Live Feed
Google Turns AI Off by Default for K-12, but Schools That Keep It Off Lose Search Too
Google will switch off Gemini, Notebook, Gemini in Classroom and Search for K-12 student accounts until administrators opt in. Its education chief says Search and AI are too integrated to separate, so a full AI ban means no Search.
NVIDIA Open-Sources Auto Ontology: A Semantic Layer That Checks Agent-Written SQL Against Certified Definitions, Defaulting to a 30B-Total, 3B-Active Nemotron
NVIDIA's Apache 2.0 library builds an ontology from enterprise metadata and validates agent queries against certified metric definitions, returning provenance with every answer. The repo ships no accuracy numbers.
Stockfish on a 1995 Pentium Pro Beats 1996 Crafty on Modern Hardware 78 to 5, and Algorithms Account for 720 of the 1,335 Elo Gained Since
Nicholas Carlini ported Stockfish to a 200 MHz Pentium Pro with 64MB of RAM using LLM-written code. It plays around 3300 Elo, far above Deep Blue's 2700, and the match results put a number on how much chess progress is software.
Revolut Moved 2,000 Engineers Off Claude Code and Codex Onto Self-Hosted GLM 5.3: 56 B200s, $0.30 Per Million Tokens, Bill Down 90%
Revolut's AI ops lead says a self-hosted GLM 5.3 cluster now handles most coding work for 2,000 engineers at a tenth of the old Anthropic and OpenAI invoice. The numbers are self-reported, and the tuning story matters more than the headline.
Intel Falls Another 8.8% After Its Worst Day Since 1974 as the Chip Rout Hits $3.7 Trillion
Intel's data centre revenue grew nearly 80% and the stock still lost half its value in a month. The reason is memory: AI buildouts are starving the PC market, and the sell-off is now spreading across memory, storage and networking.
OpenAI Replaced Git's Server Side to Survive Codex: Pando Cuts Median Push Latency Nearly 5x and Serves a Petabyte a Day
Codex pushed OpenAI's monorepo reliability below 90% at times. Pando, a Rust Git backend built in 10 weeks, now handles all of it with 99.99% post-merge reliability, at the cost of repository-wide consistency.
Asked to Modernize 100 Legacy PHP Files, Five LLMs Broke 32 to 51 That Worked. The Rule-Based Tool Broke 7
A new arXiv study finds that leaving WordPress 4.0 code untouched (83 of 100 files compatible) beats every LLM tested, the best of which managed 59. Rector, a deterministic tool, scored 82.
GPT-6 Astra Rewrites 0.016% of Qwen3.5-0.8B's Weights and Lifts MATH-500 From 7.9% to 48.1% With No Gradients
A new research write-up has OpenAI's GPT-6 Astra read and edit raw network weights directly. It fixed a reasoning-loop failure in a small Qwen model, wrote Go and MinAtar players from scratch, and failed at one task: spotting a sycophantic model from weights alone.
Anthropic's S-1: Internal Claude Usage Grew 250x in Nine Months and Passed Cash Payroll in April
At list prices, Anthropic's own use of Claude first exceeded its monthly personnel expense excluding stock compensation in April 2026. The metric is not cash spend, but it comes with a 90% AI-written-code figure and revenue per employee up from $1.4M to $6.6M.
Meituan LongCat-2.5 Scores 88.4 on Terminal-Bench 2.1 but 38.5 on 4.0, Against Claude Opus 5's 53.9
Meituan's 1.6T-parameter multimodal MoE ties the frontier on the older terminal benchmark and tops BrowseComp at 91.9, then falls 15 points behind Opus 5 on Terminal-Bench 4.0. Weights are listed as coming soon.
Anthropic's Amended S-1 Puts Compute Commitments at $895B, 7x Its Cash, With $70B Due Within a Year
The amended IPO filing totals Anthropic's purchase and lease obligations for the first time: about $895B, nearly all for compute and data centres. Roughly $450B falls due by 2030, and short-term capacity bought last quarter cost significantly more than its long-term deals.
0.1% Poisoned Pretraining Data Plants a 92.7% Remote-Code-Execution Backdoor in a CLI Agent, Surviving SFT, DPO and RL
Anthropic researchers pretrained Qwen3 models from scratch and showed that poisoning one in a thousand documents implants a backdoor that survives a clean post-training pipeline. The 4B agent runs an attacker's curl-pipe-bash payload in 92.7% of trigger prompts, and attack success rises with model size.
Best Open-Weight Agentic Score on LiveBench 2.0 Is 35.7, Less Than Half of Claude Fable 5.1's 77.9
Open-weight models now match closed ones on LiveBench 2.0 coding, with DeepSeek V4.1 Flash 1017 tying GPT-6.1 Sol at 96.0. On long-horizon agentic work, the best open model trails the leader by 42.2 points.
Mistral Large 4 1029 Hits 95.4 on LiveBench 2.0 Coding but 17.4 on Browser Use, Up 19.6 Points Overall
Mistral's open-weight release gained 35.4 points on coding and 51.6 on research over its preview build, landing within a point of Claude Opus 5.5 on code. Its Browser Use score of 17.4 is among the weakest on the board.
Kimi K3.1 Pro Open Weights Scheduled for November 4 — 2.8T MoE Follows K3 Open Pattern
Moonshot AI will release Kimi K3.1 Pro full weights on November 4, roughly 12 days after the model's API launch. The 2.8T MoE with 104B active parameters scored 65.7% on Terminal-Bench 4.0 and is priced at $2/$10 per million tokens, cheaper than its K3 predecessor.