Korva is a scroll of the internet’s best ideas.
AI's boy-who-cried-wolf problem
Dramatic statements about white-collar work going away or AGI arriving by 2027 feed a frenzy. People trying to make sense of the world get pulled into it, and then, time and time again over the last few years, it hasn't turned out that way.
Over time that gap between expectations and delivery has produced a boy who cried wolf dynamic. For people deep in tech, the progress looks different: every release feels like a step change, the capabilities being unlocked keep growing, and the technology is evolving faster than ever.
Someone outside tech hears that AI is about to become super intelligent and reach a singularity. Then a new release comes out, it still hallucinates, and it turns out to be unable to do a lot of things. That person is left wondering what the tech people are talking about.
Why do AI predictions keep following a boom and bust?
Gartner analyst Jackie Fenn introduced the hype cycle model in 1995 to describe how new technologies are received. It traces a peak of inflated expectations, followed by a trough of disillusionment, and then a slower climb toward practical use. Many technologies pass through the whole curve while steadily improving underneath it. Gartner has published hype cycle charts for artificial intelligence topics in many years since.
Did the dot-com bubble follow the same pattern?
In the late 1990s, forecasts said the internet would remake retail, media, and communication. The Nasdaq Composite peaked in March 2000 and then lost roughly 78 percent of its value by October 2002. Much of what was predicted did happen over the following two decades, through companies such as Amazon and Google. The Nasdaq did not regain its March 2000 peak until 2015.
Why does AI still hallucinate after each new release?
Language models are trained to predict plausible next words, which means a fluent answer can be produced even when the model has no solid basis for it. OpenAI researchers argued in a 2025 paper that common training and evaluation methods reward guessing over admitting uncertainty. Many benchmarks score only accuracy, so a confident guess earns more than an honest abstention. Their paper proposed scoring that penalizes confident errors more than it penalizes expressions of uncertainty.
More from Sourcery with Molly O'Shea
