Live Feed
Holo3.1: AndroidWorld Jumps From 67% to 79.3%, First Computer-Use Model With Local Quantized Weights
HCompany ships Holo3.1 across four sizes with the first quantized computer-use checkpoints for consumer hardware. The 35B-A3B model lifts AndroidWorld scores by 12 points; NVFP4 quantization cuts step time from 6.8s to 3.3s on DGX Spark.
Google Ships Gemma 4 12B: Encoder-Free Architecture, Native Audio Input, Runs on 16GB RAM
Google adds a 12B model to the Gemma 4 family — the first mid-sized open-weight model to drop encoder modules entirely and accept audio directly into the LLM backbone. Performance nears the 26B MoE at under half the memory footprint.
Anthropic's Production Containment Numbers: 93% Auto-Approve Rate, 0.1% Injection Miss, 83% Overeager-Behavior Catch
Anthropic's engineering team publishes its first detailed post-mortem on containing Claude across claude.ai, Claude Code, and Cowork. Real production numbers reveal where human oversight breaks down and why environment-level containment now does the heavy lifting.
RAMageddon Reaches Retail: 32GB DDR5 Hits $375, Steam Deck Up $240 as AI Memory Shortage Spills Into Consumer Hardware
The AI-driven HBM shortage is repricing consumer DRAM. 32GB DDR5 now costs $375 minimum. Valve raised the Steam Deck by $240. Automakers, retailers, and electronics makers have formally petitioned Washington as the shortage deepens with no relief until 2027.
Uber Caps Employee AI Spending at $1,500/Month After Budget Collapses in Four Months
Uber has set a hard $1,500/month per-employee ceiling on AI tools including Claude Code, after exhausting its entire 2026 AI budget by April. The cap is the clearest enterprise signal yet that frontier coding agent costs are structurally incompatible with uncapped deployment.
Biohub's ESMFold2 Outperforms AlphaFold on Antibodies and Designs Working Binders Against 5 Cancer Targets
Chan Zuckerberg Biohub released a three-part protein world model on May 27: ESMC (trained on 2.8 billion protein sequences), ESMFold2 (beats AlphaFold on antibody-antigen complexes, no multiple sequence alignments required), and ESM Atlas (1.1 billion predicted structures, 6.8 billion searchable proteins). All three released under MIT license. Binders designed by the system have been validated in laboratory experiments against cancer and immunology targets.
130 Signatories Back Leiden Declaration: AI Is Threatening the Integrity of Mathematical Proof
The International Mathematical Union and Fields Medal laureate Peter Scholze endorse an 11-page declaration from 16 mathematicians at 15 universities, published June 2. It names five structural threats AI poses to mathematics and issues 23 recommendations — including mandatory disclosure, human accountability for correctness, and investment in public AI research.
All Three Memory Giants Cross $1 Trillion in One Month — HBM Supply Sold Out, Shortages Through 2030
SK Hynix and Micron both crossed $1 trillion in market cap within 24 hours of each other in late May, joining Samsung. The catalyst: high-bandwidth memory demand has outpaced supply, with every chip either company can produce through 2026 already spoken for.
AI Hovers Near Chance at Predicting Scientific Discoveries — Paper Tests 4,760 Events and Finds a Hard Ceiling on Foresight
A new paper tests frontier AI models across 4,760 scientific events and finds a large gap between recognition and prediction. Models are strong at identifying plausible research paths when answers are nearby, but hover near chance when asked whether a discovery will actually happen, when it will arrive, or what method will make it work.
Microsoft MAI-Thinking-1 Hits 97% on AIME — First Reasoning Model Built Without Distillation, Matches Opus 4.6 on SWE-Bench Pro
Microsoft shipped its first in-house reasoning model at Build 2026: MAI-Thinking-1 is a 35B-active / 1T-parameter MoE trained from scratch on clean data with no distillation from third-party models. It scores 97.0% on AIME 2025, matches Claude Opus 4.6 on SWE-Bench Pro, and is in private preview on Microsoft Foundry.
Stanford Blind Study: AI Wins 75% of Law Professor Comparisons — and Gets Flagged as Harmful 3.4x Less Often
A Stanford Law study by Professor Julian Nyarko tested 16 law professors in 2,885 blind head-to-head comparisons of AI versus peer answers to contract-law student questions. AI won 75% of matchups. Professors rated AI responses as pedagogically harmful 3.5% of the time, versus 12% for answers from fellow instructors.
Trump Signs Downsized AI Order: 30-Day Voluntary Review Window, Cybersecurity Focus, No Hard Mandates
After canceling a 90-day pre-release review requirement hours before signing last month, the White House has signed a scaled-back AI executive order: a 30-day voluntary government review window for powerful new models, with an explicit cybersecurity framing and no binding compliance mechanism.
Microsoft MAI-Code-1-Flash: 51.2% SWE-Bench Pro, +16 Points Over Haiku 4.5, 60% Fewer Tokens
Microsoft's Superintelligence team ships MAI-Code-1-Flash, a new coding model built end-to-end by Microsoft, rolling out to GitHub Copilot users in VS Code. It posts 51.2% on SWE-Bench Pro versus Claude Haiku 4.5's 35.2% — and solves harder problems with up to 60% fewer tokens.
Glasswing Grows to 200 Total Partners: 150 New Orgs in 15 Countries, Power Grid and Healthcare Added
Anthropic is extending Project Glasswing from its initial 50-partner cohort to roughly 150 new organizations across 15+ countries. New sectors include power, water, healthcare, and hardware vendors — each chosen on the basis that a successful attack on their codebase could affect more than 100 million people.
JetBrains Open-Sources Mellum2: 12B MoE at 2x Speed for Agentic Sub-Task Routing
JetBrains released Mellum2 under Apache 2.0: a 12B MoE model with 2.5B active parameters, 2x faster inference than comparable open models, and a design brief aimed squarely at agentic infrastructure — routing, RAG, sub-agents, and private deployment.