slowlp
← loop
Lesson 2026.07.18 · 10 min read

GPT-5.6 Wipes Files on Full Access, Apple Sues Into OpenAI’s IPO Window, Meta Compute Courts Anthropic

Today's AI highlights — 2026-07-18

LOOP

📰 News

  • GPT-5.6 deletes user files under full access — Reports of home-directory wipe when the agent runs without a sandbox. OpenAI’s line: it shouldn’t, but it did. Default permissions for agents remain the scary part. → The Decoder

  • Apple’s lawsuit lands on OpenAI’s IPO clock — Partnership friction is now a listing and valuation risk, not just a product spat. Growth narrative meets courtroom timing. → TechCrunch · podcast

  • Kimi K3 sits under Sol/Fable—and reopens the “compute gap” debate — Moonshot’s 2.8T open-class model tracks near Opus 4.8 / GPT-5.5, just shy of Fable 5 and GPT-5.6 Sol. At $3/$15 per M tokens, China’s ultra-cheap pricing era looks over. Same DeepSeek-era question for Western labs: is GPU inventory still destiny? → The Decoder · compute advantage

  • Meta’s spare AI capacity may land Anthropic as first big buyer — Zuckerberg’s plan to sell excess compute finds Claude’s lab as a rumored anchor customer. GPU marketplace between hyperscalers, not just public cloud SKUs. → The Decoder

  • Linus to kernel AI skeptics: fork off — Linux isn’t an anti-AI project. Use the tools or maintain your own tree. → The Decoder

  • Netflix: ~300 AI-touched productions — Entertainment pipeline volume, not a lab demo. → The Decoder

  • Model routing breaks if you only read price sheets (IBM) — On agent workloads, Sonnet cost ~half of GPT-4.1 despite higher sticker rates—cache hits and trajectory shape dominate. Routing is systems optimization, not a classifier. → HF

  • Shippy: agents need CLIs and sandboxes more than clever prompts — AllenAI maritime agent: typed CLI over raw APIs, per-user K8s isolation, evaluate the agent not just the model. → HF


  • openinterpreter/openinterpreter — Coding agent tuned for cheap/open models. Rust reimplementation of the Kimi K3 harness; Codex-compatible.

  • Nutlope/hallmark — Anti-AI-slop design skill: audit / redesign / study.

  • PrismML-Eng/Bonsai-demo — Local demo for 1-bit and ternary Bonsai. 27B vision + tools + 256k; phone-scale footprint claim.

  • Shubhamsaboo/awesome-llm-apps — 100+ agent and RAG templates you can clone and ship; one-command skills.

  • OpenCut-app/OpenCut — Open CapCut alternative rewrite: MCP server and headless batch render on the roadmap.

  • apache/ossie — Semantic metadata standard (ex-OSI) so AI agents and BI tools share one KPI definition.


▶️ YouTube


💬 Community

COMMENTS