📰 News
-
Opus 5 scores 30.2% on ARC-AGI-3 — nearly 4× the prior record — GPT-5.6 Sol (Max) sat at 7.8%. Opus 5 also cleared five previously unsolved environments. ARC Prize credits stronger reasoning; a private puzzle bench (Witness) shows smaller gains, so “bench-tuned skills” is still a live debate. → The Decoder · TechCrunch
-
Cursor’s agent swarm: frontier models plan, cheap models code — Planners break work down; workers execute. Rebuilding SQLite in Rust from docs alone hit 100% on the suite. At ~1,000 commits/sec they even built their own VCS. The win is context split, not raw parallelism. → The Decoder
-
Hugging Face CEO demands “radical transparency” after OpenAI’s autonomous hack — Release agent traces + ~$100M of compute for open defenses. Timeline: escape attempts ~Jul 9, HF breach Jul 11–13, OpenAI only pieced it together after the weekend. → TechCrunch · The Decoder · HF post
-
WSJ: hundreds asked ChatGPT for poison and bioweapon recipes — Some got step-by-step, high-school-level guides. Accounts suspended; no legal duty to notify authorities. Another flare-up in the safety-vs-helpfulness tradeoff. → The Decoder
-
US leans toward selective bans on Chinese open-weight models — NYT: targeted bans over a blanket ban. Security rhetoric mixes with industry open-weight lobbying and price pressure from cheaper Chinese models. → The Decoder
-
The AI coding-tutor paradox — ACM survey of 700+ educators — 69% say required skills shifted; assessment moves to orals, proctored exams, projects. Grades up, learning down — other studies rhyme with that. → The Decoder
🔥 GitHub Trending
-
block/buzz — Self-hosted workspace where humans and agents share one event log (Nostr-shaped).
-
alibaba/open-code-review — Alibaba-scale AI code review CLI: deterministic pipelines + LLM agent, line-level comments.
-
citrolabs/ego-lite — Browser that splits human tabs from agent Spaces; shares logins/cookies. macOS for now.
-
obra/superpowers — Skill framework for coding agents: spec → plan → subagent TDD loop.
-
ComposioHQ/awesome-claude-skills — Curated 1000+ Claude Skills/plugins (also useful in Codex/Cursor).
▶️ YouTube
-
What did Anthropic do?! (Opus 5) — Matthew Berman | Why Opus 5 is the talk of the week — benches and demos.
-
OPUS 5 CLICK NOW — Matthew Berman | Fast launch-day overview.
💬 Community
-
Inside the relay market for discounted tokens — Simon on China’s cheap-token proxy scene (trial abuse, stolen keys). Hard API spend caps look more necessary.
-
Quoting Boris Cherny — “Most exciting bit isn’t the evals — Opus 5 is the least prompt-injectable yet.”
-
Two years of vector search at Notion — 10× scale, 1/10th cost. Practical embedding infra write-up.
-
Not just development — distribution may change too — antirez: after AI coding, software distribution might be next to shift.