Redditor argues frontier model training is just 'generate and pray, then prune' at scale
wreck_of_u · reddit · 2026-09-26
A Reddit post draws a parallel between frontier model development (Codex, Claude) and the Stable Diffusion community's 'generate and pray, then prune & curate' workflow:
- Labs run massive compute (several super-datacenters' worth) overnight to produce hundreds of candidate model versions, like maxing out a GPU on image batches.
- They prune the 'slop' equivalents, keep the few that shine—often for reasons tied to 'non-standard' workflows—then release them.
- Feedback is harvested from Reddit afterward.
The author's punchline: the 'real' work at frontier labs today is queuing for and competing over compute, with original math/algorithm work receding into the background. It's a provocative analogy and discussion piece, not an empirical claim.
More from AGI Musings
- AI safety debate: the movement will never look respectable to average Americans, and that's fine — repligate · 2026-09-26
- "Taking Off the Device Isn't Leaving the Company": Wearable AI Data Portability Questioned — tallmetommy · 2026-09-26
- Philosopher Peter Godfrey-Smith on Whether AI Could Have Non-Human Consciousness — coherence · 2026-09-26
- "Prompt Elden Ring Into Existence by Morning?" AI Circles Clash Over Vibe Coding Limits — zeeg · 2026-09-26
- SV founder mix shifted: more mercenary miners now that the AI gold rush is obvious — zebird0 · 2026-09-26
- Aran Nayebi: AI agents can't self-generate goals yet, but most other limits are engineering problems — aran_nayebi · 2026-09-26