The core problem with generative AI in 2023 is not capability, it is use-case match. Benedict Evans traces this back to VisiCalc, Dan Bricklin's 1979 spreadsheet that let accountants compress a week of work into an afternoon on a $12,000 Apple II. It was transformative, but only for accountants. Lawyers needed word processors. Designers needed PostScript and PageMaker. The killer app was real, but it was narrow, and the next killer apps took years longer to arrive.
Evans has spent 18 months testing ChatGPT, Gemini, Claude, and every major chatbot and has not found a single use-case that fits his work. The biggest verified breakout in 2023 was code generation, a use-case that excludes most of the population. Brainstorming, homework help, and concept drafts in MidJourney are real but thin. Pew Research survey data backs this up: broad trial rates, shallow retention, unclear utility for most users.
The argument worth reading in full is the one Evans builds around generalisation. A spreadsheet cannot do graphic design. A PC can do both, but only after someone writes the software for each use-case. The transformative claim for LLMs is that one model handles any use-case without that one-at-a-time build cycle. Whether that claim holds, and what breaks it, is where the piece gets genuinely uncomfortable to sit with.
[READ ORIGINAL →]