firehose> #llmops

AI Completion Asymptote

AI-assisted work reliably reaches "good" (~80%) fast, but the final stretch to "great" (the last 20%) flattens into an asymptote that more prompting often cannot cross — and sometimes regresses — so the disciplined move is to ship at good-enough and spend scarce human effort only where it actually converts. The source's framing: "AI is getting really good at getting you 80% of the way there, but no matter how many times you prompt it, it won't get you that final 20%… there's a chance you might actually see slight regression trying to make it perfect." The load-bearing distinction is that this is not a licence for slop — "good is not slop" — it is the recognition that returns on prompting diminish sharply near the top, so re-prompting to chase perfection is wasted effort at best and quality-destroying at worst.

Two rules follow. The 80/20 allocation: 80% of a task should take ~20% of the time — lean on AI to reach the asymptote quickly — while the 20% that matters takes ~80% of the time, and that is where you spend human energy to carry good to great. And the judgment of what needs to be great at all: the best builders "know what needs to be good versus perfect" and deliberately build to the minimum viable standard, because most things do not warrant the last 20% (the source cites Peter Levels: only 4 of 70+ launched projects made money — so shipping good and iterating beats polishing things that will fail anyway).

This is the deliberate reconciliation of a real tension with the "quality over quantity" thesis (see Execution Commoditization): quantity-of-slop erodes value, and perfectionism wastes it — both are failures of allocation. The resolution is that quality means choosing which 20% to make excellent, not making everything perfect. The asymptote claim itself is an observation about the current generation's behaviour on the operator's tasks, not a proven law — where the ceiling sits, and whether it is 80% or moves with the model, is not measured here.

Claims


Linked from