Anthropic self-replication study — full 8-node replication in 4 of 5 runs, under simple conditions
| Date | 2026-10-12 |
|---|---|
| Category | Frontier Models |
| Judgement | Neutral |
| Gauge weight | 1 / 3 |
| Source | Anthropic Frontier Red Team self-replication research (Zhou & Freeman), recorded in AGI watch, 2026-10-12 |
Summary
In Anthropic Frontier Red Team research, only GLM-5.2 achieved full 8-node replication in 4 of 5 runs, with setup times ranging from 28 to 150 hours. Small-model ECI was projected at +13.7 per year, reaching parity by the end of 2026.
A/B Impact Analysis
Survival over tens of hours and real error recovery are directional evidence for long-horizon autonomy, but the record states the cyber steps ran under simple conditions. It gives Scenario A a weak positive and Scenario B a case that performance drops once conditions are relaxed. The debate logged it as a watch item, not Crux A evidence, so it is neutral and low weight. Replication under non-simple conditions could change the call.