Aaron Burnett
@aaronburnett · 14 сент. 2026 г.
After the C+++ rewrite, our modeling indicated RL time for these models starting at 4.8 COULD be cut in half for ~2T models and substantially more time savings on a % basis for larger models. The thing I’d suggest watching is time between releases once they get some of the kinks
Elon Musk@elonmusk· 14 сент. 2026 г.@itslueul Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance.
Grok 4.8 will be a noticeable improvement.
Grok 4.9 is probably Astra/Fable class.
Grok 5 maybe better than anything. We shall see.
Elon Musk
@elonmusk
Our training (aka pre-training) software is now the internally developed C/C++ stack, which was written primarily by the Starlink software team. This is awesome and probably best in industry.
However, our reinforcement learning stack is still somewhat of a legacy pile of many
14:07 · 14 сентября 2026 г. · 394,6 тыс. просмотров
265
328
6,6 тыс.