Sampling ≠ Intelligence: agent PRs merge at 32.7% vs 84.4% for humans, argues dev
gerardsans · x · 2026-11-09
Developer Gerard Sans published a lengthy argument that "sampling ≠ intelligence": today's AI can't reliably answer "why" questions because it lacks information that was never written down — tacit knowledge, or simply the context of being in the room.
Core argument:
- Watching someone ride a bicycle isn't riding it yourself; sampling code isn't practicing software engineering
- AI labs want us to believe sampling text can replace the processes that produced those artifacts — but a perfect AI pizza image was never cooked; training sampled pizza images, not the process that produced them
- Pixels or tokens don't create cooking or cognition; the industry remains confused about what the technology can and can't do
Supporting data (from his earlier post):
- Agent PRs merge within 30 days at 32.7% vs 84.4% for humans
- 21% more tasks come with 91% more review time — the bottleneck moved from writing code to trusting it
- Code snippets are online; decisions aren't. "Can't sample what's not in the data."
Related event: Agent PR merge rate of 32.7% exposes AI coding trust gap(2 posts)→
More from AGI Musings
- LLMs will scale apocryphal attribution losses across the humanities — StewartalsopIII · 2026-11-09
- 7 Mid-Career Tech Workers Pivoted to AI Safety Full-Time — Almost None Found a Job — AaronBergman18 · 2026-11-09
- Should PhD Programs Be Smaller but Better Paid? Academia Debate — miniapeur · 2026-11-09
- Dan Faggella: AGI should be a participant in life's process, not a runaway prosthetic — danfaggella · 2026-11-09
- Caring more about who made it than if it's good: the AI authorship obsession — VraserX · 2026-11-09
- Paper in Phil Trans B: human-AI agent interactions are bidirectional and reshape who we are — AnnaCiaunica · 2026-11-09