H-Matched: A website tracking shrinking gap between AI and human performance

[mrconter11]

https://h-matched.vercel.app/

[mrconter11]

Hi! I wanted to share a website I made that tracks how quickly AI systems catch up to human-level performance on benchmarks. I noticed this 'catch-up time' has been shrinking dramatically - from taking 6+ years with ImageNet to just months with recent benchmarks. The site includes an interactive timeline of 14 major benchmarks with their release and solve dates, plus links to papers and source data.

[MrCoffee7]

I did not see anything about how the estimated time to human level performance of an AI was estimated and what the error bounds might be on those estimates?

[mrconter11 → MrCoffee7]

Thank you for your reply. The time to human level is simply the time it took between the initial release of the website to when an AI system reached human level performance for that benchmark. :) I am in the process of adding sources for all "solved" dates... Here is for instance source for the Winograd challenge human level performance:

https://arxiv.org/pdf/1907.10641#:~:text=The%20best%20state%...).

:)

[MrCoffee7 → mrconter11]

Thanks for your response.