AI Topic
AI-driven discoveries in math, physics, chemistry, biology, materials. Curated and summarized from dozens of sources by AIBriefs. RSS
Aaronson says his blog draws near-daily emails claiming to solve P versus NP, Goldbach's conjecture or quantum gravity, which he has long dismissed. He now says an unknown amateur's proof of the Riemann hypothesis would leave him less certain it was wrong, since an AI model might be behind it.
Roman Yampolskiy's podcast episode features UT Austin quantum computing and complexity theorist Scott Aaronson discussing AI systems cracking math problems that went unsolved for centuries. Aaronson founded UT Austin's Quantum Information Center and previously taught at MIT.
Anthropic's life sciences team has deep research partnerships with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. It has also begun in-house drug discovery, including programs for rare childhood muscular dystrophy, with Claude handling much of the scientific and operational work.
In an MLST clip, Lessard argues machine-generated proofs that are merely correct fall short of real mathematics — he wants provenance, the reason a machine reached a given answer. Human understanding remains central once the work moves beyond proof-checking.
Ai2's "Old Ideas, Novel Problems" tested six judges — GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Opus 4.5 and Opus 4.6 — and found small changes to the evaluation setup produce dramatically different novelty verdicts on the same ideas. Changing just a few words in a prompt shifted a model's judgment.
C-BRAIN, a consortium of 18 institutions launched in November 2025 and led by Dr. Randall Bateman, is working with NVIDIA on open-source AI tools to mine previously inaccessible Alzheimer's data, including unpublished findings and negative results. It targets 57 million people living with dementia and 10 million new cases a year.
Kevin Hartnett, author of The Proof in the Code, says LLMs now turn a math paper into Lean code in a fraction of the time a human team would take. The Mathlib community calls the resulting flood of Lean a "Trojan horse" because much of it isn't verified.
Both labs' models reportedly solved 3-SAT with a 2^(n/√log n) bound on Oct 16, beating the Exponential Time Hypothesis — not P=NP, but faster than thought possible. An internal Anthropic model did it with 14M output tokens and no human input.
Two physicists reportedly asked an AI to pick its own problem, solve it, and draft a paper suitable for a prestigious journal; the run produced a novel, significant result in two hours. No model, journal, or paper title is named in the post.
Mathematicians Ankur Moitra, Steven Brunton and Alex Kontorovich discuss AI solving problems including Navier-Stokes, the quasi-Riemann hypothesis and the unique games conjecture, plus hundreds more.
Scott Aaronson tells how his nine-year-old son told his wife, complexity theorist Dana Moshkovitz, she got "cooked" by AI after OpenAI released an AI-found proof of the unique games conjecture — the problem she has worked on since 2009. She has since spent nearly all her time on it.
AI safety startup Resolution used orchestrator and sub-agents to autoformalize the resolution of singularities, spending ~100B input and ~300M output tokens (~$135,000) over about a week to produce ~600k lines of Lean. The work began with a 112-page atomic-step plan and aims to link loss-landscape geometry to generalization.
OpenAI's internal model found a counterexample to Erdős's unit distance conjecture (open since 1946) in May, then published proofs of the Cycle Double Cover problem, ten more open problems by July, and Navier-Stokes in September. Scott Armstrong called the Navier-Stokes construction "unbelievable," possibly beyond what a human could devise.
The UC Berkeley mathematician now grades homework only on completion, saying AI use is unavoidable and there is "no point" paying readers to grade for correctness. He said if AI keeps improving, "in a few years, there'll be nothing human mathematicians can do that AI can't do better."
Tao says OpenAI published answers to 722 unsolved maths problems without building a new model, using the conundrums as raw training material. He argues AI firms chase benchmarks, producing "terrible proofs" that skip elegance or understanding.
Latent Space essay traces OpenAI's internal model finding a counterexample to Erdős's unit distance conjecture (open since 1946), then a Cycle Double Cover proof and ten more open problems by July, plus a Navier-Stokes proof in September. It uses WT Gowers' hint-sequence complexity formula, 2^k + t, to predict which problems LLMs can crack.
Artificial Societies' behavior-prediction model outperformed nine leading AI models at forecasting how people answer surveys, cutting error by 74%. Its validation framework, the Artificial Societies Benchmark, is described in an arXiv preprint by Chidichimo, Jung, Wallis and He.
Scientists grew a human mini-brain and used it to control a humanoid robot, which learned to grasp objects with 94% accuracy, chase a laser, and transfer learning between tasks. It outperformed a conventional neural network in limited tests.
In a study of 124 Harvard social science papers spanning statistical analyses, formal models, and qualitative work, Opus 4.8 outperformed human experts as a peer reviewer by every measure. Ethan Mollick argues academics must adapt the practice of science to AI quickly.
Anthropic is inviting principal investigators at academic institutions, nonprofits, and government labs to apply for 8–12 week projects using Claude on causal human biology, structural and computational molecular biology, medicinal chemistry, drug discovery, and clinical science. Anthropic runs its models and harnesses at no cost, and outputs are published with open-source data and code.
PLANTEUR extends the PASTIS benchmark to six tropical regions — Réunion, Guadeloupe, Martinique, Guyane, Mayotte and Saint-Martin — with 2,750 Sentinel-2, Sentinel-1 and Landsat-8 time series and 4,506 km² of parcel labels across seven crop classes. Baselines cover leave-one-island-out, few-shot and PASTIS-in-training splits.
OlmoEarth v1.3 outputs one 128-dimensional vector per 10-meter pixel from Sentinel-1, Sentinel-2 and Landsat time series, with the first 64 dimensions usable alone. Ai2 precomputed annual embeddings for the whole planet and shared them on Source Cooperative, plus a no-sign-in browser Embeddings Explorer.
The Association for Human Mathematics formed after OpenAI released 722 papers at once; members split into "vegetarians" who won't work for AI companies and "vegans" who eschew AI in research entirely. Terence Tao warned such results are often "dumped" on the field in impenetrable form.
Carbon-X is a decoder-only model built on Qwen3-14B that reads raw DNA in a chat and answers in English, scoring 66.7% average across five DNA benchmarks (self-reported). It matches GPT-6.1 Sol within a point at roughly 1,000x lower inference cost, per Hugging Face.
StormScope EU is a regional short-term forecasting model trained on European weather data, predicting cloud cover across 16 Meteosat channels at 10-minute intervals from EUMETSAT satellite observations. NVIDIA demoed a live pipeline with Earthmover at GTC Berlin; AXA is using FourCastNet 3 to study regional cyclones.
Stubbs, Harvard's Samuel C. Moncher Professor of Physics and Astronomy and Dean of Science from 2018 to 2024, helped discover the universe's accelerating expansion with the High-Z Supernova Search Team in 1998. He was also inaugural project scientist for what is now the Vera C. Rubin Observatory and is a member of JASON.
Researchers resolved the Kannan-Lovász-Simonovits conjecture, open since 1995, showing the best flat cut matches the best cut of any shape within a universal constant regardless of dimension. The result came from a human-directed agentic process where AI explored possibilities and researchers pruned and redirected, then machine-checked formal verification.
Thousands of GPT Rosalind agents analyzed millions of mouse brain cells in 32 minutes, work OpenAI estimated would take human researchers months to a year. The agents reproduced prior human findings and surfaced new leads; blinded human experts preferred their findings 75% of the time.
Computer scientist Scott Aaronson describes a bet with colleague Ernie Davis that an AI would reach gold-medal level at the International Math Olympiad by 2026. He compares it to chess and Go, each once called the real test until computers beat them.
RFD4-Proteina, a flow-matching model from the Baker Lab and NVIDIA, co-designs protein sequences and all-atom structures, skipping the separate inverse-folding handoff to models like ProteinMPNN. It ships as an open BioNeMo model plus a NIM microservice, with a tutorial targeting PD-L1 binders via reward-guided inference-time search.
NVIDIA's CUDA-X library moves large sparse-system solves onto GPUs; Dassault Systèmes reports 8-10x speedups in Abaqus and Applied Materials up to 10x in Ginestra versus CPU-only. ZEISS SMT ran an EUV lithography plasma benchmark 4x faster, finishing in six days and 14 hours.
RFD4-Proteina combines programmable protein design with search capabilities, released as an open model at GTC Berlin. Siemens Healthineers is also using NVIDIA intelligence for adaptive radiotherapy with agentic co-planning and robotic control.
Radical Numerics used its Omnii-1 model to engineer viruses targeting an antibiotic-resistant infection; lab tests cleared the patient's bacteria before compassionate-use administration.
In a SAIR interview, Tao argues big tech once gave mathematicians useful tools, but recent events show community-driven models are needed to keep AI aligned with the mathematical community.
Agent4Sci evaluates frontier coding agents on scientific ML benchmarks and finds they improve results mainly through engineering tricks rather than domain insight. Targeted plug-in skills significantly boost performance.
Scott Aaronson discusses the debate over AI-generated mathematics and what human mathematicians do if AI can prove theorems. He covers controversy around recent AI results and responses from mathematicians including Terry Tao.
Electric Plant raised an $11.2M seed led by Onto Ventures and Climactic to commercialize sensors plus Bombadil, an LLM trained on three years of plant electrical-signal data. CEO Justin Mast says the model relays a plant's water, nutrient, and light needs to growers.
Anthropic's internal research model designed minibinders against scorpion neurotoxins over 64 hours, working mostly autonomously from one prompt and four follow-ups. Adaptyv Bio and Twist Bioscience tested 211 designs; 17 met binding criteria. All 211 designs and data are being released.
In a clip, mathematician Terence Tao argues the AI era is revealing that the common notion of intelligence is not really accurate. No specific model, benchmark, or result is cited in the excerpt.
Podcast interview with Alex Rives, Head of Science at Biohub and former lead of the ESM protein language model project, on building a predictive world model of life. The effort is backed by an expanded $1.8 billion Virtual Biology Initiative with partners including the Department of Energy.
Among 377 results OpenAI announced on October 6, 2026, the quasi-Riemann hypothesis proof stands out; Stanford's Jared Duker Lichtman calls it "the most significant AI-for-math achievement so far, by quite a margin." NYU's Maksym Radziwiłł says the implications are "staggering" since most RH consequences follow from quasi-RH.
Researchers at Saint-Gobain are using Microsoft Discovery to accelerate research and identify promising new molecules for cement additives. The goal is stronger, lower-carbon cement by reducing reliance on clinker, a major source of emissions in cement production.
A DoubleAI paper claims an AI agent constructed explicit regular graphs with girth (8/5) log n, described as the first improvement in 38 years and a leap toward the theoretical maximum.
Pardon says AI catches subtle errors in even his messiest proofs and proves nontrivial results, calling the progress "world-changing for mathematics." He says he would have said the opposite a year ago.
The statement of intent targets a single pretrained Large Physics Model, the first foundation model for collider physics, trained on billions of simulated proton-proton collisions at the LHC. It would handle reconstruction, particle identification, anomaly flagging and detector-design evaluation in one model.
Fields Medalist Terence Tao argues math's next breakthroughs will come from collaboration between humans and AI rather than a lone genius. He also frames math as the check on AI answers that merely sound right.
STAT's profile traces Thrive Capital's 18-month courtship of Alphabet's Isomorphic Labs, which had kept outside investors at bay and wasn't actively raising. Isomorphic is built on AlphaFold, the protein-modeling tool behind a Nobel Prize in chemistry, and aims to turn it into an engine for new medicines.
A non-astronomer reports a 0.21% dip every 36.34 days in TIC 402710280, 218 light-years away, implying a planet 3.75x Earth's size; a separate Reddit user cites a 0.05% dip every 3.18 days at 116 light-years. Both are unconfirmed candidates with pre-registered transit predictions.
A Bristol-led team fed Braille tactile sensor data into forebrain organoids on multielectrode arrays; a decoder distinguished letters from the organoid responses. Single organoids decoded accurately 61% of the time, a team of three hit 83% — better than decoding the noisy raw signals directly.
OpenAI published over 700 model-proved math results, but statistics, ML theory, optimization and numerical analysis make up under 1% of them, versus roughly 20% of arXiv theorems. The gap stays statistically significant even when baselined against arXiv papers that solve open problems.
Odyssey placed a simulation of a fruit fly's brain, built from Google's connectome map of the fly's actual neurons and connections, inside an AI-generated world. The simulated brain navigates the imagined environments on its own.
A human-AI team reports a proof for a long-standing open problem in chaos theory, with Shanghai AI Lab's Shusheng-S2 contributing to the work. Details of the theorem and the model's specific role are not given in the available snippet.