Investor: boutique data training may beat strip miners amid model collapse risk
StewartalsopIII · x · 2026-09-26
Investor Stewart Alsop reacts to the Oxford/Cambridge "model collapse" research: indiscriminately training on AI-generated web text causes irreversible defects, losing the tails of the original distribution.
- He argues the "data strip miners" hire annotators who likely just use ChatGPT to produce training data — and can lie about it anyway.
- His take: trust-centric boutique data training may be the real play, an idea he had in 2024, though VC money in the space has become "insane — not the good kind."
More from AGI Musings
- Unsafe AI is less useful: safety is a prerequisite for faster adoption — mchorowitz · 2026-09-26
- Sriram Krishnan says DHH's AI speech for engineers will 'be seen as prophetic' — rseroter · 2026-09-26
- Personal agents today: build it yourself, trust a startup, or hand data to an ad giant — neal_lathia · 2026-09-26
- How Law Firms are Becoming AI Firms — Coolnumber11 · 2026-09-26
- Sull: ARGs and transmedia concepts get cheaper and more viable every year — sull · 2026-09-26
- AI is fast but organizations are slow: the bottleneck has shifted to the weakest link — khinhooi · 2026-09-26