Last week a video went viral in which a kid asks OpenAI's voice mode to time his run. The AI starts the timer, the kid immediately says he's back and asks for his time: a bit over 10 minutes, says ChatGPT. Fun clip. Good laugh. The sentiment in the comments: see, AI isn't all that good after all.
But that's a bit like writing off Google because it won't make you coffee. The tech behind ChatGPT is, at its core, a word predictor — it wasn't built to run timers. But good luck explaining that.
It's looking more and more like two groups are forming who both talk about AI but aren't talking about the same thing. One group has seen clips like this, watched ChatGPT botch a simple question once, and built their whole picture on that. The other group, which I count myself in, uses paid AI tools like OpenAI's Codex or Claude Cowork every day, and experiences something completely different.
Both groups talk about "AI," but they're using totally different products. OpenAI's free voice mode runs on a model from the GPT-4o era, which came out well over two years ago. Meanwhile the same company also has Codex: a model that can spend an hour autonomously restructuring a codebase or hunting for vulnerabilities in systems.
Group one uses a voice assistant to pull up a recipe. Group two has an agent run security audits on a codebase. Programmers using Codex build with AI. My relatives on free ChatGPT play with AI.
No coincidence
There are two structural reasons for this gap.
The first is simple: money. Professionals are quicker to pay for the efficiency AI gives them. So naturally, AI companies are focusing more and more on this audience. Of ChatGPT's nearly 1 billion weekly users, most stay stuck on the free version, which is pure bleed for OpenAI.
The second reason is technical: code has explicit, verifiable feedback. A piece of code passes or fails. You can make a model measurably better by running through that process over and over. It's different with tasks like writing, planning, or coming up with recipes. Those are objectively much harder to judge, so they get less attention.
In my own circle too, I notice it often stops at: "I use it as a faster Google," or "I run my email through it real quick before I send it." AI slots neatly into the lineup of innovations we already had: first there was the book for knowledge, then the internet with Wikipedia, and now ChatGPT. Faster answers, same idea. Meanwhile I'm putting agents to work every day that go off and do their thing while I grab a coffee. I'm just using a completely different product that happens to also be called AI.
On the right side of the wall
In Silicon Valley, this realization of two groups has created a mood of its own. Among employees at startups and tech companies there's a sense of urgency, of grinding, of 90-hour workweeks and startups where 996 is the norm: nine to nine, six days a week. The belief is that the gap we're seeing now isn't temporary, it's permanent. Anyone who isn't on board now — who doesn't learn how to work with agents, who isn't paying 200 dollars a month for a Claude subscription — is too late. And not a little too late. Too late for good. Work hard now, and you land on the right side of the wall. On the other side waits what they call the permanent underclass.
A wild frame, I get it. But I'm desperately looking for reasons to soften it. And the gap that's opening up now is only getting bigger. AI companies are pivoting from generating pictures for consumers to professional tools. Earlier this month OpenAI already pulled the plug on their video generation model Sora, because they need all their compute for other products.
On top of that there's an economic gap growing right along with it. The group that can drop 200 dollars a month on an AI subscription right now is building a head start. They learn to work with agents, which makes them more productive, and they spin up companies with fewer and fewer people in between. This group is literally buying extra thinking capacity.
Sure, everything democratizes eventually. But until then we're stuck with two AIs, two groups, and one brand name. Still seems like a raw deal to write off AI because it timed your run wrong, then walk into the office one morning and find that same AI has taken your job.