2019 Pruning Experiment Cited to Claim 96% of GPT-5's Weights Are Useless
TinfoilTricorn · x · 2026-09-12
A widely shared thread recommends an 18-minute video arguing trillion-parameter models don't need most of their weights. It cites a 2019 MIT pruning study showing 96% of a neural net's weights can be removed without performance loss — the "winning lottery ticket" subnetwork does the real work.
The author claims this explains why Claude and GPT-5 are mostly "empty scaffolding" and why labs waste 90%+ of their NVIDIA budgets. Note this is a hype-framed take: results from small 2019 experiments don't directly transfer to today's dense frontier models.
More from Infra
- Running out of context on a large codebase: how to auto-handoff long-running local LLM tasks — Developer-Y · 2026-09-12
- What MoE/LLM runs well offline on a 24GB M5 MacBook Air? — itis_whatit-is · 2026-09-12
- Draft model hits ~60 tok/s running Qwen3.8-27B at 131k context on a 16GB GPU — pneuny · 2026-09-12
- DeepSeek v4.1 Flash on-device test: q2 runs at 16 tok/s but tool calls go off the rails — challis88ocarina · 2026-09-12
- smolbenchmark ranks sub-8GB models by speed, tok/J and heat on your own hardware — East-Muffin-6472 · 2026-09-12
- Ex-Tenstorrent exec's startup ai& deploys non-NVIDIA AI hardware in Japan — DavidBennett__ · 2026-09-12