Pretraining is elegant science, posttraining is cowboys yelling hyperparameters
himanshustwts · x · 2026-10-17
tszzl's viral post contrasts the two halves of AI research: pretraining as "elegant science" done by mathematicians and distributed-systems engineers, while posttraining is "hair-raising cowboy research" where people chug Diet Coke and yell hyperparameters across the room — culminating in the cursed checkpoint name "gpt-9-final-v320-restart4". Quote-replier adds the checkpoint name itself is calling people names.
More from Fun
- Will Chen: 'Twitter is the Bloomberg of AI research' — willcb · 2026-10-17
- Burkov's one-line model ranking: 5.5 Haiku far below 6 Luna, on par with 3.8 Flash — burkov · 2026-10-17
- The "does Claude have a soul" 2x2 meme chart makes the rounds — sergeykarayev · 2026-10-17
- 'EU regulates flying cars' post is a meme mocking preemptive tech regulation — cgarciae88 · 2026-10-17
- Beff Jezos says his brain hit token limits: 16-hour days, 80-hour weeks — beffjezos · 2026-10-17
- AI Industry Worker Says Nobody He Knows Uses xAI's Grok Build — SimonBalmain · 2026-10-17