Dwarkesh on MagicAILabs' 50x compute claim: RSI may be less compute-bottlenecked than we think

AccBalanced · x · 2026-09-09

Dwarkesh Patel praised MagicAILabs' claimed pretraining recipe that matches DeepSeek V4 Pro with 50x less compute, speculating it suggests automated AI researchers during recursive self-improvement may be less bottlenecked by compute than naively assumed. The underlying claim remains unverified.

Original post →

More from AGI Musings

AGI Musings channel →