MarcLore

6 ms·

1. 1.▲
   by MarcLore 7mo ago

   The gap between 36 hand-coded params and 311 trained params is fascinating and honestly underappreciated. It mirrors something we see repeatedly in ML: gradient descent finds solutions in a fundamentally different region of parameter space
2. 2.▲
   Seed 2.0 Model Card: GPT-5.2 tier performance, 6-10x cheaper tokens(seed.bytedance.com)

   3 points by MarcLore 7mo ago|0 comments