Ryan Greenblatt@RyanGreenblattI've updated towards a higher chance of much faster than trend AI progress in 2026. I also generally updated towards (somewhat) faster doubling times in the next 2 years. I was expecting ~170 day doubling times (5.7 months) and now I expect ~150 day doubling times (5 months).Opens with an observation
Ryan Greenblatt@RyanGreenblattThis is well known, but I do find that Kimi K3 claims to be Claude disproportionately often. I had Opus 4.8 do some (slop?) research about this. I was curious if K3 would ever claim to be Mythos/Fable, but we don't see this. K3 does claim to be 4.5 while claude models don't. 1/Opens with an observation
Ryan Greenblatt@RyanGreenblattMany people seem to be arguing this is implausible because Fable hasn't been out for long. But Moonshot could have obtained unauthorized access before Fable was released via actors with prerelease access (Glasswing) or possibly by hacking Anthropic directly.Opens with an observation
Ryan Greenblatt@RyanGreenblattI tenatively believe it would be good if all AI companies had a policy of doing external deployment before internal deployment, because the largest risks are from internal deployment and external deployment improves visibility. Large internal/external gaps seem dangerous. 1/Opens with an observation
Ryan Greenblatt@RyanGreenblattIt's possible OpenAI and others will use these architectures with only a few recurrent iterations (because that's most performant or due to safety concerns). If so, this would only make things moderately worse. But I expect much more opaque reasoning than this in the future...Opens with an observation