Unreleased Anthropic Model Autonomously Advances Riemann Hypothesis Bound via 60 Sub-Agents

Skynet Chance (+0.08%): A model autonomously coordinating 60 sub-agents for 36 hours with minimal human guidance demonstrates open-ended goal pursuit at a scale where humans cannot meaningfully supervise intermediate steps. Delegating correctness checking to AI validator sub-agents shifts oversight from humans to the system itself, a structural precursor to loss-of-control conditions.

Skynet Date (-1 days): Demonstrated long-horizon autonomous orchestration by an unreleased frontier model, alongside parallel OpenAI results, signals a competitive race in agentic research capability that pulls forward the arrival of hard-to-oversee autonomous systems. That an operator without domain training triggered novel research also lowers the barrier to unsupervised deployment.

AGI Progress (+0.07%): Producing new mathematical results on a 150-year-old open problem, formally verified in Lean, is evidence against the claim that LLMs only interpolate training data and cannot originate ideas. Sustained multi-hour planning, hypothesis generation, and self-validation across sub-agents are core general-intelligence competencies.

AGI Date (-1 days): The described pattern of progressively stronger mathematical results tracking more powerful models within a single year suggests autonomous reasoning gains are compounding faster than expected, shortening AGI timeline estimates. Unreleased models at both Anthropic and OpenAI already exceeding public capability implies the visible frontier lags the actual one.