Musk: xAI's pretraining runs on in-house C/C++ stack, RL stack still legacy mess
jsuarez · x · 2026-09-14
Elon Musk revealed that xAI's pretraining software now runs on an internally developed C/C++ stack, written primarily by the Starlink software team and "probably best in industry," while the reinforcement learning stack remains a legacy patchwork of third-party systems awaiting a rewrite. A reply recommends PufferLib, a fast 100% CUDA C RL stack for small models, noting its design tricks are useful even for non-LLM workloads.
More from Infra
- Attestable Builds a Verification Layer to Check If Compute Follows AI Rules — ml_hardware · 2026-09-15
- Nina Schick: the power gap becomes an intelligence gap, Europe most exposed — NinaDSchick · 2026-09-15
- Forge launches with Arcee, Microsoft, Vercel to push open weight models to the frontier — inkko44 · 2026-09-15
- Polymarket prices 13% odds of AI bubble bursting by end of 2026, with strict resolution rules — Polymarket · 2026-09-14
- François Fleuret: MiniMax runs well on a $500 16GB 4060Ti when you prompt per the docs — francoisfleuret · 2026-09-14
- The power gap is an intelligence gap: why Europe's grid threatens its AI future — NinaDSchick · 2026-09-14