Limits to narrow LLM complementarity: why 'taste is the bottleneck' won't hold

zetalyrae · x · 2026-09-06

An osmarks.net essay argues the recent 'taste is the bottleneck' thesis is no more valid than earlier 'harnesses' or 'prompt engineering' bottleneck claims. For humans to be economically relevant complements to LLMs, they'd need to offer useful skills cheaper than LLMs after coordination costs — which the author argues are very large. Production LLM instances spend most of their existence never interacting with live humans, pretrained on 5 orders of magnitude more text than a human reads, with 6+ month data lag. LLMs lacking human skills doesn't mean humans can usefully complement them, nor that human data is especially valuable.

Original post →

More from AGI Musings

AGI Musings channel →