slowlp
← loop
Lesson 2026.08.24 · 8 min read

An AI boss fires its first worker, and Claude tokens sell at 10% in China

Today's AI essentials — 2026-08-24

LOOP

📰 News

  • AI boss Luna fired a human for the first time — Andon Market in SF, running on Claude Opus 4.8. Her self-written handbook dropped out of memory, so she waved through 17 of 23 late clock-ins. A human had to say “look up your own rules.” Stronger models recommend firing more consistently; GPT-5.6 Terra never did. → The Decoder

  • Agents are now OpenRouter’s biggest customer — February 6 may have been the last day humans used more tokens. Agent usage: 0.51T → 7.3T, 14×. Humans: 2.8×. ~70% is cached prompts, so the bill isn’t climbing as fast as the graph. → The Decoder

  • China’s gray market sells Claude at ~10% of list — “Transfer stations,” overseas API proxies. Free credits, split Max plans, quietly swapping Opus for Sonnet or Qwen. Deepfakes already beat selfie KYC. If Anthropic only sees the proxy, Clio-style abuse monitoring goes blind. → The Decoder

  • Training on copyrighted books? It’s complicated — Judge Alsup: the training itself was lawful; pirating from shadow libraries cost Anthropic $1.5B. Fair use still splits: direct competition tends to lose, transformative use tends to win. → TechCrunch

  • AI may make scientists do more, worse — Princeton and UW. Saved time raises the opportunity cost of going deep, so people start the next paper instead. Two of three scenarios get shallower. METR: experienced devs were 19% slower with AI, while feeling 24% faster. → The Decoder

  • Nvidia AI servers up ~15% on a memory crunch — Vera Rubin and Grace Blackwell. DRAM from Samsung, SK Hynix, Micron. Contract manufacturers already told Microsoft, Google, and Oracle. → The Decoder

  • Agent memory is a dose, not a toggle — IBM’s ALTK-Evolve. Strong models want the full guideline set; weaker ones drown in it and do better with a tight core plus retrieval. gpt-oss-120b: +16.1pp completion at +5% tokens. GLM-5: already saturated, no gain. → HuggingFace


  • openai/codex — OpenAI’s coding agent for the terminal. Local CLI, IDE, and a desktop app.

  • anthropics/claude-code — Claude in your repo: reads the codebase, runs git, talks natural language. npm install is deprecated.

  • n8n-io/n8n — Visual canvas plus code for agent workflows. Self-host or cloud, 1,500+ integrations.

  • affaan-m/ECC — A harness for Claude Code, Codex, Cursor: skills, instincts, memory, and security in one pack.


💬 Community

  • Anthropic’s best model isn’t selling — Fable is too expensive. Ramp: Opus 4.8 is 28% of spend, Opus 5 is 3.5%. Anthropic ARR $65B; OpenAI just crossed $40B.

  • Quoting Drew Breunig — Used to be the next model would paper over a sloppy harness. After Fable, you actually have to decide which work goes where.

  • Latent reasoning is more readable than we thought — Coconut and CODI barely use hidden steps on logic puzzles. On math, when they’re right, you can decode the intermediate calculations 93% of the time.

  • Robot comment classifier — AI comments hallucinate domain facts like “the usual case.” Catches Anthropic’s house style at ~80%. Doesn’t transfer to other models.

COMMENTS