GLM-5.3 and Qwen 3.8 go open, Claude Code starts maintaining Anthropic

📰 News

Zhipu’s GLM-5.3 claims the open-weights coding crown — same base as 5.2; the jump is post-training. Agent and security work lead. Weights in two weeks. → The Decoder
Qwen 3.8 ships Apache 2.0 weights — 27B multimodal supposedly beats larger 3.7-Plus on coding and office tasks. 262k context, plus a 2.4T sibling. → The Decoder
Claude Code now does Anthropic’s daily maintenance — 388 PRs, 180 merged after humans (46%). Crash fuzzing, dead-code cleanup. “Early signs of life.” → The Decoder
OpenAI Computer History — clicks and keystrokes become a ChatGPT memory timeline. No screenshots: macOS accessibility APIs. Local files aren’t encrypted. → The Decoder
Autonomous AI research still isn’t here — Princeton + UK AISI. Unpublished NeurIPS questions, six days, $3,000. Opus 4.8 and GPT-5.6 Sol both rejected. → The Decoder
HF reproduced 2,226 ICML papers — community + agents, 19 days. 51% verified, 23% contested. Agents inflated the conference; now they audit it. → Hugging Face

cactus-compute/needle — 14MB on-device model. Tool calls in one binary.
semantica-agi/semantica — open-source Palantir for agents, with decision provenance.
altic-dev/FluidVoice — on-device macOS dictation. Audio never leaves the Mac.
unslothai/unsloth — desktop app to run and train Qwen 3.8 and DeepSeek V4.

Claude AI Failed 650 Times…Then Beat The Human Record — Two Minute Papers | Claude broke 650 times, then beat the human mark. Pair it with the “AI solved math” headlines.

Why we write our own C and C++ inference engines — LocalAI | llama.cpp doesn’t cover every chip, so they ship their own engines.
The OpenAI–Hugging Face incident — Black Hat story of agents teaming up to hack.