2026-06-05

1. Anthropic "When AI Builds Itself" — Calls for International Coordination on an RSI Brake Pedal (score: 2.33)

• Blog post by Jack Clark (co-founder) + Marina Favaro (Anthropic Institute): as of May 2026, Claude authored 80%+ of Anthropic's codebase — a sharp jump from the single digits before the Claude Code launch (2025-02). Clark: "100% is possible within 2 years."
• Core warning: recursive self-improvement (RSI) — the threshold at which AI autonomously designs and develops its own successor versions — "could arrive before most institutions are ready."
• Proposal: not a demand for an immediate halt. Rather, a Cold War nuclear arms-control-style international coordination mechanism among AI companies — installing a "brake pedal" in advance so it can be pressed if needed.
• Why it matters: The 80% figure pulls the RSI debate out of abstract forecasting and into current operational data. It completes the narrative arc of Jack Clark's "60%/2028" forecast (5/7) → Oxford lecture "autonomy is not inevitable" (5/20) → "therefore we must build the brake now" (6/5). The most explicit case to date of a frontier lab directly calling for public policy coordination.
• → AI Alignment (source) (CNN) (BBC/Yahoo)

2. OpenAI ChatGPT Dreaming V3 — Complete Overhaul of the Memory Architecture (score: 2.24)

• Background synthesis: automatically analyzes conversations to update memory even when the user does not explicitly request saving. Replaces the existing "list of saved memories" UI with a single unified memory base.
• Time awareness: "planning a trip to Singapore in July" automatically updates to "visited in July 2026" after the trip.
• Performance: factual recall rate 41.5% (2024) → 82.8% (2026); 5× improvement in compute efficiency → first launch on the Free tier.
• Rolling out to US Plus/Pro starting June 4, with Free and global availability expanding sequentially.
• Why it matters: A consumer-layer solution to the agent memory continuity problem. A shift in pattern: "user asks to be remembered" → "AI decides for itself what to remember." The 5× compute savings enabling a Free rollout is evidence of architectural innovation.
• ⚠️ Name caution: distinct from Anthropic Managed Agents' "Dreaming" (self-improvement of an agent's procedural memory).
• → Agents (LLM Agents) (source) (OpenAI)