Live Feed
Gemini 3.5 Flash's 1M-Token Context Window: 77% Retrieval Accuracy at 128K, 27% at Full Scale — and Google Skipped General Reasoning Benchmarks
Independent analysis of Gemini 3.5 Flash finds its headline 1M-token context window degrades sharply in practice: MRCR v2 retrieval accuracy falls from 77.3% at 128K tokens to 26.6% at 1M. Google published only agentic benchmarks at I/O — no MMLU, GPQA, or competition math results.
Glasswing Month One: UK AISI Confirms Mythos Solves Both Cyber Ranges End-to-End, 2,100 Vulnerabilities Patched — Public Release Still Blocked
Anthropic's first Glasswing progress report: UK AI Security Institute confirms Mythos Preview is the first model to complete both multistep cyberattack simulations end-to-end. Claude Security patched 2,100 vulnerabilities in three weeks. Anthropic will not release Mythos-class models publicly until substantially stronger safeguards exist.
Pwn2Own Berlin 2026: OpenAI Codex Hacked Three Ways, Claude Code Flagged for Known Bugs, AI Tools Now a First-Class Attack Category
AI coding agents debuted as an official Pwn2Own category at OffensiveCon Berlin. Codex was exploited by three independent teams using three separate techniques. Claude Code attempts returned bug collisions — meaning known vulnerabilities were already in the codebase, unpatched.
Glasswing: 10,000 Vulnerabilities in One Month, Verizon Joins as First Telecom
Anthropic's Project Glasswing has found more than 10,000 high- or critical-severity vulnerabilities across critical internet infrastructure in its first month. Cloudflare found 2,000 bugs at 10x its normal rate. Verizon just became the first telecom to join. And the bottleneck is no longer finding vulnerabilities — it's patching them.
Newsom Signs First-in-Nation AI Labor Executive Order as Meta Cuts 8,000 Jobs
California Governor Gavin Newsom signed an executive order on May 21 directing state agencies to develop early-warning systems, revise WARN Act rules, and explore worker ownership models in response to AI-driven job displacement. It is the first state-level policy framework in the US to treat AI labor disruption as a government problem rather than a company problem.
Pentagon Tests OpenAI, Google, and Grok to Replace Claude — $200M Contract in Dispute
The US military began testing AI models from OpenAI, Google, and xAI on March 1, three days after designating Anthropic a supply-chain risk. Twenty-five power users are now running competitive evaluations on GenAI.mil. Anthropic is fighting the designation in court. The NSA is still using Mythos internally.
BitCPM-CANN: First Open-Source 1.58-Bit LLM Built End-to-End on Huawei Ascend 910B
ModelBest, Tsinghua University, and OpenBMB have released BitCPM-CANN, the first ternary large language model trained entirely on Huawei Ascend 910B NPUs — not ported, but natively built on Chinese hardware. At 1.58-bit weights, an 8B model retains 95-97% of full-precision accuracy with a 6x lower memory footprint.
Google Genie Connects 280 Billion Street View Images to Its World Model — Waymo Trains on It Already
Google DeepMind's Genie 3 world model now draws on 280 billion Street View images across 110 countries, turning two decades of real-world photography into an interactive, promptable simulator. Waymo has been using Genie to train its autonomous vehicles on rare-event scenarios.
DeepSeek Locks In V4 Pro's 75% Price Cut as Permanent — $0.87/M Output From June 1
DeepSeek has confirmed that its 75% promotional discount on V4 Pro becomes the permanent price on June 1, 2026. Output drops from $3.48 to $0.87 per million tokens — one-tenth the cost of GPT-5.5 at $30/M and a fraction of Claude Opus 4.7.
AI's Memory Appetite Has Doubled DRAM Prices — Samsung's 18-Day Strike Is Now in Progress
PC DRAM up 100% in a single quarter. LPDDR5 heading from $10/GB to nearly $20/GB. Samsung's 18-day strike, which controls 40% of world supply, started May 21. Gartner says the sub-$500 laptop segment disappears by 2028.
ArXiv Bans Authors 1 Year for Unchecked AI Output — Fake Citations Up 10x Since 2023
The preprint server powering ML research now has teeth: a 1-year submission ban for papers with incontrovertible LLM artifacts. Hallucinated citations hit 1 in 277 papers in early 2026, up from 1 in 2,828 three years ago.
Tri Dao's CODA Fuses Every Non-Attention Transformer Op Into the GEMM — Training's Last Memory Bottleneck Has a Fix
A paper from FlashAttention creator Tri Dao and collaborators introduces CODA, a GPU kernel abstraction that reparameterizes normalization, activations, and residual updates as GEMM epilogue programs. The operations run while the output tile is still on-chip, eliminating the memory round trips that have become a primary bottleneck in otherwise optimized training stacks.
Google Sold So Much TPU Capacity to Anthropic and Meta That DeepMind Researchers Are Queuing for Scraps
Bloomberg reports that Google DeepMind researchers compete with paying customers for access to the chips they built. Andrew Dai left the company after failing to secure compute for a Gemini improvement he discovered while playing a board game. Ioannis Antonoglou, a long-tenured contributor, also departed.
France's AION Consortium Places $10B Bid for EU AI Gigafactory — 288,000 H100s, 1GW Target
Eight French heavyweights — Iliad, EDF, Capgemini, Orange, and Ardian — have lodged the largest single-country bid inside the EU's 20B InvestAI programme. The 200MW initial phase targets 1GW of total capacity and would double France's current compute footprint.
Max Planck Paper Proposes Multi-Stream LLMs to Break the Single-Thread Agent Bottleneck
Researchers at MPI Tubingen show that replacing sequential message formats with parallel computation streams lets language models think while reading, act while thinking, and monitor themselves while generating — fixing a structural limitation present in every deployed agent today.