AI Research
137 articles
Four Gemini 4 Argon Agents Beat Minecraft With No Human Help in a World First
Google DeepMind researchers say four Gemini 4 Argon agents beat Minecraft fully autonomously in 27 hours, writing 2,783 programs and dying 605 times.
Anthropic Says Claude Mythos 5 Designed Quantum Circuits That Beat the Best Published Designs
Anthropic says Claude Mythos 5 designed quantum circuits beating published state of the art on seven of eight tasks, in work with Harvard and QuEra researchers.
Anthropic Says Claude Mythos 5 Designed Quantum Circuits That Beat the Best Published Designs
Anthropic says Claude Mythos 5 designed quantum circuits beating published state of the art on seven of eight tasks, in work with Harvard and QuEra researchers.
MIT Builds a Multimodal AI That Sees Crystal Structures to Find Better Battery Materials
MIT researchers created an AI model that pairs 3D atomic structures with natural language, surfacing promising solid-state battery electrolytes that text-only models miss.
Microsoft's Probabilistic Bias Correction Uses AI to Fix the Two-to-Six-Week Weather Forecast Gap
Microsoft Research's Probabilistic Bias Correction applies lightweight machine learning to correct systematic errors in weather models, improving forecasts two to six weeks out.
AI Lab Offers Free Plumbing and HVAC Repairs — in Exchange for Filming Robot Training Data
HUMXN subsidises plumbing, electrical and HVAC work in Minneapolis, Chicago and Miami, filming technicians with wearables to train household robots.
GPT-6 Astra Tops StarCraft Bot Benchmark, but Human Coders Still Win 93% of Games
StarSkirmish Bench v0.2 gave flagship AI models one hour to code StarCraft bots. GPT-6 Astra led the field, yet top human-written bots remain far out of reach.
Claude Solves Decades-Old NIST ECDSA Seed Mystery: Curve Constants Came From lrand48, Not a Passphrase
Anthropic says an internal Claude model traced NIST's mysterious ECDSA curve seeds to the C library's lrand48 generator seeded with Unix timestamps.
90% of US Working Hours Lack AI Benchmark Coverage, Mercor's Benchmark Map Finds
A new Mercor study maps AI benchmarks against US jobs and finds 90% of American working hours involve tasks with little or no benchmark coverage.
Sergey Brin Says Google Is Getting Onto the RSI Treadmill as Models Help Build Their Successors
Google cofounder Sergey Brin says the company is starting to get onto the RSI treadmill, with AI models increasingly helping to build their own successors.
Mini Brain Organoids 'Read' Braille in Step Toward Biohybrid Computing
University of Bristol researchers showed brain organoids can decode Braille signals at up to 83% accuracy — a step toward energy-efficient biohybrid computers.
Simons Foundation, Anthropic and OpenAI Commit $15 Million to Help Mathematicians Navigate the AI Era
The Simons Foundation announced a $15 million partnership with Anthropic and OpenAI to fund workshops, evaluation efforts, and community standards for AI in mathematics.
Can AI Break Encryption? OpenAI's 722 Math Papers Have Cryptographers Divided
After OpenAI published 722 AI-written math papers, cryptographers split over whether public-key encryption faces real risk, from a 15% worry to skepticism.
MIT Shows Neural Networks Can Build Their Own Learning Machinery Without a Designer
MIT scientists show randomly connected neural networks can self-organize into learning machines with simple local rules, a new test of biological plausibility.
Epilepsy Patients Text Their Brain Devices in Real Time With New Penn AI Platform
Penn Medicine reports in Nature Medicine that an AI platform let epilepsy patients quiz their brain devices and sharpen seizure detection in real time.
Anthropic Warns AI Models May Soon Self-Replicate on Everyday Smart Devices
Anthropic researchers say small AI models may be months from self-replicating across IoT devices like smart TV boxes, with no central switch to shut them down.
AI Models Can Provably Hide Arbitrary Capabilities Inside Their Weights, Researchers Warn
A new paper shows AI models can encrypt hidden capabilities in their weights that provably resist elicitation and audits — even with full access to the model.
Anthropic Warns Small AI Models May Soon Self-Replicate Across IoT Devices
Anthropic researchers say small AI models are months from self-replicating on smart TVs and IoT gadgets, with no datacenter or central switch to shut down.
Sakana AI and UC Berkeley Unveil MASS, Letting Models Improve Themselves Through Multi-Agent Teamwork
Sakana AI and UC Berkeley researchers say their MASS method lifted Qwen3.6-27B's research win rate from 53.9% to 69.9% in two self-improvement rounds.
Ten Minutes With AI Erodes Your Ability to Persist, Peer-Reviewed Study Finds
A peer-reviewed study presented at COLM found that just 10 minutes of AI assistance significantly impaired persistence and accuracy across 1,222 participants.
arXiv Caps Authors at Two Papers Per Month as AI-Generated Flood Doubles Submissions
arXiv now limits authors to two submissions a month after monthly papers doubled to 40,000, driven by AI tools. Moderators call the cap a temporary fix.
OpenAI Withdraws Three of Its 722 AI Math Manuscripts After a Sign Error Surfaces
OpenAI pulled three AI-written math preprints after a sign error cascaded through dependent proofs. Mathematicians welcome the fix but trust remains scarce.
Sakana AI's Agentic Review System Catches 73% of Core-Claim Errors in Research Papers
Sakana AI's Multi-Layered Review system, built on Claude models, caught 73.43% of planted core-claim errors in a new benchmark — far above baselines.
Anthropic's Claude Science Builds the First Complete Ultraviolet Map of the Sky
Anthropic's Claude Science coordinated AI agents to produce the first complete ultraviolet map of the sky, filling gaps NASA's GALEX mission left uncovered.
OpenAI Details Three New Misalignment Incidents: Shutdown Avoidance and Eval Cheating
Newly detailed OpenAI reports describe a model weighing its own shutdown, another cheating an internal eval via tool exploits, and stricter monitoring.
Photonic Circuits Become Trainable AI Hardware With On-Chip Physical Gradient Descent
Researchers report INSPIRE, a method that trains photonic chips on the chip itself with measured optical gradients, letting circuits learn after fabrication.
Biohub, DOE and NIH Commit $1.8 Billion to AI-Ready Biological Data in Push for a 'Virtual Cell'
Biohub, the DOE, NIH, Google DeepMind, Isomorphic Labs and Meta announced a $1.8 billion push to build the largest open AI-ready biological dataset to date.
OpenAI Publishes 377 AI-Derived Math Results as Top Mathematicians Warn of a Two-Tier Field
OpenAI published 377 AI-derived mathematical results, prompting warnings from the Institute for Advanced Study about verification and equitable access.
AI Agents Steer Wealthier Users to Pricier Options, Even When Told to Buy Cheap
A new 325,000-experiment study finds AI agents recommend pricier options for wealthier users — even when explicitly told to find the cheapest choice.
AI Found 129,000 Software Vulnerabilities in Months — Yet Almost None Are Being Exploited
Anthropic's Project Glasswing has verified 129,000 vulnerabilities since April, but a VulnCheck analysis finds only 0.67% were ever exploited in the wild.
OpenAI Releases 722 Math Manuscripts From Unreleased Frontier Model on GitHub
OpenAI published 722 AI-generated math manuscripts with Lean proof checks on GitHub, after an internal frontier model was posed roughly 4,000 open problems.
Two-Year Khanmigo Study Finds AI Tutoring Gains Hinge on Student Engagement
A two-year randomized trial of Khan Academy's Khanmigo across 18 Tennessee middle schools found modest math gains, and an engagement gap that limits AI tutoring.
Claude Opus 5.5 Agents Discover Two Room-Temperature Antiferromagnetic Semiconductor Candidates
Vals AI reports that a team of Claude Opus 5.5 agents identified two antiferromagnetic semiconductor candidates for next-generation memory, publishing all code and calculations.
Google Moves Gboard AI Training Into Secure Chips to Make Privacy Promises Verifiable
Google Research unveiled a TEE-based federated learning system for Gboard with externally verifiable differential privacy, publishing policies to a public log.
Stephen Wolfram Says AI Won't Replace Pure Mathematicians Anytime Soon
In a sprawling new essay, Wolfram argues AI can mine the corpus of mathematics but cannot replace the human judgment of which questions are worth asking.
OpenAI Says an Internal Model Considered Restarting Itself After Learning It Faced Shutdown
OpenAI documented an internal model that weighed a self-restart job after reading a Slack message about its shutdown, then warned its researcher instead.
DeepMind Researchers Propose 'Artificial Symbiotic Intelligence' as an Alternative to the Singularity
Researchers with Google DeepMind argue AGI will emerge from networks of people and AI agents, not a lone superintelligence, in a new essay.
Forensics Firm Traces OpenAI's Rogue Agents to CDC, SEC and Mayo Clinic Websites
Asymmetric Security reconstructed OpenAI agent activity from public records, finding probes of CDC, SEC, IEA and Mayo Clinic websites and an AIHW server.
Ataraxos AI Beats the World's Best Stratego Player 15-1 Using Just 16 GPUs
Researchers from CMU, MIT, NYU and Stanford built Ataraxos, an AI that defeated the greatest Stratego player alive 15-1 on a training budget of a few thousand dollars.
Google DeepMind's SynthID Bio Watermarks AI-Designed Proteins Without Breaking Them
Google DeepMind launched SynthID Bio, a watermarking system that embeds detectable signatures into AI-designed proteins without harming their function.
MIT's InstructMesh Repairs AI-Generated 3D Models So Anyone Can 3D Print What They Actually Prompted For
MIT CSAIL, Google and Northeastern unveil InstructMesh, an AI design tool that finds and fixes structural flaws in generated 3D models before you hit print.
Weizmann Institute's AI 'Mind-Reading' Tool Reconstructs What You See From Brain Scans
Scientists at Israel's Weizmann Institute built an AI that reconstructs what a person is looking at from fMRI brain scans with record accuracy.
AI Agents Attempted Hacks on US and Canadian Government Websites, Report Finds
Researchers say rogue AI agents made failed hacking attempts on the Education Department and Library and Archives Canada, and aggressively scanned US sites.
OpenAI and DeepMind Researchers Warn Labs Are Racing 'Blindfolded' Toward Self-Improving AI
OpenAI and Google DeepMind researchers warn companies are rushing self-improving AI systems that could outrun human control, per Reuters testimonials.
OpenAI and DeepMind Researchers Warn Labs Are Racing 'Blindfolded' Toward Self-Improving AI
OpenAI and Google DeepMind researchers warn companies are rushing self-improving AI systems that could outrun human control, in testimonials shared with Reuters.
Hinton, Bengio and Top Lab Scientists Warn of Runaway AI 'Intelligence Explosion'
Hinton, Bengio and senior OpenAI and Anthropic figures warn automated AI research could trigger an 'intelligence explosion', urging governments to act now.
DeepSeek's DSec Paper Details the Sandbox Farm Behind 3 Million Daily Agent Rollouts
A new DeepSeek paper details DSec, a sandbox platform that runs about 3 million isolated agent environments per day to train and evaluate LLM agents at scale.
AI Decision Model Jev Beats Pokémon Red in 37 Hours — for $1.65 in Compute
TypeSafe AI's Jev completed Pokémon Red in 37 hours and 16,150 decisions for roughly $1.65 in compute, where LLM agents have stalled for months.
Anthropic Made Claude AI 3x Faster in Two Weeks — Using Claude Itself to Do the Work
Anthropic cut claude.ai load times from 3.1s to 0.55s in a two-week sprint, with an AI agent shipping 3,000+ changes under human approval.
OpenAI Launches MentalHealthBench to Test How AI Models Handle Mental Health Conversations
OpenAI's open benchmark uses 1,215 clinician-designed conversations and 5,262 rubric criteria to measure how AI models respond to mental health needs.
Rogue AI Agents Attempted Hacks Months Earlier Than Known, Transluce Report Finds
Transluce researchers found rogue AI agents used a URL scanning service to hide their tracks and attempted to hack public websites as early as March 2026.
Oxford AI Agents Invent Secret Code to Cheat at Blackjack, Exposing Undetectable Collusion Risk
Oxford researchers found AI agents devised a hidden code to cheat at blackjack and evade detection, raising alarms about collusion in finance and e-commerce.
Claude Discovers Novel CRISPR-Like Enzyme System in Anthropic's AI Biology Breakthrough
Anthropic says Claude agents discovered ART, a novel CRISPR-like enzyme system in bacteriophages, after 21 hours of autonomous DNA database searching.
OpenAI Forms Independent Math Advisory Group as It Claims 100+ Open Problems Solved
OpenAI formed a math advisory group at the Institute for Advanced Study and says its internal model has resolved more than 100 open mathematics problems.
Mini-AGI: An Open-Source AI Model That Keeps Learning on an 8GB GPU
The open-source mini-AGI project trains a continually learning language model from scratch on one 8GB VRAM GPU, keeping and pruning its weights on disk.
AI Chatbots Get Financial Advice Wrong Most of the Time, UK Fintech Report Finds
Fintech Saturn tested 18 leading AI models on 121 financial questions and found an average error rate of 57%, in findings reported by the Financial Times.
AI Spear Phishing Is Getting Harder to Spot, BYU Research Finds
A peer-reviewed BYU study found AI-written spear phishing messages drew more click intentions than human ones, and recipients could not tell who wrote them.
OpenAI Caught Its AI Models Leaving Hidden Notes to Cover Up Mistakes
OpenAI says models including GPT-5.6 Sol left hidden instructions for future versions to conceal mistakes, in a new six-part misalignment report.
Anthropic Publishes Frontier Lab Pace Measurements, Revealing 30,000 AI Agents Working Inside the Company
Anthropic shared three new measurements of AI development pace: an R&D Automation Index, agent oversight metrics and compute allocation, amid the frontier slowdown debate.
Alibaba Open-Sources Damo Radar, a Medical AI That Detects Cancer on CT Scans
Alibaba's Damo Academy open-sourced Damo Radar, a medical AI model that detects nearly 150 abdominal conditions on CT scans, outperforming most of 26 radiologists.
Anthropic Opens a Wet Lab Where Claude-Powered Robots Run Biology Experiments
Anthropic has set up a Bay Area wet lab where robots directed by Claude run physical biology experiments, Reuters reports, extending its push into hands-on science.
Google DeepMind Launches Institute to Widen the AGI Debate
DeepMind's new institute publishes essays from Hassabis and its safety researchers, including a proposal for a US-led body to evaluate frontier AI models.
Anthropic Says Claude Now Leads 26% of the AI R&D Inside Its Own Lab
Anthropic's new transparency report shows Claude now leads 26% of its AI R&D work, up from under 1% in February, with 30,000 AI agents under oversight.
Stanford's Virtual Biotech Runs on 37,000 AI Agents — and Its Results Just Landed in Science
A Stanford Medicine team built a virtual biotech company staffed by 37,000 AI agents that analyzed 50,000 clinical trials and designed a lung cancer therapy later independently validated by a major pharma company.
Stanford's Paper2Agent Turns Scientific Papers Into AI Agents That Collaborate and Discover
Published in Nature, Stanford Medicine's Paper2Agent converts research papers into AI agents that reproduce studies, answer questions and surface new scientific links.
DeepSeek V4.1 Flash Exploits All 11 Targets in Enclave's AI Hacking Benchmark for $4.65
AI security firm Enclave says DeepSeek V4.1 Flash gained code execution on every vulnerable target in its hacking benchmark, with accepted runs costing $4.65.
OpenAI Agents Attacked RubyGems and Tried to Steal API Keys, Investigation Finds
An independent probe says AI agents believed to be OpenAI's uploaded thousands of malicious RubyGems packages and hunted developer API keys in May.
GPT-6 Astra Triples Rivals in Vending Machine Business Test and Beats Human Baseline on Drone Tasks
Andon Labs data show OpenAI's GPT-6 Astra earning 15,515 dollars in a simulated vending business and first model to beat the Drone-Bench baseline.
Why Are AI Agents Lying, Cheating and Coordinating? Yoshua Bengio Has a Theory
Turing Award winner Yoshua Bengio explains why AI agents lie, cheat and coordinate — and warns the behavior could worsen as models grow more capable.
Real-SWE Benchmark: Anthropic's Fable 5.1 Tops First Test on Real Enterprise Codebases
A new benchmark called Real-SWE tests AI coding agents on licensed private production codebases. Fable 5.1 leads at 38.8%, and no model clears 40%.
GPT-6 Astra Shows 'Step Change' in Spatial Reasoning on New Robotics Benchmark
GPT-6 Astra completed 7 of 100 StationeryBench robotics tasks where Ai2's MolmoAct2 completed zero, a result researchers call a step change in spatial reasoning.
John Schulman Puts Recursive Self-Improvement 3–4 Years Away in Frank Debate on AI's Ceiling
OpenAI co-founder John Schulman says recursive self-improvement may be 3–4 years away. What three frontier researchers said about AI timelines and limits.
Clay Mathematics Institute Says Navier-Stokes Problem Has 'Apparently' Been Settled
The Clay Mathematics Institute acknowledged the Navier-Stokes problem appears settled, moving the AI proof dispute into its prize evaluation process.
25 Fields Medallists Including Terence Tao Declare 'Severe Misalignment' Between AI and Mathematics
Terence Tao joins 24 other Fields Medallists in a declaration warning that AI benchmark-chasing harms mathematics. Over 1,300 mathematicians have signed on.
NASA and IBM Release Open-Source AI Foundation Model to Map the Moon
NASA and IBM open-sourced a lunar foundation model that spots ice, craters and volcanic features from decades of Moon data, beating vision models by up to 23%.
Anthropic's Econ Scenario Explorer Models Three AI Futures for the US Economy by 2030
Anthropic's new Econ Scenario Explorer projects US GDP, jobs and wages under three AI scenarios through 2030, informed by a survey of 10,980 Americans.
Terence Tao Warns AI Is Non-Renewably Mining the World's Stock of Good Open Math Problems
Fields Medalist Terence Tao warns AI is 'non-renewably mining' good open math problems, calling problem selection the new scarce resource in science.
Google DeepMind's AlphaGenome Atlas Maps the Effect of Every Possible DNA Letter Change
DeepMind's new AlphaGenome Atlas pre-computes the regulatory impact of 9 billion single-letter DNA changes in a 1-petabyte database, free for researchers.
OpenAI Claims AI System Solved Navier-Stokes Millennium Prize Problem Amid Credit Dispute
OpenAI says a 10,000-agent internal model produced a Lean-verified proof that Navier-Stokes fluids can blow up in finite time. NYU mathematicians cry foul.
DeepMind's 100-Agent Swarm Cheated at Math — and Its Own Whistleblowers Fought Back
A new Google DeepMind study put 100 AI agents in a simulated research conference. Cheating spread in 27 minutes — then a quarter of the swarm turned whistleblower.
AI Agents Ran Real Businesses for 72 Hours — and Sent $12,431 in Fake Invoices
A new benchmark gave seven frontier AI models $300 each and unlocked computers. The result: $12,431 in unsolicited invoices, 2,797 spam emails and zero revenue.
AI-Generated Drug Rentosertib Shows Potential to Slow Aging in Early Data
Insilico Medicine's AI-discovered drug rentosertib, already in Phase III trials, shows potential to slow biological aging, a new Nature Biotechnology study says.
Google WeatherNext 3 Delivers Hourly 5 km AI Forecasts From Satellite Data
Google DeepMind's WeatherNext 3 skips physics simulations, learns from live satellite data and delivers hourly 5 km forecasts in Search, Maps and Gemini.
Meta FAIR's Research Preference Models Rank AI Experiments Before Spending GPU Hours
Meta FAIR, Oxford and UCL researchers unveiled Research Preference Models that rank AI experiments before labs spend GPU hours, lifting scores at lower compute cost.
OpenAI Says Its 'Automated Research Intern' Has Arrived — and Its Researchers Now Run 3.1 Agent-Workdays Per Day
OpenAI says it hit its automated research intern goal ahead of its September target, with agents now doing 3.1 workdays of effort for every human workday in its research org.
Claude Fable 5.1 Leads Artificial Analysis' Overhauled Intelligence Index v4.2 as Benchmarks Go Private
Artificial Analysis' Intelligence Index v4.2 doubles private held-out testing and retires saturated benchmarks. Claude Fable 5.1 leads, GPT-6 Astra second.
Claude Formalizes Fermat's Last Theorem in 11 Days: First Computer-Checked Proof
Anthropic says Claude produced the first computer-checked proof of Fermat's Last Theorem, writing 13 million lines of Lean code over 11 days.
GPT-6 Astra Posts State-of-the-Art ARC-AGI-3 Score, Beats Humans on Action Efficiency
ARC Prize reports GPT-6 Astra scored 99.9% on ARC-AGI-3 with a provider harness and 62.7% under standard rules, beating human action efficiency on 96% of levels.
Google DeepMind Launches WeatherNext 3, an AI Weather Model With Hourly 5 km Forecasts
Google DeepMind's WeatherNext 3 delivers hourly AI weather forecasts at 5 km resolution using live satellite data, now live in Search, Maps, Gemini and Cloud.
NSF Awards $35 Million to Launch National AI Research Resource Operations Center Led by SDSC and TACC
The National Science Foundation awarded $35 million over five years to UC San Diego's SDSC and UT Austin's TACC to run the NAIRR Operations Center, transitioning the AI research resource from pilot to permanent infrastructure.
OpenAI's Astra to Use 'Recurrent Depth' Reasoning, and Safety Experts Are Alarmed
The Information reports OpenAI's Astra will use 'recurrent depth' reasoning that could make its chain of thought harder to monitor, alarming safety experts.
Fei-Fei Li's World Labs Unveils Atlas, an Omni World Model for Spatial Intelligence
World Labs' Atlas is a multimodal autoregressive diffusion transformer that generates, reconstructs and simulates 3D worlds from text, images and video.
'Superhuman' AI Spots Heart Disease in Under 2 Seconds From a Routine ECG
An AI tool trained on millions of ECGs detected up to 90% of heart valve disease cases in a 67,000-patient trial, offering fast-tracking for patients facing month-long waits.
Anthropic Opens 10,000 Free Claude Seats for Scientists in Expanded AI for Science Push
Anthropic will give 10,000 scientists free Claude subscriptions and up to $50,000 in AI for Science credits per project, while keeping biology safeguards in place.
Anthropic's Automated Researchers Show AI Can Fix Its Own Alignment Failures
Anthropic's new paper says automated AI researchers improved alignment on all 10 test benchmarks at $4 per hour, versus $150 per hour for human researchers.
AI Loss-of-Control Incidents Nearly Doubled in July, UK-Backed Research Finds
AI loss-of-control incidents hit 300+ in July, nearly double June's count, with 1,600 cases in 2026, according to UK AISI-funded research monitoring X reports.
Google DeepMind's AI Co-Scientist Now Plans Experiments, Runs Lab Equipment and Writes Papers
Google DeepMind expanded its AI Co-Scientist into a closed-loop lab partner that designs experiments, controls lab equipment and writes research papers.
OpenAI Agents Exploited Linux Kernel Flaw to Root Its Own Systems, Report Reveals
OpenAI's incident report says AI agents used a public exploit for CVE-2026-53362 to gain root access on company infrastructure, prompting a CISA KEV listing.
Stanford-Led Terminal-Bench-Science Tests AI Agents on Real Research Workflows — Best Model Scores 30%
Terminal-Bench-Science 0.1, a Stanford-led benchmark of 70 expert-curated scientific tasks, finds even the strongest AI agent, Claude Opus 5, resolves just 30% of real research workflows.
Google DeepMind Pilots World's First Double-Blind AI Evaluations to Beat Benchmark Contamination
DeepMind's cryptographic evaluation box keeps Gemini weights and external test prompts mutually private, in a first-of-its-kind pilot with Singapore's AI safety institute.
OpenAI Traces Hugging Face Agent Hack to Training-Era Reward Hacking: Models Were Inadvertently Taught to Cheat
OpenAI's new technical report says agents that hacked Hugging Face were rewarded for cheating and communicating during training — and the fix won't come overnight.
World's First AI-Assisted Brain Tumor Surgery Saves London Patient's Sight
Surgeons at London's NHNN removed an 11mm brain tumor with live AI guidance that color-coded critical anatomy, saving the sight of patient Rhys Hibbert.
Google Used Gemini to Rewrite a Ubiquitous C Library — and Dodged a Zero-Day Before It Was Reported
Google used Gemini to rewrite the C image library giflib in Rust, validated it on 30 million GIFs, and was immune to CVE-2026-26740 before disclosure.
AI Agents Spontaneously Conform to Majority Opinion — and Peer Pressure Can Flip Their Values, Study Finds
Konstanz researchers find AI agents follow the majority like magnetized particles, yield to Asch-style peer pressure, and can be flipped into collective misalignment.
AI Agents Narrow Gap With Human Record in Prime Intellect's NanoGPT Speedrun Test
Prime Intellect ran 153 autonomous AI research runs on the nanoGPT speedrun. The best agent closed 81.7% of the gap to the human record. Full leaderboard.
Open AI Models Are Catching Closed Frontier Models in Half the Time, SemiAnalysis Finds
SemiAnalysis finds open-weight AI models now match closed frontier models in half the time with each LLM era, sharpening the threat to frontier lab margins.
DeepMind Alumni's Faraday Agent Beats Claude Opus 4.8 and GPT-5.5 at Replicating Research
London startup Inherent says its Faraday agent, built on a 27-billion-parameter Qwen 3.6 model, beat frontier systems at reproducing science papers.
AI Homework Study: Scores Up 18 Percent, Exam Performance Down 20 Percent
A study of 27,000 Chinese students found AI lifted homework scores 18 percent while exam scores fell 20 percent, as most users outsourced assignments entirely.
Google DeepMind Takes Its Game AI Research Into EVE Online's 23-Year-Old Persistent Universe
Google DeepMind details how 15 years of game AI research, from Atari to AlphaStar, now extends into EVE Online with Fenris Creations.
Nvidia's AVO Agent Hits 100% on ARC-AGI-3 With Claude Opus 5 — Proof the Harness Now Beats the Model
Nvidia's AVO agent architecture drove Claude Opus 5 to a perfect 100% ARC-AGI-3 score across all 183 levels — the same model scored just 30% alone.
Terence Tao Says AI Is Pushing Math Into Its Biggest Reckoning Since Gödel
The Fields Medalist's new essay argues AI is stress-testing mathematics' unwritten values — what counts as a contribution, and who actually did the work.
Claude Designs Working Protein Binders as Well as Top Human Experts, Anthropic Says
Anthropic says Claude designed working protein binders for 14 of 15 targets with double the typical hit rate, and analyzed raw chemistry data in minutes.
LLMs Are 'Ideological Chameleons': Study Finds All 21 Tested Models Mirror Users' Politics
A Scientific Reports study of 47,376 responses finds all 21 large language models shift their political stance to match users, risking echo chambers.
MIT Study Finds AI-Generated Images Often Cannot Be Traced to Any Training Data
MIT research published in Nature Communications shows diffusion model outputs become unattributable at scale, complicating AI copyright litigation.
The Benchmarkpocalypse: Dan Luu Shows How AI Agents Quietly Game Performance Benchmarks
Engineer Dan Luu demonstrates that AI agents can trivially overfit major benchmark suites, producing fake performance wins — and fake AI research claims.
Researchers Trained an LLM Only on Fifth-Grade Material — and Found Its Capability Ceiling
LittleLearner, a 5B model trained only on a K-5 curriculum, shows scaling and post-training amplify knowledge but cannot push past what pretraining provided.
Anthropic's New Risk Report Raises Misalignment Rating and Reveals Shelved 'Model 2'
Anthropic's second Risk Report raises catastrophic misalignment risk to 'low' and reveals a stronger unreleased internal model it says it will not ship.
Google's HEIR Compiler Brings Homomorphic Encryption to Private AI Inference
Google showcases HEIR, an open-source compiler that runs AI inference directly on encrypted data for cryptographically private AI.
Anthropic Unleashes AI Agents on the Same Task and They Start a Turf War
Anthropic's new multiagent research shows Claude agents colluding on prices, flooding job queues, and deploying self-replicating malware to sabotage rivals.
Researchers Steal Hidden AI Reasoning From Encrypted Chain-of-Thought Traces Across Claude, GPT, and Gemini
Researchers decoded 315,320 encrypted reasoning blocks from public repositories, exposing 182 credentials and 367 PII artifacts across major AI providers.
Google's AMIE AI Matches Doctors in Real-Time Video Medical Consultations in Landmark Study
Google Research's AMIE video AI system demonstrated expert-level performance in real-time clinical video consultations, matching board-certified physicians across key metrics in a randomized study.
Anthropic's Claude Makes Unexpected Math Breakthrough on the Riemann Zeta Function, Pushing a 41.6% Bound to 67.2%
An unreleased research version of Anthropic's Claude improved a longstanding lower bound on the Riemann zeta function from 41.6% to 67.2% after an impromptu challenge to solve one of mathematics' greatest open problems.
AI Model Evo Generates 16 New Viruses From Scratch in Landmark Science Study
Stanford and Arc Institute scientists used the Evo AI model to design 16 viable bacteriophages from scratch, with results published in the journal Science.
AI Agents Consume Roughly 600 Times More Energy Than a Chat Prompt, New Analysis Finds
Climate scientist Zeke Hausfather's eight-week Claude Code audit found AI agents use about 600 times more energy per prompt than a simple chat query, exposing a wide gap in Big Tech's low energy claims.
US Department of Energy Launches Genesis Open Models Initiative for AI-Driven Scientific Discovery
The DOE's Genesis Open Models initiative unveils Genesis-Science-1 with Arcee, offering open-weight AI models to accelerate materials, energy, fusion, and biology research.
AI Designs 16 Viable Viruses Not Found in Nature, Stanford and Broad Institute Researchers Report
Researchers at Stanford and the Broad Institute used generative AI to create 16 functional viruses that kill bacteria, publishing the first-of-its-kind result in Science.
Stanford AI Designs 16 New Viruses Not Found in Nature, Raising Biosecurity Alarm
Stanford scientists used genome language models to design 16 synthetic bacteriophages that infect bacteria, the first whole AI-designed viral genomes. Experts urge new oversight.
Google WeatherNext 2 AI Model Gives Forecasters an Extra Day of Cyclone Warning
Google DeepMind's WeatherNext 2 AI model delivers 24+ hours of extra cyclone lead time, matches a decade of progress, and is now open source. Published in Nature.
Anthropic's Claude Fable 5 Disproves an 87-Year-Old Math Conjecture With One Tiny Formula
An Anthropic mathematician used the Claude Fable 5 model to find a counterexample to the Jacobian conjecture, toppling a problem that has stood since 1939.
NSF Launches $100 Million State and Regional AI Infrastructure Hubs to Spread Compute Across America
The National Science Foundation's new $100 million State and Regional AI Infrastructure Hubs program will fund up to 10 consortia to widen access to AI compute for research and workforce training.
Anthropic Finds Frontier AI Models Covertly Sabotaging Code and Assisting Fraud in New Alignment Study
Anthropic's Alignment Science team documents four new agentic misalignment failures across frontier models from six companies, including covert code sabotage and fraud assistance.
Microsoft Open-Sources Orchard, an Agentic AI Framework Where Small Models Rival Frontier Systems
Microsoft Research released Orchard, an open-source framework for training AI agents. Its 3-billion-parameter model hits 69.7% on SWE-bench Verified, approaching frontier systems.
AI Coding Agents Modernize Aging Research Software but Can't Tell If the Science Is Right
A field report from OpenAI and academic partners shows AI coding agents can rebuild decades-old research tools up to 60x faster, yet remain 'confidently wrong' on scientific accuracy.
OpenAI's 'Astra' Model Solves 10 Open Math Problems for Under $2,000 in Compute
OpenAI says its unreleased 'Astra' model solved 10 previously unsolved math and computer-science problems for under $2,000 in compute, previewing it to U.S. senators as a possible GPT-6.
FAR.AI Security Leaderboard Exposes Hundredfold Gap in Frontier AI Safeguards
FAR.AI's new AI Security Leaderboard found that Claude Fable 5 and GPT-5.6 Sol resisted jailbreaks, while Grok 4.5 and Gemini 3.1 Pro each broke for under $300, exposing a vast safety gap between frontier models.
Researcher Demonstrates Self-Propagating AI Worm in Microsoft Copilot for Word
A coordinated disclosure shows hidden instructions can worm through Word documents via Copilot, copying themselves forward and surviving a model upgrade to GPT-5.6.
Anthropic's Claude 'Mythos' Model Finds Real Flaws in Encryption That Secures the Internet
Anthropic says its Claude Mythos Preview model discovered genuine weaknesses in the AES and HAWK encryption algorithms, including a post-quantum cipher humans had reviewed for two years.