Reasoning model reports I recommend reading: 2025-01-22 - DeepSeek R1 - arxiv.org/abs/2501.12948 2025-01-22 - Kimi 1.5 - arxiv.org/abs/2501.12599 2025-03-31 - Open-Reasoner-Zero - arxiv.org/abs/2503.24290 2025-04-10 - Seed-Thinking 1.5 - arxiv.org/abs/2504.13914 2025-04-30 - Phi-4

It's good for academics to publicly experiment with new AI tools, but important to report both successes and failures when doing so. Audience capture incentivizes only doing one or the other, which is part of the reason the information environment around capabilities is so bad.

One’s breakthrough is another’s follow-through.

It has learned the subtle art of being attentive and available while providing little service.

Positive thinking

Francesco Orabona Let me explain why "given access to the same data, prefer the tighter bound" is not a path to "explaining generalization". I'll tie thi back to the methodology "In Search of Robust Measures" paper. If your goal is estimating risk for the classifier you've learned, then you