slowlp
← loop
Lesson 2026.08.31 · 8 min read

Claude Code limits rise on paper, fall in practice — Ox-Alpha is GLM-5.3-Flash

Today's AI essentials — 2026-08-31

LOOP

📰 News

  • Claude Code: a raise on paper, a cut in the chair — On September 14, Pro, Max, Team, and Enterprise baselines go up 25% from the original. The catch: a temporary 50% boost is live now, so weekly capacity actually drops about 17%. Anthropic confirmed it on X. New consolation prize: when a run fails, Claude can draft a bug report for you. → The Decoder

  • Coding agents have no sense of time — MATS work on Claude Code and Codex. Ask how long a task will take, then how long it took: both guess ~90 minutes regardless of difficulty, and miss worst on short jobs. Same model runs 2.5× more steps in Claude Code than in Codex. They score themselves 70% on work that scored 7. Give them an elapsed-time tool and they nail it. → The Decoder

  • Workplace AI love fell from 81% to 43% — Glassdoor. Negatives hit 53%. Gen Z women are the skeptics at 21% positive. Insurance claims adjusters are almost all negative; writers and CS lean 4-to-1 against. Job-loss fear is only 20% of the complaints — forced tools and worse customer experience follow close. Execs remain the cheerleaders. → The Decoder

  • The skills that earn A’s are the ones GPT can fake — 1,053 Bocconi freshmen, GPT-4o, a 180-word marketing memo. Nearly a full point on a 1–5 rubric. A causal-reasoning lesson didn’t raise grades; it made answers weirder and more diverse. The rubric rewards polish inside the expected box. No follow-up without ChatGPT, so we don’t know if anyone learned. Several authors are OpenAI. → The Decoder

  • Meta’s India VP walks to OpenAI — Sandhya Devanathan, Singapore-based, reporting to APAC MD Kiran Mani, covering SEA and Australia. A decade at Meta. Days after Uber veteran Prabhjeet Singh took India. Timing: a Modi-post restriction apology and a CSAM-ad mess. → TechCrunch

  • Hindi lands on the Open ASR Leaderboard — Voice Arena + Hugging Face. Monsoon en-IN and hi-IN, 4,888 speakers. Half a billion Hindi speakers, and the multilingual tab was Europe-only. Public/private splits to stop benchmark fitting. → Hugging Face


  • p-e-w/heretic — Automatic safety-alignment removal. Directional ablation plus Optuna. Gemma-3-12B refusals 97→3, with lower KL than hand-tuned abliterations.

  • THU-MAIC/OpenMAIC — One prompt, a whole course. Multi-agent classroom; v1.0 adds an agent workbench, uploaded materials, 20 skills.

  • K-Dense-AI/scientific-agent-skills — 163 science skills. Any agent on the Agent Skills standard, not just Claude. There’s a BYOK desktop co-scientist now.


▶️ YouTube

  • Cancel your subscriptions, Ox-Alpha is here! (GLM 5.3 Flash) — Matthew Berman. The anonymous OpenRouter chart-topper Ox-Alpha is Z.ai’s GLM-5.3-Flash: 320B total, 18B active, preview traffic served entirely on Chinese chips. Coding benches near Claude Opus 4.8, DeepSeek-like prices. → YouTube

💬 Community

  • Tencent Hy4 Preview — 770B total, 49B active, 1M context, 1.56TB on Hugging Face. Big jump from Hy3 (295B). reasoning_effort is binary: high or no_think.

  • Robot comment classifier — Is this code comment a human or Claude? Em dashes, trailing periods, “so” and “whether.” ~80% accurate — and other labs’ models look human to it.

  • The turbulent AI era is here — Bill Gates on the fork in the road. One lap on Lobsters.

COMMENTS