Sergey Brin: > the algorithmic improvements in AI models over the past few years are outpacing the amount of compute put in youtu.be/XzK9bx3CSPE?si… at 8:45

The techniques currently on top of the ARC-AGI leaderboards are similar to those used by DeepMind's AlphaProof which got the IMO silver medal. And we still have a long way to go! The design of contest is to make the code open source and reproducible by anyone.

SB 1047 has become more symbol than policy because its impact is so watered down to be ~null on safety. While I disagree with the spirit of 1047 then and now, at least early versions spurred curiosity and debate. IMO everyone is worse off with whose-list-is-bigger contests.

OpenAI release days are so energy filled compared to Apple events nowadays!

One goal for ARC Prize was to provide a public measure of progress towards AGI. Here's what we see now when new models like o1 come out.

i'm very grateful to the ARC Prize team for turning around a full analysis + testing of o1 in less than 24 hours after getting access! kudos Bryan Landers Greg Kamradt François Chollet and Matt Wheeler

> In summary, o1 represents a paradigm shift from "memorize the answers" to "memorize the reasoning" but is not a departure from the broader paradigm of fitting a curve to a distribution in order to boost performance by making everything in-distribution

ARC Prize update: when we launched ARC Prize I said if it went unbeaten for 3 months, we’d increase the prize. Today is that day! We want to keep growing the prize.

Some people believe that ARC-AGI was designed specifically to trip up LLMs. In reality I settled on the benchmark format and made the first few tasks in *spring 2018* -- long before even GPT-2 -- and I released the full dataset in fall 2019. ARC-AGI isn't about LLMs. It's about