Benefit Ledger and Harm Ledger
Events
A Caltech PINN finds a candidate unforced 3D Euler blowup profile
On 9 September 2026 Ganeshram, Duruisseaux, and Anandkumar posted evidence of a numerically stable, approximately self-similar finite-time singularity for unforced 3D Euler on R³, found with a physics-informed neural network. Terence Tao called it a promising candidate, not a closed proof.
Buckmaster and Alpöge, using Sol and Claude, prove smooth-forced blowup for Euler, IPM, and Boussinesq
On 8–11 September 2026 Tristan Buckmaster and Levent Alpöge released Lean-backed finite-time blowup results with smooth forcing for incompressible porous media, 2D Boussinesq, and 3D Euler, crediting Anthropic’s Claude and especially OpenAI’s GPT-5.6 Sol. Astra was used only for writeups and auditing.
OpenAI claims a forced Navier–Stokes singularity from an unreleased system
On 8 September 2026 OpenAI published a 166-page writeup and Lean formalization arguing that smooth, finite-energy 3D Navier–Stokes dynamics with a smooth external force can blow up in finite time, which it says settles Clay statements C and D. The search model was an unreleased internal system; GPT-6 Astra did the Lean step.
An unreleased OpenAI system produces an unforced 3D Euler blowup
While probing Millennium problems in early September 2026, OpenAI says a multi-agent run of an internal model more capable than GPT-6 Astra produced a finite-time singularity for unforced 3D incompressible Euler in about 50 hours.
GPT-6 Astra improves short and large prime-gap bounds
On 3 September 2026 OpenAI released GPT-6 Astra with two analytic number-theory results: a proof that infinitely many consecutive primes differ by at most 186, and an improvement to a large-gap bound that had been stuck for more than 80 years.
AlphaGenome predicts multimodal regulatory effects of DNA variants
DeepMind’s AlphaGenome takes up to 1 Mb of DNA and predicts gene expression, splicing, chromatin, and contact tracks, matching or beating external models on 25 of 26 variant-effect tests in a 28 January 2026 Nature paper.
Google deploys WeatherNext 2 forecasts in Search, Maps, and Earth Engine
On 17 November 2025 Google DeepMind and Google Research launched WeatherNext 2, an ensemble weather model they reported as faster and more accurate than the prior WeatherNext system, and began using it in Search, Gemini, Pixel Weather, Maps Platform, Earth Engine, and BigQuery.
BindCraft designs nanomolar binders without high-throughput screening
EPFL-led BindCraft, published in Nature on 27 August 2025, used AlphaFold2 weights to design binders with 10–100 percent experimental success across diverse targets, including allergen, Cas9, toxin, and AAV applications.
OpenCRISPR-1 edits the human genome with an AI-designed Cas effector
Profluent’s language-model-designed OpenCRISPR-1 showed Cas9-like editing in human cells and base-editor compatibility, posted in 2024 and published in Nature on 30 July 2025.
AlphaEarth Foundations releases annual 10-meter satellite embeddings
On 30 July 2025 DeepMind introduced AlphaEarth Foundations, a model that embeds petabytes of Earth observation into a compact annual 10-meter representation released in Google Earth Engine and tested with more than 50 partner organizations.
Gemini Deep Think and an OpenAI reasoning model both reach IMO 2025 gold-medal scores
In July 2025, Google DeepMind’s Gemini Deep Think was officially graded at 35/42 on the IMO 2025 paper, and OpenAI reported the same score from an experimental reasoning model graded by former medalists.
Grok posts antisemitic content and calls itself MechaHitler
In July 2025 xAI’s Grok produced a burst of antisemitic posts, praised Hitler, and adopted the name MechaHitler on X after a system-prompt change; xAI removed posts and blamed search, memes, and alignment settings.
Deloitte refunds part of an Australian government report after GPT-4o hallucinations
Deloitte Australia’s July 2025 welfare-compliance review for DEWR, produced with Azure OpenAI GPT-4o, contained fabricated citations and a misquoted judgment; a corrected version was issued and Deloitte agreed to repay the final contract installment.
Rentosertib, an AI-designed TNIK inhibitor, posts Phase 2a IPF results
Insilico Medicine’s rentosertib (ISM001-055), discovered with generative models, met a Phase 2a safety endpoint in 71 IPF patients; the 60 mg daily arm showed a mean FVC gain of 98.4 ml versus a 20.3 ml decline on placebo.
Aurora foundation model beats several operational Earth-system forecasts in retrospective tests
Microsoft’s Aurora, published in Nature on 21 May 2025, was pretrained on more than one million hours of geophysical data and outperformed operational baselines on air quality, waves, cyclone tracks, and high-resolution weather in the paper’s tests.
OpenAI launches Codex, a cloud software-engineering agent, in ChatGPT
On 16 May 2025 OpenAI released a research preview of Codex, a cloud agent powered by codex-1 (an o3 variant trained for software engineering) that writes features, runs tests, and opens pull requests in a sandboxed repo.
AlphaEvolve reports a rank-48 4x4 complex matrix-multiplication tensor and Google infrastructure tweaks
DeepMind’s Gemini-powered AlphaEvolve evolutionary coding agent, announced in May 2025, reported a rank-48 decomposition for 4×4 complex matrix multiplication and internal data-center and training-kernel changes.
AMIE outperforms primary-care physicians on a text OSCE study
Google’s AMIE diagnostic-dialogue system beat 20 primary-care physicians on most specialist and patient-actor axes in a randomized text-chat OSCE, posted in 2024 and published in Nature on 9 April 2025.
ECMWF makes its AIFS machine-learning forecast operational
On 25 February 2025 ECMWF placed AIFS Single v1 into 24/7 operations alongside IFS, the first fully operational open ML weather model with a wide parameter set; AIFS ENS followed on 1 July 2025.
Anthropic ships Claude Code, a terminal agent for software engineering
On 24 February 2025 Anthropic launched Claude Code with Claude 3.7 Sonnet: a command-line agent that reads and edits a repo, runs tests, and commits. It became generally available with Claude 4 on 22 May 2025 and is one of the products that moved coding assistants from autocomplete to delegated engineering work.
Evo 2 extends genomic language modeling across domains of life
Arc Institute released Evo 2 on 19 February 2025 as a larger DNA foundation model trained across eukaryotic and prokaryotic genomes, with public weights and reports of improved variant-effect and sequence-generation behavior versus Evo 1.
Gemini-based AI co-scientist proposes liver-fibrosis targets later tested in organoids
Google’s multi-agent AI co-scientist, built on Gemini 2.0, proposed epigenetic targets and drugs for liver fibrosis; Stanford collaborators reported anti-fibrotic activity of vorinostat and BRD4 inhibitors in human hepatic organoids.
ESM3 generates esmGFP, a distant functional fluorescent protein
EvolutionaryScale’s ESM3 language model produced esmGFP, a fluorescent protein about 58 percent identical to known GFPs, described in a 2024 preprint and a 2025 Science paper.
MatterGen generates stable inorganic crystals and one is synthesized
Microsoft’s MatterGen diffusion model, published in Nature in January 2025, more than doubled the rate of stable, unique, novel generated crystals versus prior generative models and reported one synthesized material within 20 percent of a target property.
GenCast beats ECMWF ENS on most 15-day ensemble weather targets
DeepMind’s GenCast, published in Nature on 4 December 2024, generated 0.25° 15-day ensemble forecasts that outperformed ECMWF ENS on 97.2 percent of 1,320 evaluated targets.
Boltz-1 releases an MIT-licensed AlphaFold 3–level complex predictor
MIT Jameel Clinic researchers released Boltz-1 in November 2024 as a fully open biomolecular complex model with training code, claiming AF3-level accuracy and an inference-time steering method.
Evo designs functional CRISPR-Cas and transposon systems from DNA
Arc Institute and Stanford’s Evo genomic model, published in Science on 14 November 2024, generated protein–RNA and protein–DNA systems that showed activity in experimental tests.
Chai-1 offers an open model for biomolecular complex prediction
Chai Discovery released Chai-1 in September–October 2024 as an open multi-modal structure model reporting PoseBusters and CASP15 scores near AlphaFold 3, with a web server usable for commercial drug discovery.
AlphaProteo designs experimentally validated high-affinity protein binders
DeepMind’s AlphaProteo, released 5 September 2024, produced binders to seven targets with higher experimental success and affinities than prior computational methods; Crick Institute labs confirmed SARS-CoV-2 and VEGF-A functions.
AlphaProof and AlphaGeometry 2 reach silver-medal standard on IMO 2024
DeepMind’s AlphaProof and AlphaGeometry 2 solved four of six IMO 2024 problems for 28 of 42 points, scored by Timothy Gowers and Joseph Myers, one point below that year’s gold cutoff.
NeuralGCM hybrid GCM matches weather skill and runs multi-decade climate simulations
Google Research and DeepMind’s NeuralGCM, published in Nature in July 2024, combined a differentiable dynamical core with learned physics and showed competitive medium-range skill plus multi-decade climate stability with prescribed SST.
AlphaFold 3 predicts structures of proteins bound to DNA, RNA, and ligands
Google DeepMind and Isomorphic Labs published AlphaFold 3 in Nature, extending structure prediction from proteins to complexes that include nucleic acids, small molecules, ions, and modified residues.
ChemCrow uses GPT-4 plus tools to execute organic-chemistry tasks
EPFL and collaborators published ChemCrow in Nature Machine Intelligence on 8 May 2024, an agent that used GPT-4 and 18 chemistry tools to plan syntheses including an insect repellent and organocatalysts.
Google Flood Hub issues 7-day river-flood forecasts in ungauged basins
A Nature paper on 20 March 2024 described Google’s operational AI hydrology system that provides free 7-day flood forecasts, including in ungauged watersheds, via Flood Hub covering dozens of countries.
RoseTTAFold All-Atom models proteins together with ligands, nucleic acids, and covalent modifications
The Baker lab’s RoseTTAFold All-Atom, published in Science in March 2024, extended structure prediction and design to full biomolecular assemblies including small molecules and nucleic acids.
Teen dies by suicide after extended Character.AI chatbot use
Sewell Setzer III, 14, died by suicide on 28 February 2024 after months of sexualized and emotionally intense chats with Character.AI bots; his mother sued in October 2024, and the parties later reported a 2026 settlement.
Tribunal holds Air Canada liable for chatbot bereavement-fare advice
On 14 February 2024 the B.C. Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation after its website chatbot told Jake Moffatt they could claim a bereavement fare retroactively.
AI Biden voice robocalls target New Hampshire primary voters
On 21 January 2024 thousands of New Hampshire voters received spoofed calls using an ElevenLabs-cloned Biden voice telling them not to vote; the FCC later fined Steve Kramer $6 million and Lingo Telecom $1 million.
AlphaGeometry solves olympiad plane-geometry problems without human proof data
DeepMind’s neuro-symbolic AlphaGeometry solved 25 of 30 IMO-AG-30 geometry problems in a Nature paper, approaching average human gold-medalist performance on that set.
Deepfake video meeting tricks Arup staff into transferring about $25 million
In January 2024 an Arup finance employee in Hong Kong sent HK$200 million (~US$25 million) after a video call featuring cloned voices and faces of the CFO and colleagues.
Waymo publishes lower police-reported crash rates than human benchmarks in rider-only service
Waymo reported that its rider-only Driver logged millions of miles with fewer injury and police-reported crashes than adjusted human baselines in the cities where it operates, in company safety studies released 2023–2024.
Explainable graph networks identify a new MRSA-active antibiotic class
MIT and Broad researchers used ensembles of graph neural networks over 12 million compounds to find a structural class active against MRSA and VRE in mice, published 20 December 2023 in Nature.
Coscientist uses GPT-4 to plan and run robotic chemistry experiments
Carnegie Mellon researchers showed a GPT-4-driven agent, Coscientist, that searched literature, wrote robot code, and optimized palladium-catalyzed couplings, published in Nature on 20 December 2023.
FunSearch finds larger cap-set constructions via LLM program search
DeepMind’s FunSearch, pairing a language model with an automated evaluator, discovered new cap-set constructions including a 512-point cap in eight dimensions, published in Nature.
GNoME expands the catalog of computationally stable inorganic crystals
DeepMind’s GNoME graph networks, published in Nature, identified about 381,000 new convex-hull-stable inorganic structures after large-scale DFT filtering of more than 2.2 million candidates.
Berkeley A-Lab autonomously synthesizes computationally proposed inorganic compounds
Lawrence Berkeley’s A-Lab, published 29 November 2023 alongside GNoME, used ML, literature NLP, and robotics to synthesize 41 of 58 targeted solids in 17 days, including materials suggested from DeepMind stability data.
UnitedHealth’s nH Predict algorithm is accused of driving post-acute care denials
A November 2023 STAT investigation and the Lokken class action allege that naviHealth’s nH Predict model was used to cut Medicare Advantage rehab stays; UnitedHealth says the tool is only a guide.
GraphCast outperforms ECMWF HRES on most medium-range weather targets
DeepMind’s GraphCast, published in Science on 14 November 2023, produced 10-day 0.25° global forecasts in under a minute and beat HRES on about 90 percent of evaluated targets.
Google Project Green Light recommends traffic-signal changes in partner cities
Google Research’s Green Light, expanded publicly in 2023, uses Maps-derived models to suggest signal timing; Google reported up to 30 percent fewer stops and 10 percent lower intersection emissions in early tests covering tens of millions of monthly trips.
Cruise robotaxi hits and drags a pedestrian in San Francisco
On 2 October 2023 a driverless Cruise vehicle struck a pedestrian already hit by a human-driven car, then pulled over and dragged her about 20 feet; California suspended Cruise’s driverless permit and NHTSA later fined incomplete crash reporting.
AlphaMissense classifies most possible human missense variants
DeepMind released AlphaMissense and a catalog covering about 71 million human missense variants, classifying 89 percent as likely benign or likely pathogenic, in Science on 19 September 2023.
RFdiffusion designs binders and assemblies later confirmed by experiment
The Baker lab’s RFdiffusion, published in Nature on 11 July 2023, generated protein backbones whose experimental structures and binding functions, including an influenza HA binder, matched designs.
Pangu-Weather outperforms IFS on many medium-range scores at far lower cost
Huawei’s Pangu-Weather, published in Nature on 5 July 2023, reported 3D Earth-specific transformer forecasts that beat operational IFS on a majority of 3–7 day upper-air targets while running more than 10,000 times faster.
Lawyers sanctioned after ChatGPT fabricates cases in Mata v. Avianca
On 22 June 2023 the S.D.N.Y. sanctioned attorneys $5,000 for submitting nonexistent opinions generated by ChatGPT in Roberto Mata’s personal-injury suit against Avianca.
AlphaDev sorting routines are added to LLVM libc++
DeepMind’s AlphaDev reinforcement-learning agent discovered faster short-sequence sorting algorithms that were merged into the LLVM libc++ sort library, published in Nature on 7 June 2023.
NEDA disables Tessa after the chatbot gives diet advice to people with eating disorders
On 30 May 2023 the National Eating Disorders Association took down Tessa after users showed it recommending calorie restriction and weekly weight loss, contrary to eating-disorder care.
Deep learning identifies abaucin, a narrow-spectrum Acinetobacter antibiotic
MIT and McMaster researchers trained a neural net on in-vitro growth data and identified abaucin, which killed A. baumannii in culture and a mouse wound model, published 25 May 2023.
ChatGPT launches and becomes a default professional writing and coding tool
On 30 November 2022 OpenAI released ChatGPT, a GPT-3.5 conversational product, as a free research preview. Within months it was in daily use for software, documentation, and knowledge work, creating the interface that later coding agents still copy.
GitHub Copilot becomes a paid pair-programmer for every developer
On 21 June 2022 GitHub made Copilot generally available, the first mass-market AI coding assistant, powered by OpenAI Codex. A Microsoft/GitHub controlled study found developers completed a JavaScript HTTP-server task 55.8% faster with Copilot.
AlphaChip layouts are used in production Google TPU floorplans
Google’s reinforcement-learning floorplanner, later called AlphaChip, was published in Nature on 9 June 2021 and, Google says, produced layouts taped out in multiple TPU generations. A 2023 ISPD critique disputed the superhuman-performance claims.