AI Deserves Smarter Criticism
That the models miss on niche facts about a Norwegian writer is about as surprising as Yr.no occasionally getting the weather forecast for Astana wrong.
Thanks to Anne Kat Hærland for coining the “meh” man | Canva
In Aftenposten, writer and programmer Bjørn Stærk declares that language models are “disappointing, frustrating and marginally useful.” His test for judging a model? He asks ChatGPT who he is, and when the model gets it wrong, he concludes that AI does not live up to the hype.
This is an odd yardstick. Language models are trained on trillions of text fragments. Only a fraction of these are in Norwegian, and of those, a microscopic share is about Norwegian writers. That the models miss on niche facts about Bjørn Stærk is about as surprising as Yr.no occasionally getting the weather forecast for Astana wrong.
Ask the models instead about topics where there is plenty of data – the Second World War, say, or photosynthesis – and they are surprisingly good. In areas like programming they have become so good that over 25% of all code at Google is now written with AI.
That said, the newest models actually pass Stærk's test too, when used properly. With instructions like "double-check the facts and cite sources" and models that reason (ChatGPT o3 or Claude Opus 4), most of the errors disappear.
I tested it on myself and got near-perfect results. This despite there being far less training data about me than about Stærk.
I understand the instinct many have to declare AI a hype, especially when faced with what Stærk humorously identifies as “LinkedIn people” – the ones who jump on every new trend with punchy talks and online courses.
At the same time, the AI sceptics mostly remind me of what the comedian Anne Kat Hærland calls the “meh” man. The “meh” man was sceptical of the internet (a passing fad), of smartphones (who wants a computer in their pocket?), and now of AI.
It is a comfortable position. You always come across as the adult in the room, without having to engage in the laborious work of actually understanding what the technology has become capable of.
But it is also a position that tends to age badly. Writing off language models two and a half years after they launched is a bit like dismissing aeroplanes in 1906, three years after the Wright brothers' first flight.
