Colab here! colab.research.google.com/drive/1YkcjQ...

This paper looking at financial impacts is damning, incl. an estimated 28% rise in bankruptcies for states which legalised sports gambling papers.ssrn.com/sol3/papers....

That 500ml estimate is one of a few. Mistral estimated 10ml for 100 tokens, Google said median 0.3ml per-prompt: the 17B for the US is still only a fraction of a percent of your national water use. Why focus on water when electricity use is orders of magnitude more significant?

You're conflating two things. The emissions/electricity impact is real and significant. Water use is irrelevant. 17B gallons sounds like a lot, but global water use last year was 4.3 trillion cubic metres, around 10^15 US gallons.

Hi Hendrik, is there an archive of these visualisations? I'd love to compare eg. ICML 2024 vs 2025, either via the viewer or raw data if it's available.

It's interesting that this is a fairly local take. Stanford reported that 36% of US citizens are optimistic about AI vs 83% in China. BlueSky broadly has an anti-AI sentiment, but that doesn't reflect global attitudes.

On the other hand, Deepmind and Google are working on a much wider range of scientific problems that would benefit from powerful AI than their competitors.

Imo it's a consequence of how hard it is to measure task performance - benchmarks give you a very narrow view of performance.

I think the o1 direction is an interesting result of this. Fields with easy verification are simple fit for plugging in traditional RL. How do you verify a good suggestion for writing feedback?

If only there was more PRS on Spotify!

Looks a bit like the reasoning traces out of r1/QwQ