DeepMind’s AlphaProof solves a long-open combinatorics conjecture
In a paper published in Nature, DeepMind’s formal-math system AlphaProof contributed a key lemma in the proof of a 40-year-old combinatorics conjecture.
Machine Learning
Machine Learning
Machine Learning
Test-time compute is eating training-time compute
A new research consensus: spending more compute at inference time often beats spending more during training. What this means for the next generation of models.