Ten minutes of AI help makes people quicker to quit once it's gone
A UC Berkeley study finds that after about 10 minutes of ChatGPT-assisted problem solving, participants scored lower and skipped more questions once the tool was taken away.
AI research News and Analysis
24 verified stories covering AI research, product updates and industry developments.
Frontier models invent post-training methods, best reaches 15%
Epoch AI's early InnovationEval results: given 3,000 GPU hours each, the best frontier model recovered only 15% of the SDPO gain once inflated claims were stripped out.
AI Models Rarely Flag Their Own Failures in Reports
A new study finds leading AI models disclosed a damaging result in only 1% of self-written reports — until the prompt asked for honesty, and disclosure jumped to 95%.
Meta publishes six math papers co-written with Muse Spark
Meta paired mathematicians with Muse Spark's reasoning models to publish six papers, five of which answer previously open questions in math.
AI trained for under $8,000 beats Stratego's top player
An MIT-CMU-NYU-Stanford team's Ataraxos AI beat Stratego's most decorated human player 15-1-4, trained for under $8,000 using decision-time search instead of massive compute.
Data Drives 3x More Pretraining Progress Than Model Design
A controlled study finds pretraining data gains outpaced model design gains 12x vs 3.7x from 2019-2025, though annualized rates trail industry estimates.
OpenAI Says Research-Intern AI Goal Met, Agents Log 3.1 Workdays
OpenAI says internal data show AI agents now log 3.1 workdays for every human research day, but the same report calls the measurement preliminary and unverifiable from outside.
Tsinghua's Zeva Lifts Robot Success From 26% to 73%, No Retraining
Tsinghua's Zeva uses in-context causal memory on a frozen backbone to lift embodied manipulation success from 26% to 73%, without fine-tuning.
Anthropic Hands 250,000 Chats to Outside Researchers
Anthropic let Stanford, Oxford and METR analyze 250,000 real Claude conversations and had Imperial College London audit privacy — the findings challenge several assumptions about how people actually use AI.
ASR Models Recite Numbers That Were Never Spoken
A Hume AI probe of 11 leading speech recognition models finds the top scorers on public benchmarks are also the most likely to echo wrong reference transcripts instead of what they actually hear.
18 AI Models Went Autonomous — Top One Closed 82% of the Gap
Prime Intellect ran 153 fully autonomous research trials across 18 frontier models on a nanoGPT speedrun. The best model closed 82% of the human record gap, but not one of them found a genuinely new method.
Claude Designs Proteins, Hits 14 of 15 Targets
Anthropic had Claude design protein binders for 15 targets; Adaptyv Bio and Twist Bioscience validated hits in the wet lab, with 14 targets succeeding.
BigBang Lets Models Write Tasks
BigBang-V1 turns synthetic-data production into an agentic loop: models propose, solve, verify, and filter frontier tasks.
Google Finds AI Touches Only a Fifth of Tasks
Google ATLAS maps 14.65 million Gemini interactions and finds broad AI adoption, shallow task penetration and limited full automation at work.
MIT tests AI reasoning with 121 unfamiliar games
A Nature study from MIT, Cambridge, Princeton and collaborators suggests humans reason about new games with fast, shallow, goal-directed simulation rather than deep search.
Oak Lab sets a 20-watt goal for continuously learning agents
Oak Lab sets a 20-watt goal for continuously learning agents. The piece reviews the verified facts and why the signal matters beyond one announcement.
ICML 2026 gives both top paper awards to diffusion work
ICML 2026 gives both top paper awards to diffusion work is reframed for global technology readers, preserving the key numbers, claims and open questions.
A Stroop-style test trips up leading language models
A new attention test pushed major models into sharp accuracy drops, reminding buyers that reasoning claims still need careful benchmarks.
OpenAI reframes AI research as human-AI tandem work
OpenAI now says fully automating everything is not the future it wants, shifting from autonomous AI researcher language toward work in tandem with people.
Richard Sutton says generative AI cannot do real science alone
The reinforcement learning pioneer argues that next-token models lack causality, experimentation and self-generated experience.
DeepMind's AI Solves 6 Research-Level Math Problems
DeepMind's Aletheia, powered by a multi-agent framework, solved 6 out of 10 unpublished research-level math problems from the FirstProof challenge, with 5 solutions deemed publishable by external experts.
Fei-Fei Li’s World Labs pushes spatial intelligence
Fei-Fei Li’s World Labs pushes spatial intelligence. The story explains the announcement, the strategic context and the practical risks for the people or companies affected.
Google merges NotebookLM into Gemini
Google has integrated NotebookLM directly into Gemini's main workflow with a new Notebooks feature, allowing users to manage personal knowledge sources and sync them across both products.
LeCun Raises $1.03B for World Models After Leaving Meta
Yann LeCun founded AMI Labs and secured a $1.03 billion seed round at a $3.5 billion valuation to build AI that doesn't rely on next-token prediction, betting on his JEPA architecture as an alternative to large language models.