AI Topic
AI-driven discoveries in math, physics, chemistry, biology, materials. Curated and summarized from dozens of sources by AIBriefs. RSS
PACMAN makes plasma control decisions in about 20 milliseconds, fast enough to predict and prevent instabilities before they occur. Researchers at Princeton and PPPL report the framework was successfully tested on fusion plasma.
Apple researchers trained SimpleDesign on over 2M sequence-structure pairs with a single-stage end-to-end objective, combining discrete cross-entropy for sequences and a regression objective for structures. The model skips the usual multi-stage autoencoder-plus-latent-generative pipeline and reports competitive results on co-design and unconditional sequence/structure generation benchmarks.
Interview covers how AI, formal verification, and crowdsourcing could speed up mathematical research, which Tao describes as traditionally slow, solitary, and handcrafted. He points to the SAIR competition platform as a way to involve many more people.
OpenAI ended its sponsorship of the Caltech "Mathathon" after current and former Caltech mathematicians warned the event "is likely to have destructive impacts for the mathematical community." The letter, signed by roughly 1,000 mathematicians including Fields medalists Peter Scholze and Hugo Duminil-Copin, calls AI firms' unverified problem-solving claims "slop mathematics."
César de la Fuente's lab searches living and extinct genomes with Codex and ChatGPT for antimicrobial candidates against drug-resistant infections. OpenAI published a researcher profile and video on the workflow.
BioIR accelerated protein-complex structure generation across 4,777 proteomes — about 31 million candidate complexes, with 1.81 million released as high-confidence predictions in the AlphaFold Database expansion. It keeps the PyTorch workflow, using optimized kernels and CUDA Graphs, with Ray running a model replica per GPU.
Andreas Thom says OpenAI's non-sofic groups result built heavily on prior work by him and Gábor Kun, and that his ChatGPT and Codex interactions may have fed the model. OpenAI quietly amended its writeup after criticism; Tristan Buckmaster raised similar concerns.
Nature study builds 100+ bottom-up RNA transfer vehicles from synthetic protein assemblies, surpassing widely used delivery vehicles by several orders of magnitude in RNA transfer efficiency.
Google DeepMind's Peter Battaglia discusses how machine learning is transforming weather forecasting. AI models now require fewer resources and are accessible to anyone via Hugging Face, enabling individuals to predict weather themselves.
Tsimerman announced he is joining OpenAI hours after winning the Fields Medal, a coincidence profiled by Quanta's Kevin Hartnett the same morning. He scored a perfect 2004 IMO and says AI agents may outcompete humans in his area of problem-solving within a few years.
OpenAI says ~10,000 coordinating agents using an unreleased model more capable than GPT-6 Astra produced a solution in 88 hours, with a writeup and Lean formal proof. NYU's Tristan Buckmaster disputes the process, saying he and Levent Alpöge worked the problem for nearly a year with Claude and Codex (mainly GPT-5.6 Sol) before an August 15 breakthrough.
a16z's Lisha Li interviews OpenAI mathematicians Mehtaab Sawhney and Mark Sellke on AI's advancing math capabilities and reasoning. Another episode features University of Toronto professor Daniel Litt on what frontier models still miss about mathematical intuition.
AlphaGenome Atlas is a free-to-use web portal with precomputed predictions for all 9 billion possible single-letter DNA changes, forming a 1-petabyte dataset. It introduces the AlphaGenome Variant Impact (AVI) score, combining AlphaGenome and AlphaMissense, and is also available via Antigravity.
A Reddit post reports that researchers simulated a fruit fly's entire brain and made it play the video Bad Apple. The post has 31 upvotes and 14 comments.
Google DeepMind selected 16 organizations for its Accelerator: AI for the Planet in APAC, supporting startups, nonprofits, and researchers using frontier AI for climate, agriculture, and biodiversity. The program kicked off with a bootcamp in Singapore.
Claude produced the first complete computer-checked proof of Fermat's Last Theorem, working largely autonomously over 11 days. It wrote 13 million lines of Lean and proved 29,500 intermediate theorems.
WeatherNext 3 is the first global weather model generating forecasts every hour, drawing directly from raw satellite imagery. It predicts surface variables at 5km resolution and is integrated into Google Search, Maps, and Gemini.
Epoch AI's new benchmark, FrontierMath Erdős, features 68 unsolved problems selected by mathematician Thomas Bloom. A pre-release Astra solved 2, while GPT-5.6 Sol, GPT-5.5, Claude Fable 5.1, and Claude Fable 5 all failed.
A paper co-authored by Stanford professor Jared Duker Lichtman, linked from OpenAI, explores long gaps between prime numbers. The paper is available as a PDF on OpenAI's site.
Researchers at Okinawa Institute of Science and Technology developed Spi-Fly, an algorithm inspired by Drosophila's ~140,000-neuron brain, to overcome 'catastrophic forgetting' in electronic noses. Published in Neuromorphic Computing and Engineering, it retains old scent memories while learning new ones.
Google Research evaluated transfer learning from European cohorts to improve polygenic risk score prediction in underrepresented populations, finding it helps small cohorts but degrades accuracy as target sample sizes grow, especially for traits with population-specific genetic architectures.
WeatherNext 3 generates hourly forecasts at up to 5-kilometer resolution, roughly five times sharper than WeatherNext 2, and is now in Search, Gemini, Maps, and Cloud. It ranks as the best global weather model on Brightband's live leaderboard.
Nobel laureate David Baker launches a new accelerator using AI to map nature's design rules, part of a $95 million AI biology effort in Seattle.
Recorded at ICM 2026 in Philadelphia, hosts Janna Levin and Steven Strogatz discuss AI-generated proofs with mathematicians Akshay Venkatesh, Ravi Vakil, and Alex Kontorovich. They examine what recent AI proofs demonstrate and what changes as mathematics becomes more machine-assisted.
Yale physicist Sarah Demers discusses using AI to tune the Mu2e experiment at Fermilab, calling it "a godsend," but refuses to use LLMs for writing due to intellectual property concerns. She leads an APS effort to craft an AI policy statement.
Tao argues that pre-AI open problems are a limited supply of 'uncontaminated' benchmarks, and once solved, it's hard to tell if AI independently solved them. He wants labs to compete on new insights instead.
A paper proposes open-vocabulary mutual information as a unified metric to compare speech brain-computer interfaces across different vocabularies and conditions, revealing trade-offs between vocabulary coverage and decoding accuracy.
MAPL-EMIT, a deep learning model from Google and NASA JPL, detects 50% more methane plumes than human experts and identified over 23,000 additional plumes globally, including 24 of the 25 largest-emitting landfills. Trained on 3.6 million simulated plumes, it uses NASA's EMIT instrument data.
Musk describes using an electromagnetic linear motor on the moon to accelerate AI satellites into deep space, eliminating the need for traditional rockets.
Antigravity's Teamwork multi-agent framework, paired with Gemini 3.7 Flash, solved seven open problems in mathematics and theoretical computer science, including Knuth's Cycles Conjecture (verified in Lean), and scored 71% on TCSBench. It also produced a cycle-accurate RISC-V CPU simulator and optimizations merged into Eigen and ParlayHash.
A fact-check verdict rates TRUE Eric Schmidt's claim that AI solved the 50-year protein-folding problem. DeepMind's AlphaFold achieved this in 2020, and over 3 million researchers had used it by late 2025.
NVIDIA and Anthropic integrated the BioNeMo Agent Toolkit into Claude Science, enabling agents to call BioNeMo NIM microservices for protein structure prediction. On internal benchmarks, BioNeMo skills raise task correctness from 60% to 100% and roughly double token efficiency.
GPT 5.6 Pro, via DottedCalculator, improved the lower bound for Jacobsthal's function to Y(X) ≫ (1/L3) X L1, beating the previous record by Ford, Green, Konyagin, Maynard, and Tao. The sketch shows a proof of Y(X) ≫ (1/(L2 L3)) X L1 using new GPT ideas.
Tutorial walks through building a batched ensemble weather forecasting workflow using NVIDIA Earth2Studio, including installing components, loading the FCN prognostic model, and retrieving initial conditions from GFS.
An essay by Bruce Schneier and Kasra Rafi argues AI models are not yet as capable as experienced mathematicians, despite recent results like OpenAI's disproof of the unit distance conjecture and Anthropic's attempt at the Riemann hypothesis.
Terminal-Bench-Science is a new benchmark for evaluating AI agents on scientific research workflows. It was announced on August 28, 2026, and has gained attention on Hacker News.
Anthropic is offering 10,000 scientists free standard Claude Team seats for one year, with premium seats at $15/month (80% discount). The program expands beyond biology to fields like math, chemistry, and physics, including compute-heavy research.
Google's blog post explains how deep learning with Keras is applied to astroparticle physics, analyzing data from observatories to detect rare cosmic signals. It highlights the role of deep learning in improving instrument sensitivity and uncovering hidden patterns.
Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared spec letting AI agents operate microscopes, liquid handlers, and robotic arms in parallel. It cuts hardware integration from weeks or months to hours or minutes, and began as a collaboration with HHMI Janelia Research Campus.
Google Research's planetary prediction engine (PPE) autonomously executes the full geospatial modeling workflow from data discovery to model training, improving prediction tasks in public health, food security, environmental risk, and socioeconomics. It works directly from natural-language queries.
FrontierMath marked the elliptic curve rank problem as solved after Claude (internal Anthropic model) posted a curve of rank at least 30 on August 20, beating the previous record of 29. A rank-31 curve followed on August 23, credited to Claude with Levent Alpöge and Ava Howell.
AutoDiscovery uncovered a stronger immune signature in invasive lobular breast cancer, validated across an independent dataset and lab analysis. The finding suggests ~15% of US breast cancer patients could benefit from immunotherapy. The partnership includes a local deployment to keep clinical data secure.
New arXiv paper introduces MathAdv, a benchmark for formal theorem proving that goes beyond aggregate proof accuracy to assess reasoning, formalization, and generalization. It addresses limitations of existing benchmarks that focus on narrow math and limited robustness evidence.
The ISCSLP 2026 Real-World Audio-Visual Speech Enhancement Challenge is now open, targeting recovery of target speech from noisy or overlapping audio using visual cues. It addresses limitations of protocols that mix separately recorded sources.