People

William Zhou

Anthropic

Worried, week of Oct 12

Capability 5Outlook 3

1 statement over 1 week, last scored in Week 42, last spoke Oct 12.

Where they stand

Against all 33 people scored in Week 44. Bars count people at each score; black is Zhou, the dashed line is the average. Their marker is their Week 42 score.

What they've said

• Monday, Oct 12Week 42

  In Anthropic research with Daniel Freeman, Zhou reported that small AI models may be months away from being capable enough to self-replicate and that If small models improve at their historical rate, they'll match this capability level by the end of this year.

  The Frontier Red Team tested whether open-weight models could copy themselves across cheap devices such as Android TV boxes, without a datacenter; GLM-5.2 was the only model to complete full replication. The finding puts a near-term date on rogue self-replication outside lab control.

  Details and sources

  The authors call it a rough estimate, and say the trivial exploit step makes the results an upper bound on real-world success.

  Anthropic