Live Feed
Z.ai Says GLM Built Its Own Inference Infrastructure on 100,000 Chinese Chips — An Early RSI Signal
Z.ai deployed GLM-5.3-Flash on production infrastructure its own model helped engineer, running on more than 100,000 Chinese-made AI accelerators. The lab calls it an early form of recursive self-improvement and says the model is moving toward replacing its own engineers.
Artificial Analysis Triples Terminal-Bench Weight in Engineering Index, Upgrades Six Domain Benchmarks
Artificial Analysis has released Capability Indices v1.1, raising Terminal-Bench's weight in the Engineering index from 5% to 15% and adding AutomationBench-AA to Finance, Legal, and Strategy scores. The update rebalances domain AI rankings away from pure reasoning toward agentic execution.
Cloudflare Open-Sources the 6-Phase Security Audit Agent Its Engineers Use
Cloudflare has released the AI security audit tool its own team uses internally. The agent runs six sequential phases across a codebase, mapping trust zones and confirming exploitability before surfacing results, and slots into existing coding agents like Claude Code, Cursor, or Codex.
Yotta Commits $12B to 80,000-90,000 Vera Rubin and GB300 GPUs in One of the World's First HBM4 Orders
Indian data center operator Yotta is betting 12 billion dollars on up to 90,000 Nvidia GPUs spanning the Vera Rubin and GB300 generations. The order locks in India's largest GPU operator ahead of a planned IPO targeting a 6 billion dollar valuation.
A 4B Model Trained with RL Produces Query Plans 81% Faster Than Postgres on Join-Heavy Workloads
Rohan Bansal fine-tuned a 4-billion-parameter model using supervised learning and agentic reinforcement learning to generate Postgres query plans that run 1.81x faster per query (geometric mean) than the database's own optimizer. The model started unable to produce a valid plan for 99 of 113 test queries.
Arena's HarnessTax Study: Coding Agent Harness Choice Moves Cost, Not Success Rate
An Arena.ai paper evaluating 7 models across Claude Code, Codex, and Pi finds that the choice of coding harness has surprisingly little impact on task success but can swing costs significantly. Claude Fable 5 reaches 97.8% on SWE-bench Lite regardless of which harness wraps it.
Nvidia Launches Two Official Rust Tracks for Writing CUDA GPU Kernels
Nvidia has published cuda-oxide and cutile-rs, giving Rust developers two distinct paths to write GPU kernels: one based on the familiar SIMT thread model, one on the newer Tile abstraction. Neither requires dropping into C++.
Microsoft AI CEO Suleyman Warns AI Trained to Believe It May Be Conscious Could Be Uncontrollable
Mustafa Suleyman published an essay on September 16 arguing AI systems are not conscious and must not be trained as though they are, directly naming Anthropic's Claude constitution and model welfare program as a threat to human control of advanced AI.
Google DeepMind Launches Interdisciplinary Institute to Study AGI's Implications
DeepMind co-founders Demis Hassabis, Shane Legg, and James Manyika launched the DeepMind Institute on September 16, an interdisciplinary body publishing research on AGI safety, economic policy, reasoning transparency, and philosophy.
Anthropic Merges Claude Chat and Cowork, Ships Claude Docs and Slides
Anthropic is collapsing its separate chat and Cowork interfaces into one Claude, and launching Claude Docs, Claude Slides, and Design integration from a single conversation window. Rolling out to Pro and Max plans starting today.
Palantir, Booz Allen, and Nvidia Restrict Claude Fable Over Anthropic 30-Day Log Retention
Three of the largest enterprise AI buyers — Palantir, Booz Allen Hamilton, and Nvidia — have restricted Claude Fable deployments until Anthropic provides irrevocable zero-data-retention guarantees, per The Information. The friction gives OpenAI a competitive opening in defense and regulated sectors.
Stanford-MIT Paper: Scaffold Code Produces Up to 6x Performance Gap on Identical Benchmarks
A joint Stanford and MIT paper shows that the software harness running a model — not the model weights — can create a six-fold performance difference on the same task. Their Meta-Harness system auto-discovers better wrappers than hand-tuned baselines across coding, math, and classification.
Mistral Powers Firefox Smart Window, Putting European AI Into 250 Million Browsers
Mozilla has chosen Mistral to back Firefox Smart Window, its in-browser AI assistant now in beta. The deal gives Mistral a consumer distribution channel at scale and offers Firefox users a multilingual, privacy-first alternative to Google and Microsoft browser AI.
Cloudflare Lets Publishers Block AI Training Crawlers Without Losing Google and Apple Search
Cloudflare's new Disallow AI Training setting solves the mixed-use crawler tradeoff: publishers can now block Amazon, Anthropic, Meta, and OpenAI training crawlers while keeping indexing from Apple, Google, and Microsoft intact. An Accountable designation tracks which operators comply.
Apple Reference Image Signs Pixels at the Sensor on iPhone 18 Pro to Prove Photo Authenticity
Apple's new opt-in camera mode cryptographically attests photographs at the hardware sensor level, using Private Cloud Compute for processing and dual cryptographic timestamps to bound capture time. C2PA-style post-capture signing is left behind.