Elon Musk after finding out he can shave $4T instead of the $2T he initially estimated due to the massive fraud that was being committed by the government
Alpha Geometry 2 has now surpassed an average gold medalist in solving Olympiad geometry problems.
To achieve this, they first extend the original AlphaGeometry language to tackle harder problems involving movements of objects, and problems containing linear equations of angles, ratios, and distances.
This, together with other additions, has markedly improved the coverage rate of the AlphaGeometry language on International Math Olympiads (IMO) 2000-2024 geometry problems from 66% to 88%.
Together with further enhancements to the symbolic engine and synthetic data generation, they have significantly boosted the overall solving rate of AlphaGeometry2 to 84% for all geometry problems over the last 25 years, compared to 54% previously. AlphaGeometry2 was also part of the system that achieved silver-medal standard at IMO 2024
In just one year from silver to gold medal.
Absolutely amazing.
The Pantheon, built over 2,000 years ago in Rome, takes its name from Pan and Theon meaning all and divine
Its massive 9-meter oculus serves as a gateway between the human and the sacred
Yet, rain rarely falls inside—thanks to the famous “chimney effect”
https://t.co/KXkki0syd2
I love Mike Benz to death, but Igor Kolomoisky getting $11 billion through USAID to pay Chinese bagmento do biosurveillance on every man woman and child in the United States is a bigger story than USAID paying transgender rap bands to help in the overthrow of Bangladesh.
Intuition on why RL generalize and SFT memorize, using GRPO as an example:
If a question is too hard, no output gets reward, there's no advantage and no gradients
So it's effectively an online data selection method:
only update on data that's not too easy, and not too hard
Was just playing with the new Replit AI app creator, and it's honestly unbelievable.
I made an Anki deck with all the idiosyncrasies I wanted that most people wouldn't care about in ~10 minutes. This will give less technical people much more power to be creative.