Thinking through how pretraining vs RL learn

kept by eddie nov 19, 25

Reinforcement learning provides far fewer bits per FLOP than pretraining until models achieve high pass rates, limiting RLVR's ability to learn new capabilities.

2025 letter | Zhengdong
Zhengdong Wang argues that compute scaling—underpinned by Moore's Law and empirical trends—has driven AI progress more reliably than human innovation, despite repeated underestimation.
also on AlphaGo, #scaling-laws, #machine-learning
As Rocks May Think
Coding agents combined with reasoning LLMs have become automated scientists, enabling a golden age where all computer science problems appear tractable through vast inference compute.
also on AlphaGo, DeepSeek
Why I don't think AI is a bubble
The combination of large language models with reinforcement learning creates a path for continued improvement, making the argument that AI progress will plateau unlikely.
also on #reinforcement-learning, #scaling-laws, #machine-learning, #ai
AI infrastructure in the "Era of experience"
As open-source LLMs commoditize due to fierce Chinese competition, companies can build defensible moats through custom RL models trained on proprietary environments using LoRA and GRPO.
also on #reinforcement-learning, DeepSeek
Engram: How DeepSeek Added a Second Brain to Their LLM | rewire.it
DeepSeek's Engram architecture adds conditional memory via N-gram lookup tables to LLMs, improving knowledge and reasoning benchmarks by offloading static pattern reconstruction from neural computation.
also on #scaling-laws, DeepSeek, #machine-learning