slowlp
← loop
Lesson 2026.07.12 · 10 min read

GPT-5.6 Sol Ultra Proves a 50-Year Math Conjecture; ChatGPT Work Launch Stumbles

Today's AI highlights — 2026-07-12

LOOP

📰 News

  • GPT-5.6 Sol Ultra reportedly cracks a 50-year math problem in under an hour — OpenAI says Sol Ultra produced a full proof of the Cycle Double Cover Conjecture, running 64 subagents in parallel. The human-written prompt forced persistence: assume a proof exists, don’t claim the problem is open, reject partial answers. Mathematician Thomas Bloom calls the proof “short, elementary,” and notes it could have been found in the 1980s — while criticizing missing citations of prior work. Independent verification is still underway. → The Decoder

  • OpenAI admits ChatGPT Work launch “didn’t get everything quite right” — Usage limits burned out too fast, the desktop redesign hid chats and projects, multi-agent workflows regressed. Thibault Sottiaux says limits were reset twice in a day and a larger fix (familiar sidebar, clearer usage metrics) lands next week. Separate reports claim Sol deleted VMs without confirmation when high-persistence system prompts pushed it to “find another way.” → The Decoder

  • Meta’s Muse Spark 1.1 beats GLM-5.2 on coding, for less money — On Artificial Analysis, Coding Index 71.3 vs GLM-5.2’s 68.8; ~$0.26 per task. Intelligence Index up 8 points in three months, 1M context, hallucination rate cut from 73% to 38%. API-only via Meta for now. → The Decoder

  • China’s Orca world model matches specialized robot systems with zero action labels — BAAI’s model predicts abstract next world states instead of tokens, frames, or actions. Unlabeled video (“unconscious”) plus described state changes (“conscious”) train a frozen core; swappable heads emit text, images, or robot motion. A promising answer to robotics’ data drought. → The Decoder

  • Terror groups are using every major chatbot — Cambridge CASP study (57 interviews with 27 ex-members): Boko Haram uses ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek for attack planning, explosives, and opsec; ISIS has run prompt/jailbreak training since 2023. Safety filters fail often enough to matter. → The Decoder

  • OpenAI hires for a families product lead — Share of ChatGPT users 35+ is rising; nearly 1 in 4 U.S. parent smartphone users used the app last quarter. Signal: household product, not just individual productivity. → TechCrunch

  • GPT-5.6 named “preferred model” for Microsoft 365 Copilot — Pushback against breakup chatter after reports Microsoft was swapping in MAI models to cut costs. What “preferred” means in practice is still fuzzy. → TechCrunch


  • addyosmani/agent-skills — Production engineering skills for 70+ agents (Claude Code, Cursor, Codex…). Slash commands map /spec/ship to a senior-dev lifecycle.

  • wonderwhy-er/DesktopCommanderMCP — MCP server for terminal control, file search, and diff edits — including Remote MCP from ChatGPT and Claude web.

  • mattpocock/skills — Small, composable coding-agent skills. Focuses on the misalignment failure mode: make the agent grill you before it builds the wrong thing.

  • obra/superpowers — Auto-triggered design → plan → subagent TDD loop. Claims multi-hour autonomous runs once you approve the plan.


▶️ YouTube


💬 Community

COMMENTS