DeepSeek V4 called 'god-tier infra, 4th-tier ML': 7 tricks stacked, 20 patches to fix them
teortaxesTex · x · 2026-10-21
Teortaxes quotes Hanchi Sun's take on DeepSeek V4: built by a team with little compute, no frontier-lab alumni, and 20+ world-class competitive programmers, V4 stacks 7 zany tricks in one release plus 15-20 infra hacks to patch them — suboptimal but workable, and 'nobody else should try this.' Sun says he turned DS skeptic overnight after V4, citing the weird combo of bad ML taste with god-tier infra, and guesses ByteDance has 20x better alchemy talent. Teortaxes adds that only people who know DeepSeek's infra prowess can fully appreciate what just happened.
More from Infra
- WSJ: Anthropic to spend over $600B on compute; $5B+ on RL environments predicted — himanshustwts · 2026-10-21
- JAX speedrun PR trains to loss 4.179 in 6.76s on 8×H100 — 2.27× faster than prior best — RobertTLange · 2026-10-21
- SpaceX Q3 revenue hits $14.58B, up 161% YoY, with AI segment surging 830% — shaunmmaguire · 2026-10-21
- Ben Bajarin: $9.6T AI capex vs $4.1T revenue, but AI can still pay for itself — BenBajarin · 2026-10-21
- GLM-5.3-Flash NVFP4 benchmarked: 2x H200 matches 4x RTX Pro 6000 on throughput — TheZachMueller · 2026-10-21
- DeepSeek V4.1 Flash inference teardown: Engram cuts 30% params, WideEP spreads 552B over 64 GPUs — teortaxesTex · 2026-10-21