I think that the METR benchmark is fundamentally flawed as it is biased towards tasks for which failure and verification are cheap. For that reason I regard any post with METR charts as bait. But at the same time, I believe that the IBM analogy is valid

My For You feed is basically people constantly talking about a gc I’m not part of. If I wanted this kind of experience, I’d be on Instagram. Yet another reason not to use For You. And probably not to use this account. Zero information on ML, finance alt is more interesting atm

I really like the concept but each time I considered a graph DB, I ended up building my graph with postgresql tables. It just works at my humble level (even with pretty nasty recursions). It may have been optimal to actually use a graph DB but I’ll never know

I don’t think so

A note from my travels in France this Summer. Unusual/surprising things to report: - people now stand on the right in escalators in Paris like in London. Uncanny - a LOT of wind turbines were not operating despite windy conditions, clearly indicating overcapacity

Just realized that, 2 years ago, I bought a M2 Ultra, betting on Moore’s law to allow local frontier models in early 2025. And it absolutely did not happen. Far from it. This objective seems more remote than ever. The goal posts are moving in the wrong direction lol

Good morning with a banger resource on vLLM