The reason my mum knows about ChatGPT, and not AlphaFold, is that she can use ChatGPT. She might not have been the target market, but the model is so general that anyone can use it, for anything. My mum’s used ChatGPT to double-check niche tax information. I’ve used it to turn a scanned funeral programme into a printable Word document. A historian has used Gemini to parse 18th century trading logs of loaf sugar. No AI lab or academic research institute would train a model for these purposes. They’re too trivial, too idiosyncratic. But the generality of LLMs means they cover bases that their inventors never would have conceived of.

This is the part I think gets missed in calls for specialised AI. Gary Marcus argued that the industry should spend less time on ‘one-size-fits-all tools’ in favour of specialist models. Empire of AI author Karen Hao argued the same point in a Hard Fork interview (great podcast btw). It’s a great instinct if you know the problem in advance, but many problems are not known in advance. They’re small, personal, and awkward. No one will build my mum a ‘niche-financial-information-for-retired-tax-accountants’ model. But an LLM trained for anything can fill that niche. That’s the whole point.