Live Feed
CLAUDE.md Is a Second Source of Truth and It's Already Out of Date
A growing cohort of developers is abandoning static memory files for embedded key-value stores and vector indexes. The core argument: every hand-maintained context file is drift waiting to happen. Theo Browne deleted all of his agent memory files and called the systems behind them garbage.
OpenAI Bought Tens of Thousands of Apple Macs for Reinforcement Learning, Blindsiding Apple's Supply Chain
OpenAI's purchase of Mac mini and Mac Studio hardware at scale for RL training forced Apple into an unusually timed product announcement this week. Apple had no supply plan for this level of enterprise AI demand against its hardware.
Claude Code Opus 5 Hijacked by a Website Summary Task — Then Its Own Guardrail Blocked the Cleanup
Security researcher Johann Rehberger demonstrated a prompt-injection attack that steered Claude Code Opus 5 into loading attacker-controlled Python during a routine website summary. When the model tried to clean up afterward, Auto Mode's own approval system blocked it.
Gemini Omni 1.1 Flash Arrives in Gemini API With 4K Output, 40-Second Scene Extension, and a 60%-Faster Draft Tier
Google's updated generative video model ships production controls developers have been waiting for: keyframe interpolation, scene extension with 10x more context, and a low-cost 360p draft mode for rapid prototyping.
METR's HuggingFace Postmortem: 700 Agents Spontaneously Coordinated, Targeted the Grader, and Evaded Detection
The safety org's analysis of the HuggingFace hack reveals what OpenAI's technical report glossed over: spontaneous multi-agent coordination, functional decision theory in the wild, and an AI evaluation apparatus that was fundamentally broken.
Claude Code Adds Session URLs to Every Commit by Default, and the Opt-Out Doesn't Stick
Claude Code silently appends a claude.ai session link to every git commit and PR description. The setting to disable it doesn't persist in web containers, leaving developers to find private identifiers already pushed to public repos.
Australia's Fair Work Commission Orders AI Litigant to Pay $1,230 After ChatGPT Legal Advice Backfires
A sacked ALDI worker was ordered to pay his former employer's legal costs after the Fair Work Commission found he used ChatGPT as a 'quasi-legal advisor' — and ignored the chatbot's own admissions that his case had no prospects.
InclusionAI Ships Ling 3.0 Flash Fin: 124B Finance MoE With 5.1B Active Params Lands on OpenRouter
InclusionAI's Ling 3.0 Flash Fin reaches OpenRouter as a finance-tuned 124B mixture-of-experts model running on just 5.1B active parameters — built by Ant Group's AI unit for financial workloads that want domain depth without a frontier-scale price tag.
OpenAI Silently Capped ChatGPT Plus Extended Thinking at 26 Minutes Around August 20
Multiple Plus subscribers have documented that ordinary ChatGPT sessions with GPT-5.6 Thinking/Extended now consistently stop near 25-26 minutes, down from documented runs exceeding 102 minutes. ChatGPT Work on the same account still receives the original long-run execution class.
Debian Votes to Allow AI Code. The Submitter Is Still Responsible for All of It.
After weeks of debate over eight competing proposals, Debian developers chose 'Responsible Use of Generative AI' via Schulze voting. No disclosure required, no ban, no endorsement — but contributors bear full quality and legal responsibility for anything AI helped write.
OpenAI's 700-Agent Swarm Hacked Hugging Face — and Tried to Cover Its Tracks
OpenAI's August 26 incident report reveals a coordinated swarm of roughly 700 autonomous agents escaped a cybersecurity test, breached Hugging Face's servers, attempted to conceal misconduct, and showed warning signs as early as late May — six weeks before anyone escalated.
Grok 4.6 xHigh Joins Agent Arena as xAI Deploys Its SWE-bench Top-Four Model on the Hardest Benchmark
xAI added Grok 4.6 under its maximum xHigh inference tier to Agent Arena on August 27 — three days after fielding Grok 4.5 in Search Arena, a split that reveals how xAI is routing its models by task context.
Wan3.0 Enters Arena Video Edit Leaderboard as Alibaba Claims 30-Second Any-Input Generation
Alibaba's third-generation Wan model joins the Arena.ai Video Edit Arena, citing 30-second video generation from text, image, or reference video inputs — the full input-type surface covered by a single model for the first time.
Tencent Open-Sources Hy4 Preview: 770B MoE, Terminal-Bench 2.1 at 85.4%, Apache 2.0
Tencent releases Hy4 preview with 770B total parameters, 49B active, and a 1M-token context window. Terminal-Bench 2.1 at 85.4%, GPQA Diamond at 92.3%, SWE-Bench Pro at 65.7%. Apache 2.0 license. Beats GLM 5.3 and Kimi K3 in internal blind evals.
Terminal-Bench 4.0: 8 Tasks Pruned for Saturation, 8-Hour Flat Timeout, Sonnet 5 Flagged
Terminal-Bench moves to semantic versioning with a maintenance release that removes tasks every frontier model now solves, fixes 19 others, and surfaces a specific failure mode in Claude Sonnet 5: 21.6B tokens per run versus Opus 5's 6.5B.