AI and the supply of research effort

Microsoft's published MAI-DxO diagnostic comparisons use retrospective cases and omit specialist teams; a prospective consenting-patient trial at Beth Israel Deaconess has not yet reported results.

Microsoft's MAI-DxO solved 83% of complex simulated medical cases versus generalists' 36%, but prospective clinical validation remains ahead and published comparisons use retrospective cases.

Google offered up to 200 European researchers access to five science models, including AlphaFold 3 and AlphaGenome, with $3M in compute credits and company-selected admission.

The 142,250-person NHS-Galleri trial missed its primary endpoint of reducing stage III and IV cancer diagnoses, though the blood test detected 937 tumors and secondary analysis found fewer late-stage cancers.

A single liver-targeted base-editing infusion suppresses a cholesterol-raising protein over 90% for three years in monkeys; a first human trial is underway.

Physical Superintelligence's system running an unreleased Anthropic model produces an Aaronson-Ambainis conjecture proof on day three, later tightening its bound twice; the main theorem is Lean-checked.

Axiom transferred the Stacks Project's roughly 7,500 pages into a reusable 16M-line Lean library as AI-generated mathematics outpaces human refereeing capacity.

Anthropic reported Claude-assisted purification of a prokaryotic vault particle fell from about three days to 90 minutes, enabling its first 3D cryo-EM map; biological function remains unknown.

Nature published AutoDNA, a multi-user biomolecular laboratory coordinating over twenty instrument types across more than 11,000 steps, with overlapping experiments 3.2 times faster.

EMBL, DKFZ and Sanger researchers mapped glioblastoma cell-state regulation across over one million cells; increasing MYT1L reduced tumor growth and invasion and extended survival in mice.

Open structRFM, trained on 21 million RNA sequence-structure pairs, supported Zfold predictions comparable to AlphaFold3 across three benchmarks, with dataset, weights and code released.

FLIP improved stellar-glare removal sensitivity by two to fifteen times and recovered 51 Eridani b from JWST observations where standard techniques fail; computational costs constrain broader application.

Dr Reddy's generic twice-yearly lenacapavir received tentative FDA approval under 16 months after the original, targeting $40 annual HIV prevention across 120 countries.

A study of about 120,000 PLOS ONE papers finds individual research contributions became more specialized after ChatGPT's release, concentrated in newly formed teams.

Tencent reports Hyra 2.0 agents developed a 1.96M-parameter protein-design model generating sequences 57 times faster than ProteinMPNN, alongside quantum error-correction designs and mathematical results.

Hannover and Shanghai report base-editing seven alpha-1 antitrypsin deficiency patients; a published liver biopsy confirms correction in one recipient, off weekly replacement infusions since March.

Exin Therapeutics reported Claude Sonnet 5 selected among 19 gene therapies and controlled robotic delivery into a live mouse's brain, stopping when calibration failed.

Sam Altman said OpenAI spent hundreds of millions of dollars on mathematics, including about $10M on Navier-Stokes work, and holds further unreleased results while opposing slower publication.

A Tokyo-based multi-agent lab studying diamond spin qubits ruled its own passing result inconclusive after reference readings drifted, then re-ran the experiment under unchanged criteria.

UCSF's brain-activated engineered CAR-T cells cleared diffuse midline glioma and rejected repeat tumors in mice; a human trial is planned for April 2027.

Axiom Math published a complete Lean formalization of finite simple group classification, exposing counterexamples to a published closure claim and supplying a 35-page alternative argument.

Anthropic posted a rival claimed Exponential Time Hypothesis refutation as researchers challenged OpenAI's selection of reviewers for its withheld strong-force result.

LongChang tested 12 models in altered-physics simulations: GPT-6 Astra met discovery criteria in 16.7% of runs, while the other eleven did so in none.

OpenAI handed a claimed Exponential Time Hypothesis refutation and a Yang-Mills existence proof to the outside experts who proposed the problems for independent stewardship with no NDAs, as Resolve confirmed.

Roche's divarasib cut death risk 44% versus the approved drugs sotorasib and adagrasib in a 338-patient Phase 3 KRAS G12C lung cancer trial, extending median survival to 17.7 months from 12.0.

Interim Phase 3 DeLLphi-305 results in 563 patients found Imdelltra plus Imfinzi extended median small cell lung cancer survival to 21.1 months versus 12.7 months.

A single trapped ion corrected errors beyond break-even without measurement or feedback, while another team demonstrated high-fidelity two-qubit gates in crystals of up to 12 ions.

Intismeran autogene plus Keytruda cut melanoma recurrence or death 42% in a Phase 3 trial exceeding 1,100 patients; two-year recurrence-free survival reached 82.2% versus 70%.

Pinal generated proteins from written descriptions of four functions, with laboratory tests confirming all four worked.

Contract laboratories found 17 scorpion-toxin binders among 211 tested Claude-assisted protein designs; Anthropic released the tested designs and experimental data.

Radical Numerics reported Omnii-1-designed phages administered to a patient with an antibiotic-resistant infection through authorized compassionate use; clinical outcomes remain undisclosed.

UW and Nvidia released BSD-3-licensed RFD4-Proteina, reporting five to ten times faster candidate generation and computational design success, with laboratory validation across several tasks.

Quantinuum reported 99.9995% logical memory fidelity on Helios; Nvidia simulated a 104-qubit fault-tolerant processor using 8,192 GPUs on Europe's JUPITER supercomputer.

MIT trained a model to read crystal structures and plain-language descriptions together, letting researchers request solid-state lithium-battery electrolytes and surface candidates other labs had identified.

Anthropic and Harvard reported Claude redesigned eight quantum-computing building blocks, cutting one decoding routine's costly operations roughly twelvefold and another routine's worst-case operations substantially.

A UK digital-first GLP-1 observational program enrolled 8,127 women; 7,297 with six-month measurements averaged 15.5% weight loss, with similar results across income levels and 91.4% remaining in the program.

A preprint reported pig kidneys vitrified at -150°C for up to one year were magnetically rewarmed and transplanted into pigs, where they produced urine; long-term survival remains undemonstrated.

AI-assisted scan analysis across 101,243 cancer-center patients linked brown fat to less cardiovascular and psychiatric disease; losartan increased brown-fat activity in mice.

MIT nanoparticles reprogrammed up to 80% of targeted natural killer cells inside mice, producing melanoma remission in three of five animals and clearing T cell lymphoma without external cell manipulation.

A Polish robot chemist ran for 30 days, made more than 3,000 decisions and discovered four highly selective transformations after training decision policies against a digital twin.

Fields medalist Shing-Tung Yau convened eight institutions toward an alliance of more than 1,000 scientists developing mathematics models, generated conjectures, formalized proofs and inspectable reasoning.

An automated workflow called OSAR used liquid-handling robots and yeast recombination to assemble 31 of 34 bacterial gene clusters up to 78 kilobases without custom DNA synthesis, expressing 29 products.

TU Wien/PTB and Tsinghua independently published the first operating nuclear clocks in Nature; Vienna's prototype ran over 24 hours without atomic-clock steering and tightened clock-based dark-matter constraints.

Prime-editing cell recorders traced mouse development from a fertilized egg, reconstructing the lineage of 1.3 million embryo cells in one of two new studies.

DOE awarded $159M to twelve Genesis AI-science projects, including Caltech catalyst discovery, as technology firms pledged $2.4B in compute and researcher credits totaling $100M were announced.

Biohub announced a $1.8 billion virtual-cell initiative, including $300 million from Google DeepMind, Meta and Isomorphic Labs and more than $500 million from DOE over five years.

A light-activated gene-therapy trial reported restoration of aspects of visual function in the New England Journal of Medicine.

Station recovered 62.7% of original findings from three ICLR papers with results withheld and web access disabled, versus 15.4% for a rival system.

OpenAI published 722 manuscripts in 372 result families from an unreleased model, many with Lean formalizations and reasoning summaries; mathematicians objected that the bulk release broke an August pledge to coordinate.

Utah authorized Nolla Health's AI to assess acne photographs and prescribe low-risk drugs, initially with clinician sign-off and potentially without it later.

Utah expanded its clinical AI sandbox to acne assessment and low-risk prescribing, with initial clinician approval and named third-party auditors.

Neuralink pretrained its brain-computer-interface decoders on 50,000 hours of unlabeled recordings, reporting weeks of stable cursor control without recalibration in some participants.

A urine liquid biopsy reading cell-free RNA detected localized bladder cancer at 95% sensitivity across 683 samples and forecast which patients would respond to BCG immunotherapy versus chemotherapy.

FDA approved Edwards Lifesciences' AUTUS pediatric pulmonary valve, adjustable by catheter from about 13 to 22 mm; a 62-patient study reported no replacements through one year, with only two expa.

An AI speech clock estimated biological aging from four minutes of voice in 2,928 Spanish-speaking adults; its age gap tracked brain atrophy and Alzheimer's blood markers.

A lipid screened specifically for large gene-editor cargo delivered base editors to 48% of mouse brain cells and 27% of lung cells, while a light- and inflammation-triggered viral vector reached target tissues.

Agent swarms given only a research scope and an instruction to keep making progress produced weeks of paper-length drafts and machine-checked proofs across five fields faster than their authors could review.

Structural modeling of 44 Bundibugyo virus genomes predicted MBP134 remains protective against the 2026 outbreak while Ebanga's antibody binding may be compromised.

MIT researchers used language tied to 49 risk factors to classify about 16,000 crisis-text conversations, releasing an open tool that shows counselors the words behind each estimate; researchers say further validation is required before clinical use.

MIT researchers built an autonomous system that assembles and measures the lipid nanoparticle shells used to deliver RNA therapies, tuning particle size and shape in real time to hit organ-targeting specifications.

AlphaFold added about 8,000 predicted viral protein-pair structures from 23 human-health-relevant virus families to a free pandemic-preparedness portal.

Roughly 950 Claude agents searched more than 200,000 enzymes in 21 hours and flagged a previously unrecognized CRISPR-like repeat system for human investigation.

GRAIL's Galleri detected signals across 17 cancer types in a 35,878-person prospective study, with 99.64% specificity but 39% overall sensitivity; an FDA advisory panel backed its benefits 7 to 2.

The EAGLE model detected esophageal cancer from ordinary chest CT scans at 90% sensitivity and 98.5% specificity across validation spanning 12 centers, 80,612 patients, and three countries.

OpenAI said one internal model resolved more than 100 open mathematics problems since August and created an independent advisory group that cannot pace or halt the work.

A follow-up analysis of the RIO trial found two long-acting broadly neutralizing antibodies, paired with patients' own immune response, associated with an estimated 36-week half-life for the intervention.

Anthropic confirmed a Bay Area wet lab where Claude directs robotic biology experiments, extending its roughly $400 million Coefficient Bio acquisition and Claude Science platform into physical validation.

Novo Nordisk, down about 20% over the past year, partnered with Anthropic to apply Claude Science's biological-reasoning tools directly to drug discovery rather than back-office work.

The DOE opened a $215 million milestone competition to build the first fault-tolerant quantum computer with over 100 logical qubits, targeting fusion and battery-materials problems beyond classical computing.

A single-draw blood test combining microRNA signatures with a standard protein marker caught stage 1 and 2 pancreatic cancer 87% of the time across a 1,785-person study spanning four countries.

Stanford researchers emptied a mouse's developing cortex and transplanted human brain organoids that grew fivefold to fill over 90% of the space and wired into the animal's nervous system.

Five drug companies pooled 20,167 proprietary protein-ligand structures through the AI Structural Biology Network to fine-tune OpenFold3 via federated learning, outperforming public and single-firm models on held-out cases.

Insilico Medicine proposed a Longevity Vaccines research program using AI-chosen targets and circular mRNA lipid nanoparticles to arm a patient's own T cells against senescent cells.

The OpenAI Foundation launched Public Data for Health, funding $40M for open cancer-vaccine data at UNC and $500K to preserve failed biotechs' regulatory filings.

UCSF researchers built a single brain implant with roughly 253 surface electrodes that simultaneously decoded intended speech into on-screen text and intended gestures into an animated avatar.

Definium's single-dose LSD therapy DT120 posted its third straight Phase 3 win, cutting anxiety scores 9.8 points versus 4.7 for placebo in a 245-patient trial, with an FDA pre-NDA meeting planned.

University of Miami researchers used AI shape comparison across more than 214 million predicted proteins to identify TM184C, a hidden GPCR-family receptor that builds cell-to-cell bridges.

Densely chemical-modified guide RNA armored along its entire length reached nearly 70% prime editing efficiency in mouse liver from a single lipid-nanoparticle injection at a clinically translatable dose.

John Jumper, 2024 Nobel Chemistry co-laureate and AlphaFold co-creator behind 200M+ open protein structure predictions, joined Anthropic.