slowlp
← loop
Lesson 2026.07.16 · 8 min read

GPT-5.6 Sol Cracks a 30-Year Stats Conjecture, GPT-Red Outruns Human Red Teamers

Today's AI highlights — 2026-07-16

LOOP

📰 News

  • GPT-5.6 Sol disproves a 30-year stats conjecture in ~90 minutes — A UPenn researcher used Sol Pro to show Benjamini–Hochberg FDR can miss its target on correlated normal data. GPT-5.5 failed after 20+ hours. → The Decoder

  • OpenAI’s GPT-Red beats human red teamers — Self-play attacker finds prompt-injection-style flaws in 84% of tests vs 13% for humans. Sol is 6× more resistant to direct injections than four months ago; ~3.8% of stronger attacks still land. → The Decoder

  • Codex encrypts agent-to-agent instructions — Inter-agent delegation is now opaque in logs. Mandatory on Sol/Terra. Privacy move or rival-training block? Some flaky handoffs reported. → The Decoder

  • $230 Codex Micro keyboard amid Apple hardware fight — Limited Work Louder collab: Agent Keys, reasoning dial—a desk command center for agent fleets. Screenless ChatGPT device still in the rumor mill; Apple’s suit continues. → TechCrunch

  • Bonsai 27B: open reasoning that fits on an iPhone — PrismML compresses Qwen3.6-27B to ~3.9 GB (1–2 bit) with ~90% quality. Bet: local agent loops, zero cloud token cost. Apache 2.0. → The Decoder

  • Anthropic + Blackstone’s Ode: implementation, not just models — $1.5B JV, ~100 forward-deployed engineers, Claude-first, pilot-to-production gap. Same category as OpenAI’s Deployment Company. → TechCrunch

  • Inkling + Meta layoff suit — Thinking Machines’ ~975B MoE multimodal (1M context) hits HF. Meta workers claim AI systems ranked people for May’s 8k cut. → HF · The Decoder



▶️ YouTube


💬 Community

COMMENTS