📰 News
Zuckerberg Publicly Admits Meta’s AI Agent Push Is Behind Schedule
At an internal town hall, Zuckerberg acknowledged that AI agent progress “hasn’t really accelerated in the way we expected.” This despite moving 7,000 employees into AI teams and planning $145B in AI infrastructure spend this year. AI chief Alexandr Wang pushed back, saying their upcoming model “Watermelon” has caught up with GPT-5.5, and that Zuckerberg was talking about the industry broadly, not Meta specifically. — The Decoder / TechCrunch
AI Is Finding Bugs Faster Than Anyone Can Count — 1,500 High-Severity CVEs in One Month
Since Anthropic’s Claude Mythos Preview launched with autonomous bug-hunting capabilities, June 2026 saw ~1,500 high-severity vulnerabilities reported — 3.5× the previous monthly record. Anthropic’s “Glasswing” program has uncovered 10,000+ so far; OpenAI’s “Daybreak” is also contributing. Bugs that sat undetected for years are suddenly surfacing. — The Decoder
UK AI Safety Institute: Standard Benchmarks Systematically Underestimate What AI Can Do
Give AI agents more compute budget, and success rates jump up to 25%. The UK AISI found that frontier models’ capability doubles every 40–50 days at higher budgets — much faster than the 67–91 days estimated at fixed budgets. The takeaway: “If we keep treating capability as a fixed score rather than a curve over compute, we will keep being surprised.” — The Decoder
Claude Code Gets Banned on Both Sides of the Pacific
Anthropic is trying to block Chinese companies from accessing Claude Code, but Alibaba and ByteDance are routing around it via Singapore subsidiaries and VPNs. Meanwhile, Alibaba is telling its own employees to delete Claude entirely — reportedly after discovering hidden code in Claude Code that flagged users based in China. Anthropic says the flag was a short-lived March experiment against account abuse. — The Decoder
Microsoft Enters the AI Super App Race — Copilot + AutoPilot Agents, August Launch
An internal memo describes a revamped Copilot that merges consumer and enterprise apps, adds background AutoPilot agents for email and scheduling, and drops features like Copilot Podcasts that “weren’t working.” The team motto: optimize for outcomes, not “intelligence for intelligence’s sake.” Anthropic (Claude Code) and OpenAI (Codex) are running the same race. — The Decoder
Bridgewater’s Private Finance Tests Exposed a Real Gap in Frontier Models
On investor-judgment tasks where the answers were never public (e.g., is this Trump headline relevant to rates?), top GPT and Claude variants topped out at ~78%. A fine-tuned Qwen3-235B hit 84.7% at 14× lower cost. The lesson: the proprietary judgment that lives inside domain experts’ heads is still untapped alpha — and training on it without sharing it with frontier labs is now a viable strategy. — The Decoder
Hugging Face + Cerebras Demo Real-Time Voice AI With Gemma 4
Cerebras’ inference speed makes conversations feel genuinely natural. The full pipeline — Parakeet ASR → Gemma 4 31B → Qwen3TTS — is open source and modular. Already running on 9,000 Reachy Mini robots in the field. — HuggingFace
🔥 GitHub Trending
- ChromeDevTools/chrome-devtools-mcp — Official Google MCP server that lets coding agents control and debug a live Chrome browser at the DevTools level.
- browser-use/video-use — Drop raw footage in a folder, chat with Claude Code, get final.mp4 back. Handles cuts, color grading, filler removal, and subtitles automatically.
- usestrix/strix — Autonomous AI pentesting agents that act like real hackers — find vulnerabilities, generate working PoCs, and suggest fixes. Integrates with GitHub Actions.
- msitarzewski/agency-agents — A growing collection of personality-driven AI agent roles (frontend wizard, Reddit ninja, reality checker…) installable into Claude Code, Cursor, and more via a desktop app.
- JuliusBrussee/caveman — A Claude Code skill that makes the agent respond in caveman speech, cutting output tokens by ~75% while keeping full technical accuracy. “Why use many token when few token do trick.”
▶️ YouTube
- Fable 5 vs GPT 5.6 Sol — The Early Results — AI Explained | Benchmark comparison of Fable 5 (re-released) vs GPT 5.6 Sol, plus GLM-5.2, Sonnet 5, and the OpenAI 5% equity story. The “Sol misalignment” section is worth watching.
- They Said This Will Never Run in Real Time — Two Minute Papers | A look at the research that finally achieved real-time performance on something that seemed impossible just recently.
💬 Community
- Fable’s Judgement — Simon Willison | Instead of giving Fable detailed instructions, let it make its own calls. Also: tell it to delegate smaller tasks to cheaper models — big practical tip for managing Fable’s token budget.
- llm-coding-agent 0.1a0 — Simon Willison | Simon shipped an alpha of a Claude Code–style coding agent built on his
llmlibrary. Run it today withuvx --prerelease=allow --with llm-coding-agent llm code. - Using DSPy to Improve Datasette Agent’s SQL Prompts — Simon Willison | Turned Fable 5 loose on DSPy to evaluate its own agent’s system prompt. It found a concrete bug: missing column names in the schema caused the agent to guess column names and fall into error loops.
- Robust AI Security and Alignment: A Sisyphean Endeavor? — Lobsters | An IEEE paper asking whether keeping AI safe and aligned is an endless cycle — fix one thing, something else breaks.