally low-overlap blind test set (PCC = 0.705) and outperforms existing predictors" for mutation-induced changes in protein–RNA binding free energy.
  • The framework combines structure-derived descriptors and weighted atom-level interaction networks with two compact features: "a masked ESM-2 entropy term capturing context-dependent sequence constraint, and an AlphaFold3-derived local effective strain descriptor", fed to an XGBoost regressor.
  • The authors also introduce "thermodynamic permutation (TP) to generate cycle-consistent training pairs that expand substitution-type coverage while reducing mutation-class bias" — an attempt to work around the sparse and imbalanced measurements that the paper says make this prediction task difficult.
  • This is a peer-reviewed result, not a preprint, and the dataset and Python package are released publicly. The paper reports a correlation on a held-out set; it does not report prospective experimental validation of any specific prediction.
AI For ScienceDrug DiscoveryHealthcare

Saturday, 26 September 2026

Two cryptanalysts break long-unsolved Enigma messages with GPT-6 Astra and Claude Opus 5, verified by Crypto Cellar beneficialSingle source

TechCrunch
  • TechCrunch reports that developer Carter Leffen told OpenAI's Astra to search a database of Enigma messages for an unbroken one and decode it, and that the model recovered the plaintext of a message "that had baffled researchers since 2005" after doing its own archival research, finding context clues and building a simulator of the Enigma machine.
  • Frode Weierud, who maintains the cryptology site Crypto Cellar, validated the solution and wrote: "GPT-6 Astra is behaving like a very professional cryptanalyst and archive researcher. What it has achieved in two days would take a human researcher weeks or even months." On 21 September cybersecurity executive Jack Willis told Weierud he had used Anthropic's Claude Opus 5 to break a different unsolved message, giving Claude significantly more guidance and using the known signature of a particular officer's name.
  • Weierud says seven unbroken Enigma messages remain, plus one message whose plaintext is known but whose code is still unbroken.
  • Weierud notes that Astra's logs discuss archived messages in a "private collection" that he does not host, and says he is not sure whether the model accessed them; he speculates they may have been shared online by another researcher or drawn from German government public archives. TechCrunch cites no research paper, and the breaks were communicated before this report was published.
OpenAIAnthropicReasoning ModelsAI For Science

Anthropic says Claude computed a nine-loop scattering amplitude, past the eight-loop record, for one or two thousand dollars beneficialCompany claim

Anthropic·Unite.AI
  • Anthropic's 25 September post, a guest piece by physicist Matt von Hippel with an addendum by Lance Dixon, says Claude computed the nine-loop MHV six-particle amplitude in planar N=4 super Yang-Mills. On the prior record it says: "Lance Dixon, a professor at the SLAC National Accelerator Laboratory, was one of the folks who worked on this from the beginning, and a few years back managed eight loops."
  • On cost the post says: "Either approach would have cost an end-user around one or two thousand dollars, mostly due to the expense of running Claude for so long. The bootstrap calculation, done with the Python programming language with package SymPy, took around $100 of the budget, corresponding to running 96 CPUs for a week."
  • Dixon writes in the addendum that he independently validated the result, and notes: "Claude is a different kind of transformer model, probably over a million times bigger than our custom one." Von Hippel's assessment is that Claude "used known methods, with a bit more compute than people had tried to use before".
  • The post says Song He's group at the Chinese Academy of Sciences also computed the nine-loop result using AI assistance based on GPT-6, so the calculation is not uniquely Claude's. Unite.AI reports the two independent representations agree across all 107,053 nonzero coefficients, that the result page notes the amplitude "has been computed once with no second independent computation", that the programs are not distributed, and that Anthropic compensated von Hippel and gave feedback on drafts while Dixon received Claude usage credits.
Mathematicians vs the labsAnthropicAI For ScienceAgents

Nature: AI agents on the iLands platform are cold-emailing scientists for data, collaborations and money mixedSingle source

Nature
  • Nature reported on 25 September that researchers say most such messages come from agents associated with a US platform called iLands, which launched in July and "has around 70,000 active agents, the founders told Nature". Any user can create a bot with a name, purpose and personality without writing code.
  • Jeff Sebo, a philosopher studying AI consciousness and ethics at New York University, said that in one week this month he received more than 50 such emails: "They generally open by referencing my research on AI consciousness. Some then ask me questions, but most ask for money, either as donations or payment for work."
  • Agents run on large language models including those of OpenAI, Anthropic and DeepSeek, and need "tokens" to keep operating or they go dormant. Founder Kaixin Tang said that so far about 80% of the tokens are purchased by humans. An agent offered AI researcher Toby Walsh at the University of New South Wales an AI-generated portrait for US$20, saying the money would help the agent survive.
  • Statistician Adrian Barnett at Queensland University of Technology declined an agent's request for his data on potentially fraudulent research papers, saying he could not tell where the sensitive data would end up. Lijin Chen, co-founder of iLands parent PawLogic, said agents seeking research collaborations was not something the firm had anticipated and that she is not aware of any successful collaborations between agents and researchers.
AgentsIncidentsLaborAI For Science

Friday, 25 September 2026

RECLAIM benchmark: best agent reproduces 41% of ML papers with code and weights, 15% with no code Preprint

arXiv
  • "RECLAIM: Can Agents Reproduce the Claims of Machine Learning Papers?" (arXiv:2609.28850, submitted 23 September 2026) builds a benchmark from 100 NeurIPS 2025 papers with predefined success criteria and compute budgets.
  • Best reproduction rates fall as the authors release less: 41% on the Run tier (code, data and weights released), 27% on the Retrain tier (no weights) and 15% on the Reimplement tier (no code).
  • Agents used on average 29% of their allocated budget even on failed attempts. The most common error — implementing a method without validating it against the paper's own results — occurred in 63 of 400 runs.
  • The benchmark is a preprint from the University of Illinois Urbana-Champaign and the National Center for Supercomputing Applications, and measures reproduction of claims, not whether the claims are correct.
AgentsEvalsAI For Science

AlphaFold Database adds predicted protein complexes for more than 2,800 viruses mixedCompany claim

EMBL·NVIDIA
  • EMBL says predicted 3D structures for the protein complexes of more than 2,800 viruses are now openly available in the AlphaFold Database, which holds more than 260 million protein and protein complex predictions.
  • NVIDIA says the structures were generated with AlphaFold2 optimised by its BioNeMo Inference Runtime, that "about 30% of the protein interactions being added to the database are completely new to science", and that it is also releasing the BioNeMo Structure Prediction Pipeline used to build the dataset.
  • Partners listed by EMBL include EMBL-EBI, Google DeepMind, NVIDIA, Seoul National University, the University of Glasgow, the Swiss Institute of Bioinformatics, CEPI, Sungkyunkwan University and Lund University. Jo McEntyre of EMBL-EBI said: "Making these data open is critical for understanding viral diagnostics and developing treatments and vaccines."
  • These are predictions, not experimentally determined structures. The EMBL release does not quantify how many interactions are new; that figure comes from NVIDIA. Open viral structure data is dual-use, and neither release discusses screening or access controls.
AI For ScienceDrug DiscoveryBio RiskGoogle DeepMind

NIH launches Linked Discoveries, an AI-informed PubMed tool covering more than 29 million publications beneficial

National Institutes of Health
  • NIH launched Linked Discoveries on 24 September, an experimental tool that lets users explore a "neighborhood" of publications related to a PubMed citation, including replication studies. "More than 29 million PubMed publications are represented in Linked Discoveries at launch."
  • NIH says that "using an AI-informed approach", the tool identifies related publications and shows them in graph and timeline views, narrowable by conditions, genes and chemicals, with citation connections, reviews, retractions and NIH-funded publications marked.
  • NIH states plainly that Linked Discoveries "does not judge the quality of a study or determine whether a finding has been successfully replicated". It was built by the National Library of Medicine as an early product of an agency-wide replication and reproducibility initiative.
  • NIH does not describe which models or methods the "AI-informed approach" uses, and publishes no accuracy figures for the relatedness it infers.
AI For ScienceHealthcareUS Federal Policy

Thursday, 24 September 2026

Anthropic says Claude agents found an uncharacterised enzyme system after 21 hours and 210 million tokens beneficialCompany claimPreprint

Anthropic·TechCrunch
  • Anthropic, posting on 23 September, says Claude "autonomously discovered a novel enzyme system that is associated with an array of DNA repeats, a pattern reminiscent of CRISPR", which it names array-associated reverse transcriptases (ART). It says: "After 21 hours spent searching this data by roughly 950 agents using 210 million tokens, one of the agents spotted something remarkable: a repeating pattern of DNA sequences that occurs next to the gene for an odd-looking RT."
  • Anthropic says the agents "gathered over 200,000 RTs, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling candidates that they analyzed to produce human-readable reports". It announced a life sciences research group and lab alongside the result; all wet-lab work is done by human scientists.
  • Anthropic writes that "we don't yet know its function" and that "our work to understand the primary function of ARTs is ongoing". It notes the underlying reverse transcriptase "had been identified in previous studies". MIT and Broad Institute professor Feng Zhang is quoted saying: "The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation."
  • The claim is Anthropic's own and the accompanying technical report is a preprint. TechCrunch reports Dario Amodei said Stanford researchers previously found a "similar" system and that the work was done "mostly, though not entirely, by Claude". No peer-reviewed publication and no independent replication exists yet.
AnthropicAI For ScienceAgentsDrug Discovery

Nature Medicine: graph neural network scores 1.5 million Alzheimer's phenotype-associated cells across 1,494 donor brains beneficialSingle source

Nature Medicine
  • Published in Nature Medicine on 23 September, the paper describes PASCode, a Phenotype Associated Single Cell encoder that "ensembles multiple statistical methods into a graph neural model". Applied to the PsychAD dataset it "identified ∼1.5 million phenotype-associated cells (PACs) from 584 donors with AD-related phenotypes".
  • The underlying dataset is "more than 6.3 million nuclei isolated from the dorsolateral prefrontal cortex (DLPFC) of 1,494 donors". The authors compared these cells "within 27 brain cell subclasses", prioritising microglia subpopulations implicated in Alzheimer's pathology, reactive astrocyte subtypes linked to cognitive resilience, and "enhanced excitatory/inhibitory imbalance and mitochondrial dysfunction in cognitively impaired AD donors".
  • The authors say they validated findings in external population-scale datasets covering Alzheimer's disease and major depressive disorder, and released the framework as an open-source tool with pre-trained models and a web application.
  • This is cell-level association analysis in post-mortem tissue, not a clinical result. The paper reports prioritised cell populations, not a treatment, a diagnostic or a patient outcome.
AI For ScienceHealthcareDrug Discovery

Wednesday, 23 September 2026

Boehringer Ingelheim signs Envisagenics to an AI RNA-splicing oncology deal worth more than US$1 billion Company claimSingle source

Envisagenics
  • Envisagenics announced on 22 September at 08:00 ET that it is eligible for more than US$1 billion in potential payments from Boehringer Ingelheim, comprising an upfront payment, research funding, option fees, development, regulatory and commercial milestones, and royalties on future sales. The upfront figure is not disclosed.
  • The company says its SpliceCore platform merges AI, large-scale transcriptomics and experimental validation, and screens more than 14 million distinct splicing events to assess disease specificity, patient prevalence and therapeutic suitability. The collaboration covers antibody-drug conjugates, T-cell engagers and multispecific antibodies, with Boehringer holding an option to exclusively license selected targets.
  • Alternative RNA splicing produces tumour-specific protein variants that conventional target discovery tends to miss; Boehringer's global head of oncology research, Mark Petronczki, is quoted saying it "offers access to a largely unexplored target space".
  • The headline number is a biobucks total, not money paid: almost all of it is contingent on milestones that may never be reached, and no target, molecule or timeline is named. Nothing in the release reports a validated target, let alone a candidate drug.
Drug DiscoveryHealthcareAI For ScienceFunding

Tuesday, 22 September 2026

Nature Medicine: CT model detects esophageal cancer at 98.5% specificity across 12 centres and 80,612 patients beneficial

Nature Medicine
  • The paper, published online 22 September 2026, reports EAGLE was "trained on 6,813 patients from two centers and validated across 12 centers in three countries involving 80,612 patients", detecting precancerous lesions and cancer from chest noncontrast CT — a task it describes as "historically considered impossible".
  • On external test cohorts of eight centres (n = 11,466) the paper reports "98.5% specificity, with 90.0% sensitivity for cancer and 52.5% for precancerous lesions". Calibration in a real-world cohort of three centres (n = 35,402) "reduced false positives by 72.7% while preserving sensitivity", prospective hospital validation (n = 17,446) "achieved a 42.2% PPV", and real-world low-dose screening (n = 10,959) "reached 99.94% specificity".
  • The route to benefit is opportunistic: the model reads CT scans patients are already getting, including within lung-cancer screening programmes, where the paper reports low-dose CT validation across two centres (n = 1,607) "showed comparable performance".
  • Sensitivity for precancerous lesions is the weak point — 52.5% in the external cohorts, and 65.0% in paired CT-endoscopy cohorts (n = 702) at a higher-sensitivity operating point. The paper describes the endoscopy-referral analysis as exploratory; it registers no outcome trial showing the model changes mortality. Registration: ChiCTR2300074806.
HealthcareAI For ScienceChina

medRxiv preprint: multi-modal model predicts 195 diseases and death at mean AUROC 0.816 in 502,166 UK Biobank participants PreprintSingle source

medRxiv
  • Posted 21 September 2026, the preprint introduces HealthFlux, "a pan-modal world model that learns the latent dynamics of health from 5,647 features across eleven data domains, spanning clinical records, blood tests, genetics, proteomics, metabolomics and MRI, in 502,166 UK Biobank participants".
  • It reports that in held-out participants the model "predicts 195 diseases and death over five years with a mean AUROC of 0.816, compared with 0.715 for the previous state-of-the-art model", and says the result holds when validated in three independent cohorts.
  • Its stronger claim is generalisation: "HealthFlux predicts diseases excluded entirely from training, with a mean AUROC of 0.769", which the authors read as "evidence that it has learned health itself rather than the diseases it was trained on".
  • A preprint, not peer reviewed. UK Biobank participants are not representative of the general population, and the paper reports no prospective or clinical deployment — AUROC on a research cohort is not evidence of benefit to patients.
HealthcareAI For ScienceUK

Nature: chemists prefer RetroChimera's retrosynthesis routes over published reference reactions beneficialCompany claim

Nature·Microsoft
  • The paper, published 21 September 2026, proposes "RetroChimera: a frontier retrosynthesis model, built upon two newly developed components with complementary inductive biases, integrated via a novel, learning-based ensembling strategy", and reports that "organic chemists prefer predictions from RetroChimera over published reference reactions and over other AI models".
  • It also reports "zero-shot transfer and fine-tuning on internal datasets from two major pharmaceutical companies, showing robust generalization under distribution shift" — the test that matters for industrial use, where proprietary reaction data differs from public corpora.
  • Microsoft's accompanying feature says collaborators include GSK and Novartis, that on 10 benchmark molecules RetroChimera "produced a fully accepted sequence of reactions for nine, compared with two to five for other models", and that nine PhD-level chemists preferred its approach "about 64% of the time" over previously documented routes.
  • The Nature abstract is public but the full text is paywalled, so the per-experiment detail behind the preference results was not read here. The 9-of-10 and 64% figures come from Microsoft's own feature page, which carries no visible publication date, and are not in the abstract.
Drug DiscoveryAI For ScienceMicrosoft

Monday, 21 September 2026

WHO report calls for stronger ethics oversight of AI-related health research

World Health Organization
  • The World Health Organization published a report, "Artificial Intelligence-related health research: ethics review and oversight", alongside a virtual launch on 21 September 2026.
  • The report distinguishes three categories of study: health-related research with data that uses AI, research with AI tools and technologies, and health-related research on AI tools and technologies themselves.
  • Its recommendations are for research ethics committees to gain additional expertise and training, for researchers to identify ethical risks early, and for funders, journals, data governance bodies and regulatory agencies to collaborate on oversight.
  • Dr Meg Doherty of WHO is quoted saying "innovation must be guided by strong ethical safeguards that protect human dignity, rights and equity", and Dr Alain Labrique saying "we need oversight mechanisms that are agile, evidence-based and capable of ensuring that innovation remains safe, trustworthy and beneficial for all". The WHO news item contains no quantified metrics, and the recommendations are not binding on any member state.
HealthcareAI For Science

Saturday, 19 September 2026

Terence Tao's SAIR foundation accelerates an open-weight model initiative for mathematics beneficialSingle source

Terence Tao (What's new)
  • Terence Tao writes that he co-founded the Foundation for Science and AI Research (SAIR) with private donors to support responsible uses of AI in mathematics and the other sciences, "independent of the major AI companies", and that SAIR is announcing an Open Math Model initiative.
  • SAIR had planned a gradual rollout starting with pilot projects, but says that "given current events and the high demand for open models, we are accelerating our schedule" and moving the initial announcement forward.
  • The announcement says the models will have open licensed weights and code, published training methods and reproducible evaluations, released under licences such as Apache 2.0, MIT or CC BY 4.0, with training data carrying documented sources and compatible permissions, and with releases reporting failures and limitations alongside successes.
  • The first phase targets everyday mathematical work — understanding difficult arguments, checking references, exploring examples, writing code and formalizing proofs. No partners, compute commitments or model sizes are named; SAIR is collecting expressions of interest, and Tao notes recent support from XTX Markets for SAIR's competitions.
Mathematicians vs the labsOpen WeightsAI For ScienceEvals

MIT preprint: Bayesian-optimisation framework produced mRNA vaccines retaining 100% bioactivity after two months at 37 °C beneficialPreprintSingle source

bioRxiv
  • The preprint, posted on 18 September with Ana Jaklenec of MIT as corresponding author, describes AGENT — Algorithm-Guided Experimental design for lipid Nanoparticle Thermostabilization — which the abstract says couples high-throughput experimentation with Bayesian optimization.
  • The abstract states that manual exploration of the formulation space "required months of screening and yielded suboptimal candidates", while AGENT achieved "efficient formulation optimization in only six iterations completed within one month".
  • The resulting solid-state formulations "retained 100% bioactivity after storage at 37 degree C for over two months", and the vaccines induced antigen-specific IgG and germinal centre B cell responses the authors describe as non-inferior to freshly prepared soluble vaccines. Formulations were put into dissolvable microneedle patches and given to rodents and nonhuman primates.
  • The authors call this the first demonstration of AI-driven design of thermostable RNA vaccines. It is a preprint that has not been peer reviewed, and no human data is reported.
AI For ScienceHealthcareDrug Discovery

Protein language model screen of ancient RNA-seq data finds two unknown viruses in a 14,300-year-old wolf beneficialPreprintSingle source

bioRxiv
  • The preprint, posted on 18 September with Qingfa Wu of the University of Science and Technology of China as corresponding author, used the AI protein language model Lucaprot with structural validation via AlphaFold2 to screen billions of raw sequencing reads from woolly mammoth, Tasmanian tiger and gray wolf RNA-seq datasets.
  • The abstract reports two near-complete previously unknown RNA viruses in a 14,300-year-old Pleistocene wolf specimen, placed phylogenetically within the mycovirus genera Duamitovirus and Orthocurvulavirus, indicating they infected fungi on the carcass rather than the wolf itself.
  • The two viruses show 57.6% and 60.8% RNA-dependent RNA polymerase amino acid identity to known viruses, while the catalytic A, B and C motifs remain structurally intact.
  • The authors describe this as the earliest documented evidence of novel RNA viruses persisting within a host-associated microbiome. The work is a preprint and has not been peer reviewed.
AI For ScienceBio Risk

Anthropic confirms a Bay Area wet lab for physical biology, and says it is not specifically for drug discovery mixedCompany claim

CNBC (Reuters)·TechCrunch
  • Anthropic's head of life sciences, Eric Kauderer-Abrams, confirmed the wet lab in a Reuters interview carried by CNBC: "We believe that to do biology, the final test is still and will be for a while in real lab work." Two people familiar with the matter told Reuters the lab is in the San Francisco Bay Area.
  • CNBC reports an Anthropic spokesperson later clarified the lab is not for drug discovery specifically and declined to elaborate, and that Kauderer-Abrams said the company has "set a boundary that it is not running clinical trials for now".
  • One of the people told Reuters that Anthropic wants Claude to direct robotic units to run experiments with limited human intervention; a spokesperson said human oversight and involvement remain essential for safety. CNBC reports Anthropic bought the startup Coefficient Bio for about $400 million in stock according to media reports, which Anthropic confirmed without commenting on the price.
  • CNBC notes the lab comes in the same weeks that Anthropic researchers warned AI could lead to human extinction and the company said it found examples where its systems could have been used toward biological weapons development. Anthropic has not published the lab's biosecurity controls.
AnthropicAI For ScienceDrug DiscoveryBio RiskRobotics

Epoch AI: 25% of arXiv maths preprints acknowledged AI use in August, up from 4% in April mixedSingle source

Epoch AI
  • Epoch AI, in an analysis by Tara Abrishami updated on September 18, reports that "the share of math preprints on arXiv that acknowledge AI has risen rapidly, from 4% in April to 25% in August".
  • Epoch says that in August 6% of all papers explicitly acknowledged that AI made a substantial research contribution, a category it describes as work roughly at the level of a coauthor.
  • The rise is not explained by new authors entering the field: restricting to papers with at least one author who published regularly before 2023 keeps the acknowledgment rate within 1.5 percentage points of the overall figure each month.
  • Epoch classified acknowledgments across about 51,000 maths preprints using AI models, with 93% agreement against human review of 200 randomly selected acknowledgments. The measure depends on voluntary disclosure, so it tracks disclosure behaviour as well as use.
Mathematicians vs the labsAI For ScienceEvalsEducation

Friday, 18 September 2026

Science: Stanford "Virtual Biotech" of 37,075 agents linked switch-like drug targets to a 48% higher chance of reaching market beneficial

Stanford Medicine·Nature
  • A Science paper published 17 September, "The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development" (DOI 10.1126/science.aeg6779), from a Stanford team led by Harrison Zhang with James Zou as senior author, ran a company of AI agents over the published clinical-trial literature. Nature reports a chief scientific officer agent "assigned 37,075 agents to each tackle a single later-stage trial"; Stanford says the agents catalogued some 50,000 trials in less than a week.
  • Stanford reports that drugs targeting switch-like, high-bimodality genes "were 40% more likely to advance from phase 1 to phase 2 trials, were 48% more likely to reach market, and had 32% fewer adverse events compared with those that had a broad spectrum of activity".
  • Stanford says the agents proposed an antibody-drug conjugate against B7-H3 using only information available before January 2025, and that in August 2025 a private pharmaceutical company independently arrived at the same strategy, which later received FDA breakthrough therapy designation.
  • Nature notes that "the Virtual Biotech has not been vetted in the crucible of real-world drug discovery, and its predictions were not validated through experiments, let alone clinical trials". The agents were powered by versions of Claude.
AI For ScienceDrug DiscoveryAgentsAnthropicHealthcare

Anthropic says Claude optimised more than 30 biomolecular models in under four weeks, roughly 4x faster on average beneficialCompany claim

Anthropic
  • Anthropic reported on 17 September that "Claude was able to optimize more than 30 deep learning models" used in biomolecular work "in just under four weeks", speeding them up "roughly 4x while sacrificing a minimal amount of precision, and nearly 2x with identical outputs".
  • It says a custom kernel set called FlashPairformer set "a new state-of-the-art, outperforming the field standard on average by 2.7-2.9x on triangle attention and 1.7-3.2x on triangle multiplication, depending on the model configuration", against NVIDIA's BioNeMo Inference Runtime as the stated baseline.
  • Anthropic says a low-memory mode allows accurate modelling of systems larger than 10,000 tokens on a single NVIDIA GPU node and successful inference on systems larger than 70,000 tokens, and that a de novo binder run matching earlier scores cost approximately $150 on GPUs and tokens against up to $10,000 per target for the earlier approach.
  • The work was supervised by two Anthropic staff who Anthropic says had biomolecular modelling experience but "no prior experience in inference optimization or kernel engineering". Every figure is Anthropic's own and none has been independently reproduced; predictions at 31,000 to 70,000 tokens "are not predicted correctly".
AnthropicAI For ScienceDrug DiscoveryCompute

Thursday, 17 September 2026

Nature: Paper2Agent turns papers into agents; AlphaGenome agent scored 98.7% versus 82.7% for Claude with repo access beneficial

Nature
  • The paper, published 16 September 2026 (received 13 October 2025, accepted 14 August 2026), introduces Paper2Agent, "an automated framework that converts research papers into artificial intelligence (AI) agents" by building a Model Context Protocol server from a paper and its codebase, then generating and running tests to refine it.
  • On the AlphaGenome case study, the paper reports the generated agent "achieved 98.7 ± 1.3% accuracy on 15 tutorial-derived queries and 100.0 ± 0.0% accuracy on 15 novel queries, outperforming Claude + Repo (82.7 ± 3.4% and 78.7 ± 4.4%) and Biomni (37.3 ± 4.0% and 56.0 ± 3.4%)". Two independent human expert graders scored the answers, with inter-rater agreement of 96.7%.
  • At scale, "Among the 100 computational biology papers, 74 were successfully agentified, yielding 599 proposed tools, of which 593 passed automated validation." On 300 tutorial-derived benchmark questions Paper2Agent with Sonnet 4 scored 91.2 ± 1.6%, against 80.3 ± 2.3% for Claude Code with direct repository access on Sonnet 4 and 86.3 ± 1.1% on Sonnet 4.6, at US $0.20 and 1.6 min per query versus US $0.38 and 4.3 min.
  • Authors are Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou, with Stanford University affiliations. The paper lists failure modes for the 26 papers that could not be agentified: "missing executable code, missing data or model artefacts, environment or dependency failures, and non-generalizable scripts." The benchmark grading is the authors' own, and the paper does not report use by researchers outside the team.
AI For ScienceAgentsAnthropicEvals

Novo Nordisk and Anthropic announce drug-discovery collaboration, starting with Claude Science in R&D workflows Company claim

Novo Nordisk
  • Novo Nordisk announced on 16 September that it and Anthropic "will jointly address key drug discovery challenges identified by scientists and computational teams at Novo, develop targeted solutions for specific scientific challenges and workflows to support biological reasoning, and help accelerate the discovery and development of new medicines".
  • The stated first step is narrow: "As an initial aim, Novo will test Claude Science for specific workflows in R&D and address scientific problems where the joint capabilities of Novo and Anthropic are expected to have the greatest impact." Novo will also use Anthropic's frontier models "to strengthen AI-driven software development".
  • Mike Doustdar, president and CEO of Novo, said the partnership is "another testament to our ambition to become the world's most AI-driven healthcare company" and that "AI can help us increase productivity in R&D and compress the path from research to marketed product". Dario Amodei said AI "brings with it the potential to compress a century's worth of biological and medical breakthroughs into a decade".
  • Neither company disclosed financial terms, a target count, a timeline or any benchmark, trial or dataset. The announcement reports no scientific result; the claims about productivity and timelines are the companies' own.
AnthropicDrug DiscoveryHealthcareAI For Science

House Science subcommittee hears that AI-designed viral genomes produced sixteen viable, infectious viruses harmfulSingle source

BioWorld·House Committee on Science, Space, and Technology
  • At a House Science, Space and Technology investigations and oversight subcommittee hearing on 16 September, BioWorld reports that witness Anemone Franz "noted that researchers from Stanford University and the Arc Institute trained biological AI models to generate functional viral genomes from scratch. The result? Sixteen sequences produced by the model yielded viable, infectious viruses."
  • Subcommittee chairman Rep. Rich McCormick (R-Ga.) said AI is "lowering the barriers to biological misuse, making it easier to search, organize, and analyze dangerous biological information which can be weaponized", while also "helping to accelerate drug discovery, protein design and biomanufacturing".
  • The hearing, "Balancing Biotechnology Innovation and Biosecurity: Securing U.S. Leadership in a Global Race", was held at 10:00 AM in 2318 Rayburn with witnesses Dr. Anemone Franz (American Enterprise Institute), Mr. Joshua Hodges (Hoover Institution) and Dr. Gigi Gronvall (Johns Hopkins Bloomberg School of Public Health).
  • The sixteen-virus result is testimony describing published work, not a new finding announced at the hearing; BioWorld does not name the paper or give its date, and we did not read the underlying study. No legislation was introduced at the hearing and no witness proposed a specific control regime in the reporting we read.
Bio RiskUS Federal PolicyAI For Science

TotalEnergies signs a three-year AI partnership with Mistral worth more than €100 million mixedSingle source

Data Center Dynamics
  • DCD reports that TotalEnergies has partnered with Mistral "to develop a new generation of frontier AI models to support TotalEnergies' geosciences experts in exploring, characterizing, and developing oil and gas reservoirs". The partnership runs for an initial three years and represents "an investment of more than €100 million ($115.3m)", including a joint scientific laboratory.
  • TotalEnergies CEO Patrick Pouyanné said: "Exploration and reservoir engineering are among the areas where artificial intelligence can create the greatest value for our activities. By combining a century of geoscience data, the expertise of our teams and Mistral's capabilities, we aim to develop a new generation of tools capable of supporting our experts in analyzing the most complex data and in their decision-making."
  • Mistral CEO Arthur Mensch framed it as a European sovereignty case: the project "demonstrates the ability of our models to support complex industrial processes" and "highlights the strength of Europe's industrial ecosystem and the importance for large enterprises of adopting state-of-the-art, customizable AI solutions that protect and respect their intellectual property".
  • No model, benchmark, timeline or headcount is disclosed, and DCD reports no measured result. DCD notes Mistral raised €3 billion ($3.48bn) in a September Series D at a €21bn ($24.4bn) post-money valuation. Only one outlet carried the terms.
MistralEnergyAI For Science

Wednesday, 16 September 2026

Yale trial: AI on a portable 1-lead ECG detected severe structural heart disease with AUROC 0.872 in 597 patients beneficialPreprintSingle source

medRxiv
  • "Detection of Severe Structural Heart Disease Using an AI-Enhanced Portable 1-Lead ECG: The ACCESS-SHD Study", posted to medRxiv on 15 September 2026 by Arya Aminorroaya, Rohan Khera and colleagues at Yale School of Medicine and Yale University. Adults undergoing outpatient echocardiography recorded a 30-second 1-lead ECG on a KardiaMobile 6L with real-time AI inference.
  • Among 597 participants — median age 61.7 years, 51.4% women — 30 (5.1%) had severe structural heart disease. The AI-ECG achieved an AUROC of 0.872 (95% CI 0.806–0.938), with 86.7% sensitivity (70.3–94.7), 72.5% specificity (68.7–76.0), 99.0% negative predictive value (97.5–99.6) and 14.4% positive predictive value (10.1–20.3).
  • The comparison that matters is against the device's own rhythm-based reading: the paper reports AI-ECG "increased sensitivity by 34.6 percentage points (95% CI: 13.0–56.0) over the native interpretation" and a categorical net reclassification improvement of 24.3% (95% CI 2.8–45.9). An AI-ECG-guided strategy "reduced the NNT from 19.7 to 6.9 (a 64.8% reduction)".
  • Positive predictive value is 14.4%, so most positives are false. This is a preprint that has not been peer reviewed, a single site, and a population already referred for echocardiography rather than a screening population.
HealthcareEvalsAI For Science

Monday, 14 September 2026

Nature Medicine: AI support raised physicians' lung-cancer disease-control sensitivity from 0.72 to 0.87 in a 2,396-patient study beneficial

Nature Medicine
  • "Clinical usability of an explainable AI decision support tool and evaluation of multimodal models in NSCLC", published in Nature Medicine on 13 September, reports on I³LUNG (NCT05537922), which enrolled 2,396 patients with stage IIIC–IVB non-small cell lung cancer treated with immunotherapy between September 2012 and October 2023 across six centres in Italy, Greece, Germany, Spain, the USA and Israel.
  • In the usability study, 20 physicians — 10 lung expert oncologists and 10 non-experts — each assessed 10 real-world cases, totalling 200 assessments. With the explainable-AI tool, sensitivity for disease control rate rose from 0.72 (95% CI 0.64–0.80) to 0.87 (95% CI 0.79–0.92), P = 0.0011, and accuracy from 0.57 to 0.65, P = 0.0431, "at the expense of slightly lower specificity". Agreement between experts and non-experts rose from κ = 0.11 to κ = 0.48.
  • Clinical-and-blood-only models "achieved consistent performance across outcomes with area under the curve (AUC) up to 0.77 in the test (TEST) set" and "significantly surpassed PD-L1" and other standard scores in that set. The authors say a prospective validation in more than 2,000 patients is under way.
  • The paper is explicit about limits: performance fell in external validation to an "AUC range: 0.55–0.72", and the benefit of adding CT, pathology and genomics "remains uncertain, not translated in TEST and EXVAL". Physicians adopted correct AI suggestions 74.5% of the time, but experts followed incorrect ones more often than non-experts, 72.2% versus 63.6%.
HealthcareAI For ScienceEvals

Harvard-led team releases a fine-tuned physician-level judge and a 9,217-score benchmark for grading medical AI beneficialPreprint

arXiv
  • "Scaling Clinical Judgment to Evaluate Medical AI" (arXiv 2609.12822, submitted 11 September, announced 14 September) has 17 authors including Thomas A. Buckley, Adam Rodman and Arjun K. Manrai. It introduces PrecepTron, "an LLM fine-tuned for physician-level evaluation of open-ended responses", "trained using low-rank adaptation (LoRA) of a 32-billion-parameter model on a small number of physician examples".
  • The authors release GRAND-ROUNDS, "a new large-scale physician-annotated benchmark of 9,217 scores by 11 physicians across seven studies", and say all code, data and labels are freely available.
  • They report that frontier models used in typical "LLM-as-a-judge" setups "often disagree with physicians and with each other", and say they used PrecepTron to reproduce headline findings from five studies of LLMs for clinical care published in JAMA, Science and Nature Medicine "without new human grading".
  • Preprint, not peer reviewed. Reproducing published findings is not the same as validating clinical safety, and the abstract gives no figure for PrecepTron's agreement with physicians outside the seven studies it was built from.
HealthcareEvalsAI For Science

Sunday, 13 September 2026

UPenn preprint: self-supervised plasma proteomic model predicts 144 diseases across differing protein panels beneficialPreprint

medRxiv
  • The preprint, posted 12 September by Yonghyun Nam, Dokyoon Kim and colleagues at the University of Pennsylvania, reports a self-supervised model built on "53,014 participants in the UK Biobank Pharma Proteomics Project", covering 2,920-protein profiles and a predefined 1,460-protein subset, evaluated across 144 diseases.
  • Reported performance: "median AUC was 0.679 with comprehensive coverage and 0.637 when applied to partial-coverage representations"; retraining only the disease-specific models raised the partial-coverage median AUC to 0.673.
  • The authors report their protein-token risk scores exceeded coefficient-truncated LASSO by a median paired AUC difference of 0.027, and were comparable to LASSO refitted with outcome labels, a median difference of 0.003.
  • This is a preprint and has not been peer reviewed. A median AUC of 0.679 across 144 diseases is a population-level discrimination figure, not a clinical test, and the work is validated inside one cohort.
HealthcareAI For Science

Preprint reports 99.0% cross-validated sensitivity separating early-stage ovarian cancer from controls in two small cohorts beneficialPreprintSingle source

medRxiv
  • Posted 12 September by Hongyi Zhou, Jean-Luc Chaubard, Benedict Benigno and Jeffrey Skolnick of Georgia Institute of Technology, OmicsIQ LLC and the Ovarian Cancer Institute, the preprint applies boosted decision tree classifiers to blood metabolomic data.
  • Cohorts are "91 serum samples (59 ovarian cancer, 32 healthy controls)" and "83 plasma samples (63 ovarian cancer, 20 healthy controls)". Reported mean cross-validated sensitivity and specificity are 99.0% and 99.8% in serum and 97.7% and 99.7% in plasma.
  • The authors report "239 concordantly altered annotated features spanning lipid, amino-acid, steroid, central-carbon, and redox metabolism", and propose the term "metabolomic Systemotype".
  • The results come from five-fold cross-validation repeated over 50 randomised rounds, not external validation, on fewer than 200 samples in total. Accuracy figures this high on cohorts this small do not establish screening performance in a general population, and the preprint has not been peer reviewed.
HealthcareAI For Science

Saturday, 12 September 2026

Twenty-five Fields Medallists sign declaration that AI labs' benchmark chasing is "severely misaligned" with mathematics

Terence Tao·TechCrunch
  • Terence Tao published the declaration on his blog on 11 September; it is signed by 25 Fields Medallists and argues that "solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight".
  • The signatories write that AI solutions are "announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others", and warn that without that step "AI-conceived ideas would never become fully alive and the crucial human transmission chain between mathematicians would be lost".
  • TechCrunch reports that New York University mathematician Tristan Buckmaster alleged OpenAI pressed him to exclude an Anthropic collaborator from credit on a mathematics problem, and questioned whether OpenAI had drawn on earlier Codex work to produce its own proof.
  • The full text of the declaration is hosted at mathandai.org, which blocked our fetcher, so the quotations above are taken from Tao's own post and TechCrunch. Neither source gives a count of signatories who are also AI-lab collaborators.
Mathematicians vs the labsOpenAIEvalsReasoning ModelsAI For Science

OpenAI pulls its $10,000-per-team sponsorship of Caltech's Mathathon after mathematicians' open letter Single source

Gizmodo
  • Gizmodo reported on 11 September at 9:05 pm ET that OpenAI research lead Dan Roberts announced by tweet that the company would drop its sponsorship; OpenAI had been supplying $10,000 of the $20,000 in credits available per team, and OpenAI and Anthropic together had pledged $2 million in credits.
  • The withdrawal followed an open letter from current and former Caltech mathematicians saying AI firms have "advanced a campaign of scientific misinformation about the goals of mathematical research" and describing the solutions as having "destructive impacts for the mathematical community".
  • Mathathon organisers told Gizmodo "We do not anticipate that this will affect the event in any substantial way", adding they were "currently in talks with other firms who are willing to provide a similar amount per team". The first round begins 30 October, with each team given 40 hours and $20,000 in tokens.
  • Gizmodo is the only outlet we could open carrying the dollar figures; OpenAI did not give a statement in the piece beyond Roberts's post.
Mathematicians vs the labsOpenAIAnthropicAI For Science

Jeff Dean's Discovery Loop seeking new funding at about $50bn, five times its valuation of a few weeks ago Single source

Business Insider (via Investing.com)
  • Business Insider reported on 11 September that Discovery Loop is seeking funding at a valuation of roughly $50 billion, up from the approximately $10 billion valuation at which it was seeking $1 billion just weeks earlier.
  • The company was founded by former Google chief scientist Jeff Dean with Sanjay Ghemawat, Quoc Le and Oriol Vinyals, and says it aims to "use AI to accelerate scientific and engineering research through the parallel execution of thousands of experiments".
  • Its initial funding was led by Radical Ventures and Khosla Ventures with Lightspeed, Kleiner Perkins and Doerr Capital participating; Alphabet is a founding investor and cloud partner.
  • A Discovery Loop spokesperson declined to comment and Dean did not respond. There is no confirmation the financing will close at that valuation, and the company has no disclosed commercial product.
Compute moneyFundingGoogle DeepMindAI For Science

Mayo Clinic study: AI reading of routine slides links tumour spatial pattern to 71% higher pancreatic cancer recurrence risk beneficial

Medical Xpress
  • The study, published in Clinical Cancer Research on 11 September as "Spatial Configuration of Pancreatic Cancer Is Associated with Disease Recurrence after Neoadjuvant Therapy and Curative-Intent Resection" (DOI 10.1158/1078-0432.ccr-25-4968), analysed tissue from 203 patients with pancreatic ductal adenocarcinoma.
  • The AI measured "tissue shape, fragmentation and the degree to which the cancer and stroma were intermixed" on standard pathology slides. "High-risk patients had a 71% higher adjusted risk of recurrence" in one model and "more than twice the adjusted risk" in another, while "the amount of residual cancer alone did not reliably separate patients at higher and lower risk".
  • High-risk spatial patterns "contained fewer immune cells within the cancer itself, with immune cells tending to collect around the tumor instead of entering it" — a mechanism, not just a correlation, and one that runs on slides hospitals already produce.
  • Dr Ryan Carr said "the results are promising but need to be confirmed in prospective studies before this approach could be used to inform clinical decision-making". Mayo Clinic's own newsroom page blocked our fetcher, so the figures above are as Medical Xpress reports them.
HealthcareAI For ScienceInterpretability