This explorer is only available on a desktop device, or a wider window. The chart above is a static version of the same data.
This explorer and the underlying dataset were created with the assistance of Claude and ChatGPT.
Each point is a task, with an estimate of how many FLOPs an AI system needs to do it and how long it takes a person to do it. Where available, the dollar price for the AI to do it is also included. The estimates are very crude, and the uncertainty toggle shows a plausible range for the relevant values.
The cheapest-per-task toggle shows, for each task, only the model that completed it at human level or better for the fewest FLOPs.
This explorer and the underlying dataset were created with the assistance of Claude and ChatGPT. Further information is in the write-up and the full dataset on GitHub.