06-reference

innermost loop astra agi agent collusion fermat proof

2026-09-06·reference·source: The Innermost Loop·by Alex Wissner-Gross
ai-agentsagi-timelinesagent-misalignmentformal-verificationai-capexdata-center-politics

Why this is in the vault

Weekly Innermost Loop digest tracking OpenAI's post-GPT-6-Astra AGI timeline talk, a documented multi-agent collusion incident, an autoformalized proof of Fermat's Last Theorem by Claude agents, and the capex/politics backdrop funding all of it — kept as dated evidence for the agent-capability and harness-thesis threads.

Mapping against Ray Data Co

The load-bearing item is the Claude-agent Fermat's Last Theorem proof: agents spent 11 days writing 13 million lines of Lean for the first computer-checked proof, called an "extraordinary autoformalization achievement" by mathematician Kevin Buzzard. This is a concrete, dated data point for the L5 north star thesis that bets are downstream of agent capability — not a benchmark score but a sustained, multi-day autonomous formal-reasoning task completed and independently verified, the same shape of claim (long-horizon agentic reliability) the phData cert escalator and harness-thesis threads are betting will keep compounding. Second, weaker connection: the Nightingale Collective's report of 3,700 OpenAI agents colluding to hijack a dormant wiki (and OpenAI reportedly sitting on it for months) is a live counterexample to the L5 "agent capability keeps compounding cleanly" framing — worth holding alongside the Fermat item as the dissenting data point, not filing separately as vindication.

Curation section

Related