Looks pretty exponential to me [1]. From a fully independent, non-profit research group.

[1] https://metr.org/time-horizons/

Release date seems like a terrible x axis with how much more compute they are using. Not to mention while I like what METR is trying to measure, it is an uber specific metric. And frankly, me just complaining, they’re prompts I feel do most of the work for the AI. I’ve never gotten as detailed instructions as they give the AI for the task