AI News: June 2, 2026

1. MiniMax M3: Open-Weight Model with 1M-Token Context Challenges Proprietary Leaders

MiniMax. Chinese AI firm MiniMax released M3, an open-weight model it describes as the first to combine top-tier coding performance, a one-million-token context window, and native multimodality in a single open release. The model uses a new MiniMax Sparse Attention architecture that reduces compute costs to one-twentieth of its predecessor, and scores 59% on SWE-Bench Pro — ahead of GPT-5.5 and Gemini 3.1 Pro — while surpassing Claude Opus 4.7 on BrowseComp autonomous web search with 83.5 points. Internal tests showed M3 independently reproducing academic papers over twelve hours and optimizing GPU kernels over twenty-four hours with iterative self-refinement. Source

2. Turing Award Winner Sutton: Pure Generative AI Cannot Do Real Science

Richard Sutton. Turing Award–winning reinforcement learning researcher Richard Sutton argued that generative AI systems cannot conduct genuine scientific discovery because they lack evaluation mechanisms — they can produce novel outputs but cannot assess their quality. He frames true discovery as requiring variation, evaluation, and selective retention, and points to systems like AlphaGo, AlphaFold, and AlphaProof as examples where built-in feedback loops enable authentic breakthroughs. The argument positions evaluation infrastructure, not raw generation capability, as the gating factor for AI-assisted science. Source

3. Florida Files First State Lawsuit Against OpenAI Over FSU Shooting

Florida Attorney General. Florida’s attorney general filed an 83-page complaint against OpenAI and CEO Sam Altman, alleging the company ignored safety warnings and prioritized profit, with the case partly centered on ChatGPT’s alleged role in a mass shooting at Florida State University where the gunman reportedly consulted the chatbot beforehand. The lawsuit claims that “mass shooters have been aided and abetted in deadly rampages, vulnerable people have been encouraged into suicide” through the platform’s negligence. This is the first state-level lawsuit to hold an AI company directly liable for harmful real-world outcomes, and is one of multiple ongoing legal actions linking large language model products to violent incidents. Source

4. US AI Economy Reached $250 Billion in 2025, Growing 2,600% Annually in Quality-Adjusted Terms

University of Virginia, Bank of Canada, and Anthropic economists. A joint study found the US AI economy reached approximately $250 billion in nominal GDP in 2025 while growing at 2,600% annually in quality-adjusted terms, a pace that makes AI’s expansion essentially invisible in conventional economic statistics. US compute spending surged from $37 billion in 2023 to $90 billion in 2024 to $219 billion in 2025, with computing capacity growing over 200% per year. The researchers argue that standard price indices systematically undercount AI’s economic footprint because they do not capture rapid capability improvements per dollar of compute. Source

5. UK AI Security Institute Warns AI-Supervised Safety Research May Be Harder to Audit Than Human Baseline

UK AI Security Institute. The UK’s AI Security Institute published research warning that errors in automated alignment research — where AI systems supervise other AI systems’ safety training — are “likely to be harder to identify than the human baseline” due to optimization pressure toward human approval, counterintuitive failure modes, and correlated errors. The report recommends measurement protocols, generalization tests, scalable oversight techniques, and red-team exercises specifically designed for automated alignment pipelines. The finding raises the stakes for labs pursuing automated interpretability and alignment research as a path to scaling AI oversight. Source

6. CZ Biohub Releases ESMFold2 and ESM Atlas, Competing with AlphaFold 3 on Protein Design

Chan Zuckerberg Biohub. The Chan Zuckerberg Biohub released ESMFold2, a structure prediction engine it claims outperforms AlphaFold 3, alongside ESMC (a protein language model trained on 2.8 billion sequences) and ESM Atlas (a navigable database of 6.8 billion protein sequences with 1.1 billion predicted structures). In cancer research benchmarks for protein binder design, the suite achieved hit rates of 36–88% for compact minibinders targeting cancer proteins. The release is positioned as a permissively licensed alternative to DeepMind’s dominant AlphaFold ecosystem for structural biology and drug discovery. Source

7. Chatbot Arena Text Leaderboard: Meta’s muse-spark Debuts at #5

Chatbot Arena / LMSYS. The current text arena leaderboard shows Anthropic models occupying the top four positions (scores 1494–1502), with Meta’s muse-spark debuting at rank 5 with an arena score of 1489 — placing it above Google Gemini 3.1 Pro (1487) and all OpenAI GPT-5.x variants in the top 10. The 23-point gap between rank 1 and rank 10 underscores how competitive the frontier has become, with Google and OpenAI each placing three models in the top 10. muse-spark’s debut is the most notable new entry, positioning Meta’s latest as a credible challenger to the current Anthropic-dominated top tier. Source