I spend a good chunk of my week judging what AI models produce. I read their reasoning, watch them solve coding tasks, and score how close they got. So when people ask me whether AI is about to leave human intelligence in the dust, I have a less dramatic answer than the headlines: it's astonishing, and it isn't close. Both things are true.
The impressive part
There's no use pretending the progress isn't real. A model can read a 40-page spec, draft a working module, write tests for it, and explain its choices faster than I could open the file. On narrow problems with clear feedback, these systems already operate well past most people, myself included. That's not hype. I see it on a normal Tuesday.
If you define intelligence as "produces useful output across a huge range of tasks," AI has been quietly crossing thresholds we used to treat as science fiction.
The gap that doesn't close on schedule
But general human intelligence is a different animal, and the gap shows up in the failures, not the wins.
The same model that nails a hard task will confidently invent a function that doesn't exist. It'll ace a benchmark, then fall apart when I change one assumption the benchmark didn't mention. It has no reliable sense of when it's wrong. A junior dev who breaks the build feels the consequence and adjusts. The model just generates the next plausible token and waits for me to tell it.
That's the core of it. What these systems do is extraordinary pattern completion over an enormous amount of text. What humans do is build a messy, grounded model of the world, carry goals across years, and know the difference between "I'm sure" and "I'm guessing." We transfer a lesson from one domain to a totally unrelated one without being retrained. We notice when a problem is the wrong problem. Current AI doesn't do those things, and scaling the same recipe hasn't made them appear.
"Soon" is doing a lot of work
I'm not saying never. I'm saying not soon, and I'm wary of anyone who sounds certain in either direction. The honest position is that we don't yet know whether bigger versions of today's approach reach general intelligence or hit a wall. My bet, from the seat I sit in, is that something fundamental is still missing rather than just a few more parameters.
What this means if you build
Treat AI as a phenomenal collaborator with no judgment. Lean on it for speed and breadth. Keep a human on anything where being confidently wrong is expensive. The teams that win this decade won't be the ones waiting for the model to think for them. They'll be the ones who got good at working alongside something brilliant and strange.
What's your read? I'm curious whether people closer to the research feel the same wall, or see a way through it.
Top comments (0)