VisiCalc cost $12,000 in today's money, blew accountants' minds in 1982, and still meant nothing to lawyers or graphic designers. Benedict Evans uses that split to frame the central problem with LLMs in 2024: the technology is general, but use-cases are not. Evans has spent 18 months with ChatGPT, Gemini, and Claude and has not found a single task that maps to work he actually does.
The one documented breakout use-case from 2023 is code generation. Everything else, brainstorming, first drafts, concept roughs in Midjourney, serves people whose workflows already had that gap. Pew Research survey data backs Evans up: broad trial, shallow retention. The $12,000 barrier is gone, so the low usage numbers can't be explained away by price.
The argument worth reading in full is not the conclusion but the setup: Evans is building toward the claim that LLMs are structurally different from prior software because one model can theoretically cover every use-case without anyone writing task-specific applications. That is the transformative thesis. The question he is working toward is whether that generality is real or just a demo that hasn't met actual work yet.
[READ ORIGINAL →]