FSD is genuinely one of those “wow” moments for me. It has basically given me back my daily commute — and even some of the longer drives I’ve had this week 🫣. Time that used to be spent purely driving can now sometimes be used to catch up on work, deal with an urgent PR when a training issue needs attention, or simply enjoy the trip itself instead of being drained by the drive. That’s what really matters to me about AI: giving humans time and productivity back. Really amazing stuff from @Tesla_AI . Also very grateful to be working with them to make Grok more useful.

People keep asking me: what's different about optimization in RL? Seemingly nothing — the pre-training stack just works (Adam, even SGD 👀 @saagnikkk). Bringing some answers from my last work (sorry for the delay — been cooking 🚀). We introduce ISO: Isospectral Optimization: an RLVR-native optimization stack. Built on one simple observation, spectral inheritance: RLVR can reuse the base model's spectrum and acquire new behavior purely through the singular frames. 🧩 Offline: ISO-Merger — consolidates RL experts into one model with no data, no rollouts, no OPD. Checkpoints only. ⚙️ Online: ISO-Optimizer — a drop-in wrapper on AdamW / Muon that matches AdamW's accuracy with ~2.7× fewer steps on Qwen3-8B-Base. 📄 🌐 🧵👇