Nathan Lambert

Context: Still open — the forecast resolves at the end of 2026 — but this cycle produced the first hard reading against its named metric, and it runs against the prediction. Taking the lag as the time from a closed model first reaching a score to an open-weight model matching it, on the Artificial Analysis Intelligence Index v4.1: on 23 July the best downloadable model was GLM-5.2 at 51 (weights 16 June), a score a closed model first passed when GPT-5.5 reached 55 on 23 April — about eight weeks. On 27 July, with Kimi K3's weights published at 57, the crossing date is Claude Fable 5's 60 on 9 June — about seven weeks. Both readings are roughly a quarter of the predicted 6-9 months. Three reasons this does not yet settle it: the metric is a year-end state, not a July snapshot, and a closed-model release with no open answer would widen the lag again; Lambert's own caveat, recorded with the claim, is that this index may overstate open-model closeness because public benchmarks are easier to overfit; and the UK AI Security Institute's separate measurement of open-weight cyber capability puts the lag at 4-7 months, at or just below the bottom of his band. Recheck at resolution.

Context: Independently corroborated: MIT Technology Review reported the same week that Chinese open models had overtaken Llama in popularity, and subsequent UK AISI and Artificial Analysis measurements consistently placed Chinese models atop the open-weight rankings through mid-2026.

Context: Borne out within a year: by August 2025 MIT Technology Review reported Chinese open models (DeepSeek, Kimi, Qwen) had become more popular than Llama, and by mid-2026 the independent record — UK AISI's evaluations and the Artificial Analysis index — showed every leading open-weight model was Chinese until Thinking Machines' July 2026 entry.