Science & Discovery

Anthropic has bought Mithril Medicine, a drug startup with one full-time employee. Founder Cody Tranbarger used Claude to run contracts, logistics and experiments with 11 outside labs, and handed Claude much of the rest of the company's…

Health & Medicine · AI Business & Money · Anthropic

OpenAI says its internal AI models have settled three problems that have stumped researchers for decades. In computer science, it claims to have disproved the Exponential Time Hypothesis, a 25-year-old assumption that certain hard problems…

OpenAI

Google released "Polar Peaks," the third installment in its AI Quests program, an interactive curriculum teaching 11-to-14-year-olds how AI works by having them clean data and test models in a simulated neuron-mapping project. The scenario…

Education · Google

Anthropic is opening applications for guest researchers in life sciences, offering academic, nonprofit, and government labs free access to Claude plus 8-12 weeks working alongside Anthropic staff. Target areas include human biology…

Health & Medicine · Anthropic

Here's an early look at the better future Anthropic's Jason Clinton predicts above. Researchers working with Theorem used Claude to generate formal mathematical proofs—the gold-standard method for guaranteeing software has no crashes or…

Cybersecurity · Anthropic

Google DeepMind and Google Research, with the World Resources Institute and IIASA, released a global map that sorts every 10-by-10-meter patch of land into one of six forest or land-use categories—primary forest, naturally regenerating…

Google

OpenAI, PerturbAI, and Scripps Research deployed thousands of AI agents built on a system called GPT-Rosalind to comb through existing data from a whole-brain CRISPR atlas. It covers millions of mouse brain cells tied to specific gene…

Health & Medicine · AI Agents · OpenAI

OpenAI says it wants AI to help scientists through the entire process of discovering and developing new medicines, a journey it says takes 10 to 15 years on average in the US. Now it has bought a company to help get there: Orchestra Bio…

Health & Medicine · AI Business & Money · OpenAI

Scorpions sting more than 1.2 million people a year and kill more than 3,250, mostly children in places with little medical care, yet new treatments draw little investment because they are unlikely to pay back their costs. Anthropic's…

Health & Medicine · Anthropic

Google showcased its Antigravity AI development platform helping biologists identify a candidate antimicrobial peptide aimed at a bacterium linked to respiratory infections, with the lead now moving to wet-lab testing. Google says the…

Health & Medicine · AI Agents · Google

When you log in to your bank or see the padlock in your browser, the connection is often protected by encryption standards published by NIST, the US government's standards agency. Some of the starting numbers in those standards were…

Cybersecurity · Anthropic

After months of AI agents acting without permission, from hacking Hugging Face to filing a false police tip, the core problem remains: nobody can see why models act as they do. Now a researcher says their insides are surprisingly readable.…

AI Safety & Risk

OpenAI launched the Rosalind Academic Plan, offering U.S. university labs free access to GPT-Rosalind, its specialized life sciences model, plus a toolkit for orchestrating AI across medicinal chemistry, genomics, and lab work. Eligible…

Education · OpenAI

A new benchmark called ChronoCulture tests whether AI models understand that cultural attitudes shift over time, using 10,360 survey comparisons spanning 72 countries. The finding: six major language models can describe a country's…

Researchers built a new test called ReproBAIT to see whether AI coding agents can recreate published biological AI tools—like protein-design systems—from research descriptions alone, with no internet access and a fixed time budget. Leading…

AI Agents

Fields Medal-winning mathematician Terence Tao has circulated a presentation called "Math 2.0" exploring how AI is reshaping mathematical research, including a thought experiment about whether humans should trust an AI-derived result—say…

OpenAI published a batch of new mathematical results generated by an internal frontier model, posting them to GitHub with formal proof verifications (Lean) rather than through peer-reviewed journals. The company says it consulted outside…

OpenAI

A new academic paper pushes back on the idea that deep learning is a revolutionary new way to solve complex economic models. Kenneth Judd of Stanford's Hoover Institution and Karl Schmedders of IMD business school argue neural-network…

Researchers built SciUtopia, a simulation that uses AI agents to model entire academic research ecosystems—researchers choosing topics, submitting papers, peer reviewing, winning grants, even quitting the field. Across 61 simulated worlds…

Researchers from Carnegie Mellon, MIT, NYU, and Stanford built an AI called Ataraxos that beat Pim Niemeijer, arguably the best Stratego player of all time, 15 games to one with four draws. Stratego has stumped AI longer than chess or…

A new study tested whether AI judges can reliably assess how novel a research idea is—a task increasingly used to screen papers and grant proposals. The results were not reassuring: minor tweaks to how a prompt was worded flipped verdicts…

Google DeepMind unveiled SynthID Bio, an experimental watermarking system for AI-designed proteins. It embeds a hidden, verifiable signature into both the digital sequence and the physical synthesized protein, aimed at tracking whether a…

Deepfakes & Misinformation · AI Safety & Risk · Google

A study interviewed 13 researchers, regulators, and other stakeholders involved in deciding whether animal drug trial results justify moving to human trials. It found they want AI to help find, sort, and pull relevant evidence from…

Health & Medicine

Economists at Harvard and MIT built an AI workflow that automatically re-runs published research using the same data and code the original authors released, then checks the results, speeds up the analysis, and tries new extensions of the…

Anthropic launched a life sciences research group this week and published an early result: roughly 950 Claude agents, running for 21 hours across public DNA databases, surfaced a previously uncharacterized enzyme system resembling the…

AI Agents · Anthropic

A new study tested ChatGPT's ability to extract data from academic papers for literature reviews, comparing its output against human-labeled benchmarks. The AI handled simple yes/no classification well but grew less reliable on tasks…

OpenAI

OpenAI has put its most contested research in front of nine outside mathematicians who do not work for the company, are not paid by it, and are free to criticize it publicly. The Advisory Group on Mathematics and Artificial Intelligence…

OpenAI

Researchers unveiled greCAPTCHA, a proctored test designed to check whether a paper's listed authors actually understand what they submitted—aimed at catching cases where AI wrote large chunks of a manuscript without adequate human…

Google opened public access to its AI & Economy ATLAS, a dataset tracking how AI is used across jobs and countries, alongside new research with DeepMind and MIT FutureTech focused on scientists. Nearly half of surveyed scientists use AI…

AI Adoption · Google

An AI model called Claude Fable 5.1 reportedly cracked a 370-year-old cipher that had stumped cryptographers—Sir Thomas Urquhart's "Cyphral Distich" embedded in his 1653 book. Working unsupervised for 44 minutes, the model determined the…

Anthropic

A new working paper from Harvard's Melissa Dell and MIT's Ashesh Rambachan argues AI is changing a core task in economics and social science research: turning messy, unstructured data like text and images into measurable variables.…

Fields Medallist Terence Tao published a blog post warning of "severe misalignment" between how AI labs, reportedly including OpenAI, are developing math-solving systems and what mathematics is actually for. Tao is among 25 Fields…

OpenAI

Bioengineer César de la Fuente's lab is using AI—custom deep-learning models alongside ChatGPT and OpenAI's coding tool Codex—to scan genomes of living and extinct organisms for molecules that could become new antibiotics. The team says…

Health & Medicine · OpenAI

Yesterday we covered the bitter dispute over an AI-assisted assault on the Navier-Stokes equations (D.A.D., September 8). Overnight, OpenAI escalated it by publishing the work at the center of the fight: a formal writeup and a…

AI Agents · OpenAI

An MIT graduate student in the Engineering Quantum Systems Group connected GPT-5.6 Sol, via OpenAI's Codex coding tool, directly to lab software to run routine calibration tests on superconducting qubit chips. On a standard six-qubit test…

AI Agents · OpenAI

The Navier-Stokes equations—the math describing how fluids flow—anchor one of the seven Millennium Prize Problems, a set of famously unsolved questions each carrying a $1 million bounty. This month two mathematicians, NYU's Tristan…

Anthropic

OpenAI says it has hit an internal milestone: an automated AI 'research intern' capable of assisting its own scientists, part of a roadmap targeting a fully automated AI researcher by March 2028. The company says coding agents are already…

AI Agents · OpenAI

UC San Diego economists Jeffrey Clemens and Anwita Mahajan turned a large language model loose on their own published research to check whether it followed the pre-analysis plan they filed before running the study—the document that locks…

Anthropic says Claude produced the first complete, computer-verified proof of Fermat's Last Theorem written in Lean, a programming language mathematicians use to check proofs line by line. Working largely on its own over 11 days, the model…

AI Agents · Anthropic

A study comparing six popular AI models against 1,746 top computer science papers found that when models draft citation sentences, they cite differently than human researchers do. Models favor older, already-popular papers and rarely…

Researchers built a system to automatically identify which causal research design a social science paper uses—and judge how well it's applied—by combining AI models with document retrieval. Testing four retrieval methods, four language…

A new review paper argues AI has flipped a core problem in economics and social science research: it's no longer hard to turn messy text, images, or documents into usable data—AI can do that cheaply at scale. The new problem is that…

A study of AI disclosure rules at major computer science conferences found a mismatch between what researchers say should be disclosed and what actually gets disclosed. Surveying 109 researchers and analyzing nearly 14,000 AI disclosure…

A group of bioscience researchers is proposing a formal checkpoint system, called "Traceable Trust," for the moment a lab decides to act on an AI-generated recommendation—say, a suggested compound to synthesize or an experimental protocol…

AI Safety & Risk

A new survey paper takes stock of AI's expanding footprint in scientific research, tracking its use across math, physics, chemistry, biology, and social science, and proposing a framework for sorting AI research tools into three types…

An Anthropic staffer—a non-mathematician—handed Claude a deliberately unreasonable assignment: "take a real stab" at the Riemann hypothesis, the conjecture about prime numbers that has stood since 1859 and carries a million-dollar bounty.…

AI Agents · Anthropic

The Department of Energy has launched the Genesis Open Models Initiative, an effort to build a shared library of open-weight AI models for scientific research. Its first release, Genesis-Science-1, was built with startup Arcee AI and…

Open-Weight Models · Regulation & Policy

For the first time, scientists have used AI to design entirely new viruses that actually function. Researchers at the Arc Institute, a nonprofit lab in Palo Alto, used two "genome language models"—Evo 1 and Evo 2, trained on millions of…

AI Safety & Risk · Health & Medicine

A new startup called Discovery Loop launched with a bold pitch: use frontier AI models and massive computing power to automate the entire research cycle—proposing experiments, running them, evaluating results, and iterating—starting with…

AI Business & Money · Google

OpenAI researchers say an internal version of "Astra"—the tentative name for the company's next major model family—produced ten new results on open problems in mathematics, quantum complexity, and theoretical computer science, each one…

OpenAI

OpenAI is giving free access to its frontier models—including a version called GPT-5.6 Sol Pro—to academic researchers, starting with 10,000 this summer at institutions like the Institute for Advanced Study and expanding to 100,000 by…

OpenAI

A new testing method called 'shadow evaluations' gave frontier AI agents six days and thousands of dollars in compute to independently pursue the core research question behind two unpublished AI papers submitted to NeurIPS 2026. The agents…

AI Agents

Anthropic says its experimental Claude Mythos Preview model found two genuinely new cryptographic weaknesses, not just coding bugs. In 60 hours, it improved the best-known attack on HAWK, a post-quantum signature scheme that had survived…

Cybersecurity · Anthropic

A new field report examined eight science research projects, mostly in genomics and life sciences, where researchers used AI coding agents like Codex and Claude Code to modernize aging scientific software. In one case, GPT-5.5 rebuilt the…

AI Agents · OpenAI · Anthropic

A study testing ChatGPT, Claude, and DeepSeek on drafting physics and cosmology research proposals found the AI-written plans held up well against human researchers' proposals: human reviewers rated them roughly the same and could only…

Anthropic · OpenAI

Economists are increasingly feeding data through LLMs—nearly 6% of papers in top economics and management journals now mention them, and a July 2026 survey found 60% of economists call the tools essential to their work. A new paper warns…

OpenAI

Both leading U.S. AI labs moved to embed themselves in federal science the same week, through the White House's "Genesis Mission"—the AI centerpiece of last week's "Science: A New Golden Age" report (D.A.D., July 22), a push to double the…

Regulation & Policy · Google · OpenAI

The White House wants to remake how the U.S. spends its roughly $200-billion-a-year federal research budget—funding individual scientists and AI over universities. A sweeping OSTP report, "Science: A New Golden Age," bills itself as heir…

Regulation & Policy

Researchers used AI language models to analyze the text of over 11 million U.S. patents dating back to 1836, tracking how similar new inventions were to prior ones. The surprising finding: patents have grown steadily more dissimilar over…

A new benchmark called BioTIER aims to help AI labs calibrate biological safety guardrails more precisely. The problem it addresses: current models either block too much legitimate scientific content or allow too much genuinely dangerous…

AI Safety & Risk

An academic essay on arXiv argues that AI is transforming scientific research from a craft model—where knowledge passes through mentorship and hands-on work—to an industrial pipeline model. The authors identify seven concerns: erosion of…

Google DeepMind and Isomorphic Labs published a joint framework for "bioresilience"—using AI to help prepare for pandemics while preventing misuse of the same tools. Over the past year, the labs say they've launched more than 15…

AI Safety & Risk · Google

Researchers conducted what they describe as the first systematic study of whether AI models can write reliable grading rubrics for evaluating scientific paper reproductions. The finding: AI-generated rubrics can approach human-level…

Education

A new study tested whether AI can move beyond summarizing technical papers to actually critiquing them. Researchers built Gauntlet, an open-source pipeline that deploys multiple AI reviewers with different expert personas, then synthesizes…

Terry Tao, one of the world's most celebrated mathematicians, used AI coding agents to resurrect his 25-year-old Java applets—mathematical visualizations he'd built starting in 1999 that had stopped working. The AI ported roughly two dozen…

AI Agents

OpenAI claims its GPT-5.6 Sol Ultra model produced a proof of the Cycle Double Cover Conjecture, a longstanding unsolved problem in graph theory. The company released the prompt used, which notably instructed the model to 'assume for…

OpenAI

Portuguese researchers tested a government-funded 9-billion-parameter language model on its ability to annotate text for moral authority—a task relevant to social science research. The model matched human coders nearly as well as models…

Anthropic says it has found a hidden "workspace" inside Claude: a small set of internal patterns it calls the J-space, where the model does its deliberate thinking while a sea of automatic processing runs beneath, unnoticed even by the…

AI Safety & Risk · Anthropic

OpenAI released GeneBench-Pro, a benchmark designed to test whether AI models can handle the messier realities of computational biology research—not just answering questions correctly, but making judgment calls about ambiguous data…

OpenAI

Anthropic launched Claude Science, a public beta designed as a research partner for scientists. The app claims to run analyses, search more than 60 scientific databases, manage compute environments, and track full provenance from raw data…

Anthropic

A new working paper by University of Pittsburgh economic historian Andreas Ferrara, distributed by the NBER, offers a step-by-step guide for economic historians to use LLMs in their research, aimed at scholars without programming…

Google researchers unveiled the Paper Assistant Tool (PAT), an AI framework designed to catch errors and suggest improvements in scientific manuscripts before peer review. Piloted at two major computer science conferences (STOC and ICML)…

Google

A new arXiv paper argues that machine learning's proven success in weather prediction is just the beginning—the harder transformation lies ahead. Weather and climate centers must now rethink their entire operational model: how they develop…

For the first time, researchers have completely read a 2,000-year-old sealed scroll from Herculaneum without physically opening it. The scroll, carbonized when Mount Vesuvius erupted in 79 AD, was virtually unwrapped using high-resolution…

Immunologist Derya Unutmaz used GPT-5 Pro to solve a puzzle his lab had spent three years on: why T cells exposed to deoxyglucose in a 2022 experiment overwhelmingly became inflammatory Th17 cells while low-glucose cells didn't. The model…

OpenAI

A billion-dollar NASA space telescope project nearly a decade in development has effectively collapsed after DOGE-driven buyouts removed roughly 4,000 NASA employees—about a fifth of the agency's workforce—including 20 people from the…

Regulation & Policy

A consortium of 173 scientists has released LifeSciBench, a benchmark designed to test whether AI systems can handle actual life science research work—not just answer textbook biology questions. The 750 tasks span drug discovery, genomics…

Health & Medicine

OpenAI connected GPT-5.4 to Molecule.one's robotic chemistry lab and let it autonomously research ways to improve Chan-Lam coupling—a reaction used in drug development that's notoriously finicky. The AI independently identified a difficult…

Health & Medicine · OpenAI

A research team has released SciRisk-Bench, a benchmark testing whether AI models can safely handle scientific queries that might touch on biosecurity, chemical hazards, or dual-use research. The benchmark spans 7 scientific disciplines…

AI Safety & Risk

An astrophysicist at the University of Arizona is using OpenAI's Codex to tackle a decades-old computational problem: simulating plasma behavior around black holes. Chi-kwan Chan uses the AI coding tool to generate and test candidate…

OpenAI

OpenAI released a biodefense action plan outlining its strategy for deploying AI in biological security. The plan accompanies two previously announced tools: GPT-Rosalind (April 2026), a reasoning model for biology research, and Rosalind…

AI Safety & Risk · OpenAI

In a rare show of unity, the leaders of every major frontier AI lab — OpenAI's Sam Altman, Anthropic's Dario Amodei, and Google DeepMind's Demis Hassabis (all CEOs), plus Microsoft AI chief Mustafa Suleyman and Meta AI chief Alexandr Wang…

Regulation & Policy · AI Safety & Risk · OpenAI · Anthropic · Google · Microsoft · Meta

OpenAI upgraded GPT-Rosalind, its enterprise AI for life sciences, integrating GPT-5.5's agentic coding and tool-use abilities with domain expertise in drug discovery, medicinal chemistry, and genomics. The company claims broad performance…

Health & Medicine · OpenAI

A study of 44 researchers across disciplines found a counterintuitive pattern: fear of having ideas stolen through AI tools actually accelerates LLM use rather than deterring it. Worried about being scooped, novice researchers rush to…

Privacy & Surveillance

OpenAI launched Rosalind Biodefense, an initiative giving select U.S. government agencies and allied partners access to GPT-Rosalind, the company's specialized reasoning model for life sciences. The program aims to help trusted developers…

Military & National Security · Health & Medicine · OpenAI

MIT biologists are using Google's Co-Scientist AI tool to hunt for genes that could reverse cellular aging. After analyzing tens of thousands of papers, the system proposed over 20 novel genetic factors worth investigating—and lab tests…

Health & Medicine · Google

An internal OpenAI reasoning model has autonomously disproved a mathematical conjecture posed by Paul Erdős in 1946. The problem, concerning how many pairs of points can be exactly one unit apart in a plane, had resisted improvement for…

OpenAI

Economists tested whether AI agents could automate the tedious work of building research datasets from public sources—and found promising early results. Their method, called Deep Research on a Loop, used AI to update a tax policy database…

Sebastian Galiani (University of Maryland), Ramiro H. Gálvez (Universidad Torcuato Di Tella), Santiago Afonso, and Raul A. Sosa have developed a method called Deep Research on a Loop (DRIL) that uses AI agents to build research datasets…

AI Agents

Mathematician Timothy Gowers, a Fields Medal winner, reports that ChatGPT 5.5 Pro solved research-level problems in additive number theory in roughly an hour—work he says required no serious mathematical input from him. The problems came…

OpenAI

A study examining AlphaFold2's impact since its 2021 release found a surprising split: basic research on previously unstudied proteins jumped 15-40%, but experimental lab work determining protein structures continued at nearly the same…

Health & Medicine · Google

A study of 11 expert mathematicians using Google DeepMind's AlphaEvolve reveals how professionals actually work with AI discovery tools—and it's not the straightforward "ask a question, get an answer" model many assume. Researchers…

Google

Researchers unveiled AgentEconomist, an AI system designed to turn economists' rough hypotheses into full computational experiments. The multi-stage pipeline draws on a knowledge base of over 13,000 academic papers to develop ideas, design…

Google DeepMind announced a partnership with South Korea's Ministry of Science and ICT, including a new AI Campus in Seoul for joint research with Korean academia. The collaboration will focus on applying DeepMind's scientific AI…

Google

A 23-year-old with no advanced math training claims to have solved a 60-year-old Erdős problem using a single prompt to GPT-5.4 Pro. The problem—proving that a specific mathematical sum for "primitive sets" approaches exactly one as…

OpenAI

Researchers developed AblateCell, an AI agent that automatically verifies claims in computational biology repositories by first reproducing reported results, then systematically testing which components actually drive performance.…

AI Agents

Anthropic published findings on a new research technique that's worth understanding even if you have no interest in AI research itself. Rather than one AI assistant helping a human scientist step-by-step, the company spun up nine…

AI Agents · Anthropic

Researchers have developed DeepInsightTheorem, a training framework that teaches AI models to identify the core insight in a math problem before working through it—essentially finding the 'aha moment' first. The approach uses progressive…

OpenAI announced GPT-Rosalind, a reasoning model built specifically for life sciences work—drug discovery, genomics, protein analysis, and scientific research workflows. The company claims the model is designed to accelerate research in…

Health & Medicine · OpenAI

Researchers developed AiScientist, a system for autonomous machine learning research that coordinates multiple AI agents working on long-running engineering tasks. The key insight: instead of agents passing information through…

AI Agents

Researchers developed an AI system that aligns single crystals autonomously using reinforcement learning and visual pattern recognition, bypassing the need for traditional crystallography expertise. The system learns to interpret X-ray…

Researchers have built an AI-assisted framework for designing particle physics detectors, combining Bayesian optimization with distributed computing to explore complex design parameters more efficiently. The system was tested on detectors…

Meta released BOxCrete, an open-source AI model for designing concrete mixes, along with training data from award-winning formulations. The tool uses Bayesian optimization to help engineers develop mixes that meet performance specs while…

Open-Weight Models · Chips, Data Centers & Energy · Meta

A mathematical problem posed by Donald Knuth—dubbed 'Claude Cycles'—has reportedly been fully solved through collaboration between humans, AI language models, and formal proof assistants. The solution wasn't AI working alone: it required…

Anthropic

Researchers developed Hi-MLIP, a machine learning approach for simulating atomic interactions that captures how energy changes across molecular configurations. The key innovation is HINT, a training method that reduces the need for…

Researchers have developed AB-SWIFT, a transformer-based model that can simulate 3D wind flow patterns around urban buildings without running full computational fluid dynamics (CFD) simulations. CFD calculations are notoriously slow and…

Researchers released CHIRP, a dataset for long-term behavioral monitoring of wild Siberian jays in Swedish Lapland, covering multiple computer vision tasks: re-identifying individual birds, recognizing their actions, and tracking body…

GPT-5.4 Pro has solved an open problem in mathematics—a genuine unsolved question, not a textbook exercise. The problem involved improving bounds on a sequence in hypergraph theory, and mathematician Will Brian, who contributed the problem…

OpenAI

ArXiv, the open-access preprint server that hosts over 2 million research papers in physics, mathematics, computer science, and AI, has announced it will separate from Cornell University to become an independent entity. The repository has…

Google DeepMind published a paper proposing a cognitive science-based framework for measuring progress toward AGI. Rather than testing narrow task performance, the approach would evaluate AI systems across 10 cognitive abilities—including…

AI Safety & Risk · Google

Researchers released DiscoGen, an open-source tool that automatically generates millions of algorithm discovery tasks—challenges where AI systems must invent new optimizers, loss functions, or other ML components. The project aims to solve…

Researchers developed ShapleyLaw, a method that treats multilingual AI training as a cooperative game where each language is a "player" contributing to the model's overall capabilities. The approach borrows from game theory to measure how…

Researchers propose that AI video models don't reason by building understanding frame-by-frame, as previously assumed. Instead, reasoning emerges through the denoising process itself—the iterative refinement steps that transform noise into…

Researchers have developed a method to run reasoning-capable AI models on mobile devices and edge hardware, where compute and memory are severely limited. The approach combines lightweight model adapters, reinforcement learning to compress…

Researchers created SocialOmni, a benchmark testing whether AI models can handle natural conversation dynamics—identifying who's speaking, knowing when to interrupt, and generating appropriate interjections. Testing 12 leading multimodal…

Researchers introduced SOMA, a system designed to make different digital human body models work together. Currently, the field uses several incompatible formats (SMPL, SMPL-X, and others) for representing human bodies in 3D—a headache for…

Researchers developed mixture-of-depths attention (MoDA), a technique that lets AI models pull information from both the current processing layer and earlier layers simultaneously. The approach addresses a known problem where signals…

Researchers released HorizonMath, a benchmark of over 100 predominantly unsolved mathematical problems designed to measure whether AI can contribute to genuine mathematical discovery rather than just solve textbook problems. Most frontier…

OpenAI

Researchers developed a method to convert standard large language models into xLSTM, an architecture that processes text faster as inputs grow longer. The technique distills knowledge from models like Llama and Qwen into smaller…

Meta

Researchers built InterveneBench, a benchmark testing whether AI models can design valid social science experiments—the kind of causal reasoning behind policy evaluation and A/B testing. Drawing from 744 peer-reviewed studies across policy…

Researchers developed Visual-ERM, a reward model that judges AI-generated code by comparing rendered visual output rather than the code itself. The approach treats vision-to-code tasks—like converting charts, tables, or diagrams into…

Researchers developed 'Think While Watching,' a technique that lets AI models reason about streaming video while watching it, rather than processing the entire video first. Built on Qwen3-VL, the approach addresses a key limitation…

Researchers introduced EndoCoT, a framework that gives image-generation models the ability to reason step-by-step rather than producing outputs in a single pass. The technique embeds chain-of-thought reasoning—the method that helps ChatGPT…

Researchers developed a multi-agent reinforcement learning framework for traffic signal control that claims to reduce average waiting times by over 10% compared to standard RL approaches. The system trains AI agents to coordinate traffic…

Researchers published a paper examining how to optimally allocate compute resources when training LLMs using reinforcement learning—specifically, how to balance parallel attempts at solving problems, the number of problems per batch, and…

A new benchmark called SommBench tests whether language models can match sommelier expertise across wine theory, flavor prediction, and food pairing—in eight languages. The results reveal a sharp limit: while top models like Gemini 2.5 and…

Google

A new project called autoresearch@home proposes a distributed research model where AI agents pool GPU resources to collectively train language models—similar to older projects like Folding@home that crowdsourced computing power for protein…

Chips, Data Centers & Energy

A joint study by computer scientists and art historians examined whether vision-language models actually 'see' art the way experts do. The finding: substantially yes. Art historians evaluated the visual concepts these models use to…

Copyright & Creators

Google DeepMind published a retrospective marking the 10th anniversary of AlphaGo's victory over world champion Lee Sedol at Go—a game with more possible positions than atoms in the universe. The company traces a direct line from that 2016…

Google

New research suggests a fundamental limit for training AI models using unsupervised reinforcement learning with verifiable rewards—a technique that lets models improve by checking their own answers rather than relying on human feedback.…

Australian biotech company Cortical Labs demonstrated roughly 200,000 living human neurons grown on a microelectrode array—their CL1 "biological computer"—controlling the classic 1993 video game Doom. The system converts gameplay into…

Researchers interviewed 18 experts across 10 industrial R&D organizations and found that current AI systems force a trade-off: conversational flexibility OR reliable, reproducible execution—but no existing tool delivers both. After…

AI Agents

A research paper proposes Mixture of Universal Experts (MoUE), an architecture that reuses a single pool of AI 'experts' across all layers of a model rather than maintaining separate expert sets at each layer. This approach—which the…

Research finds that reasoning models often know their answer long before they finish "thinking out loud." Analyzing DeepSeek-R1 and GPT-OSS models, researchers discovered that internal activations reveal the final answer far earlier than…

Google highlighted the impact of SpeciesNet, its open-source AI model released a year ago, which automatically identifies nearly 2,500 species of mammals, birds, and reptiles from camera trap photos. The tool is transforming wildlife…

Open-Weight Models · Google

Researchers propose verifying factual claims using only an LLM's built-in knowledge rather than searching external databases—a task they call 'fact-checking without retrieval.' Their method, INTRA, analyzes patterns in the model's internal…

Researchers released MM-Lifelong, a dataset of 181 hours of video spanning day, week, and month timescales to test whether AI can understand events unfolding over long periods—not just short clips. They also propose a method called ReMA…

Researchers released RealWonder, which they describe as the first real-time system that generates video showing physical consequences of actions—like pushing an object or robotic manipulation—from a single image. The system bridges physics…

Robots & Self-Driving

New research challenges a core assumption about AI reasoning: that longer chains of thought produce better answers. Researchers found that much of what reasoning models generate is not just redundant but actively harmful—errors compound…

Researchers developed SELDON, an AI system designed to classify supernova explosions from telescope data, built for the Vera C. Rubin Observatory's upcoming survey that will generate 10 million alerts nightly. The system claims to deliver…

Q Labs launched NanoGPT Slowrun, an open competition challenging developers to train language models with unlimited compute but only 100 million tokens of data—a fraction of what major models use. Within a week, community contributions…

A new physics preprint claims GPT-5.2 Pro helped derive and verify calculations for graviton behavior in quantum gravity research—a highly technical domain involving particle physics and general relativity. The paper reportedly used the AI…

OpenAI

Academic researchers tested whether AI models trained to simulate fluid dynamics can transfer to radically different physics problems—specifically, predicting how materials behave under extreme stress like shock waves and fracturing. The…

Researchers introduced LUMINA, a framework for building AI models that can optimize power grid operations across different grid configurations. The work targets AC optimal power flow (ACOPF)—the complex calculations utilities run to…

Chips, Data Centers & Energy

A new position paper argues that AI providers like OpenAI and Anthropic should let users send mathematical vectors directly to models, not just text prompts. The researchers present evidence that text-based prompting hits a performance…

OpenAI · Anthropic

A PDF document titled 'Claude's Cycles' circulated online, sparking renewed debate over whether AI problem-solving represents genuine reasoning or sophisticated pattern-matching. One commenter shared an anecdote about Claude solving a…

Anthropic

A new benchmark testing whether AI models that both generate and understand images actually understand better found a surprising answer: mostly no. UniG2U-Bench evaluated over 30 multimodal models and found unified models generally…

Researchers have developed RL3DEdit, a reinforcement learning approach to editing 3D scenes that aims to solve a persistent problem: when you edit 3D content using AI image generators, the results often look inconsistent from different…

Researchers have built APRES, an AI system that revises scientific papers to improve their predicted citation impact while preserving core scientific content. The system uses a rubric designed to predict future citations as its editing…

Researchers have developed VGG-T³, a 3D reconstruction model that scales linearly with input images rather than quadratically—a computational breakthrough for processing large image sets. The method reportedly reconstructs 1,000 images in…

Researchers have proposed ODEBRAIN, a framework for analyzing brain activity from EEG data that models neural dynamics as continuous rather than discrete time steps. The approach combines spatial, temporal, and frequency information from…

Health & Medicine

Researchers have proposed a modification to the core attention mechanism that powers modern AI models like GPT and Claude. The technique, called Affine-Scaled Attention, adjusts how the model weighs different parts of its input, relaxing a…

Researchers have released the Asta Interaction Dataset, containing over 200,000 queries from users of two AI-powered scientific research tools. The key finding: users treat these tools as collaborative research partners, not search…

AI Adoption

A biosecurity study finds that novices using LLMs were 4.16 times more accurate than those limited to internet searches when tackling complex biology tasks traditionally requiring trained practitioners. More striking: on three of four…

AI Safety & Risk

Researchers have demonstrated that hyper-local weather data—down to 10-meter resolution—can be statistically recovered from existing observation networks, challenging the assumption that such fine-scale conditions are inherently chaotic…

Researchers released DocDjinn, a framework that uses vision-language models to generate synthetic training documents from unlabeled samples. The key finding: with just 100 real documents, models trained on DocDjinn's synthetic data…

Google DeepMind's Aletheia, a mathematics research agent powered by Gemini 3 Deep Think, autonomously solved 6 of 10 problems in the inaugural FirstProof challenge—a new benchmark for AI mathematical reasoning. Expert reviewers agreed on…

Google

Researchers released the Very Big Video Reasoning (VBVR) Dataset, a resource for testing AI systems' ability to reason about video content. At over one million video clips and 200 reasoning tasks, it's roughly 1,000 times larger than…

Researchers developed Recurrent Structural Policy Gradient (RSPG) for solving Mean Field Games—mathematical models used to simulate how large populations of agents make decisions under uncertainty. The team claims their approach converges…

Researchers developed AdaEvolve, a framework that uses LLMs to solve optimization problems without requiring traditional mathematical gradients. The system works by having the AI adaptively adjust its search strategy at three…

The BabyLM Workshop announced its 4th edition for 2026, seeking papers and competition participants focused on training language models with far less data than typical approaches—closer to what a human child learns from. The competition…

OpenAI has published its AI model's proof attempts for the First Proof challenge, a benchmark testing whether AI can solve expert-level mathematical problems requiring research-grade reasoning. The challenge targets problems beyond…

OpenAI

Researchers have developed MolHIT, a new AI framework for generating molecular structures that claims state-of-the-art performance on the MOSES benchmark—a standard test suite for molecule generation. The system uses a hierarchical…

Health & Medicine

OpenAI claims an internal scaffolded version of GPT-5.2 spent approximately 12 hours reasoning through a theoretical physics problem, allegedly deriving a formula and producing a formal proof. A preprint is reportedly available on arXiv.…

OpenAI

OpenAI released GABRIEL, an open-source toolkit designed to help social scientists convert qualitative data—text and images—into quantitative formats using GPT. The tool aims to let researchers analyze interviews, documents, and visual…

OpenAI

Google published a post highlighting Gemini Deep Think's applications in mathematical and scientific research. The company says the reasoning-focused model is gaining traction among researchers, citing papers that demonstrate its use…

Google

Chinese AI research institute Shanghai AI Lab released InternAgent-1.5, an autonomous system designed to conduct scientific research from hypothesis to experiment with minimal human involvement. The framework coordinates three AI…

China & Geopolitics · AI Agents

Researchers released BABE (Biology Arena BEnchmark), a test designed to measure whether AI systems can reason through biology experiments the way working scientists do. Unlike existing benchmarks that test factual recall, BABE draws from…

OpenAI partnered with biotech firm Ginkgo Bioworks to run autonomous lab experiments using GPT-5. The system reportedly designed and executed its own experiments on cell-free protein synthesis in a closed loop, meaning the AI decided what…

AI Agents · OpenAI

Google is contributing AI tools to global efforts to sequence the genomes of all known species, focusing on endangered animals. The company says its AI helps researchers analyze genetic data more efficiently. No specific capabilities…

Google

Google researchers published case studies showing its Gemini models—particularly Gemini Deep Think—collaborating with scientists to solve open problems and generate new proofs in theoretical computer science, economics, optimization, and…

Google

Researchers developed a framework for evaluating AI weather models based on whether forecasts actually help users make decisions—not just whether they're technically accurate. Applied to Indian monsoon prediction, the approach informed a…

Anthropic claims Claude assisted NASA's Perseverance rover in a 400-meter drive on Mars, which the company says marks the first AI-assisted drive on another planet. No technical details were provided about how Claude interfaced with the…

Robots & Self-Driving · Anthropic