Roko: pretraining vs RL compute split shifted massively since late 2025 — 'two different species'
yacineMTB · x · 2026-10-08
Roko Mijic shared a chart arguing that the relative compute split between pretraining and RL has undergone a massive shift starting late 2025/early 2026, claiming AI before and after this divide are "really two different species" — a notable take on the industry's pivot toward RL-heavy post-training.
More from AGI Musings
- Mathematicians' group slams OpenAI's release of 700+ documents claiming solutions to famous problems — pickover · 2026-10-08
- eigenrobot quips as entrenched rent seekers brace for the AI era's upheaval — eigenrobot · 2026-10-08
- From the printing press to LLMs: we're now mainlining synthetic linguistic entities — StewartalsopIII · 2026-10-08
- Toby Ord's interpolation-extrapolation-hyperpolation trichotomy: why AI fails off the hyperplane — tobyordoxford · 2026-10-08
- NY Fed survey: small businesses adopting AI expect higher employment and revenue — robseamans · 2026-10-08
- Lemire: AI clones of Photoshop are the boring part — the real future is entirely new UX — lemire · 2026-10-08