The Evolution of AI: From Concept to Reality
The standard story of how AI went “from concept to reality” is a capability story: the models kept getting smarter until, at some threshold, they became genuinely useful and adoption exploded. I think that story is mostly wrong, or at least badly incomplete, and the actual history makes a more useful and more uncomfortable point: usable capability existed well before mass adoption did, and what changed wasn't primarily the intelligence, it was the interface.
GPT-3 was capable of most of what made ChatGPT a cultural moment more than a year before ChatGPT existed. The underlying model wasn't meaningfully less intelligent. What GPT-3 lacked was a chat interface that let a non-technical person type a question into a box and get an answer back, instead of needing to understand prompt engineering through an API playground built for developers. The capability leap that supposedly explains AI's sudden “arrival” had already happened. The adoption leap happened when the packaging caught up to the capability, not the other way around.
Most of the “sudden” breakthroughs in AI's public story are actually packaging breakthroughs wearing a capability breakthrough's reputation.
This matters for anyone trying to predict what happens next, because if you believe the capability story, you spend your attention watching benchmark scores and model releases, waiting for the next threshold to cross. If you believe the packaging story, you spend your attention watching interfaces: what's the next context in which an already-capable model gets put in front of people in a form they don't need training to use. That's a completely different bet, and I think it's the more accurate one, based on how consistently the pattern has repeated: capability quietly outpaces adoption, then a specific interface decision makes the capability legible to people who were never going to read documentation.
I saw a version of this pattern directly on MoneyFlow. The categorisation and insight capability we built with Mistral 7B wasn't the hard-won breakthrough in that project: running a capable open-source model locally was, by 2025, a solved engineering problem if you were willing to put the work in. What actually determined whether a usability tester found the product valuable was how the insight got presented: whether it read like something a person would actually say to you about your spending, versus a technically accurate but robotic output. Same underlying capability, wildly different perceived usefulness, entirely because of packaging.
The “concept to reality” framing implies a single moment where AI crossed a threshold and became real. I don't think that moment exists as cleanly as the narrative wants it to. What actually happened, repeatedly, across different tools and different years, is that capability sat around underused until someone solved the much less prestigious problem of putting it in front of a normal person in a form that didn't require them to know it was AI at all. That's a less satisfying story than “the models got smart enough.” It's also the one that's actually held up every time I've checked it against what really shipped and when.