What August's AI Debates Tell Us About the Hype Cycle

Look back at this month and you can watch the whole AI hype cycle compress into about four weeks. Early August was full of GPT-3 demos: people wiring the OpenAI API into little tools that wrote code from comments, drafted UI mockups from plain English, answered trivia, imitated writing styles, and generally made the model look like it understood what it was doing. Twitter threads with looping GIFs of text turning into working React components. The tone was breathless. “This changes everything” energy, the kind that makes you want to cancel your weekend plans and go build something.

Then, by the 22nd, the mood had flipped. The piece that crystallized it — the one getting passed around everywhere, sometimes summarized as “GPT-3, Bloviator” — made the opposite case: that the model is a extremely good pattern completer with no grounding in what any of the words actually mean. It doesn’t know anything. It’s predicting plausible next tokens based on a staggering amount of training text, and when it looks smart, that’s because plausible-sounding text is, often, actually correct or reasonable. When it looks dumb, it’s because plausible and correct came apart, and nothing in the system can tell the difference from the inside.

Both things are true at once, and that’s the uncomfortable part. The demos weren’t fake. GPT-3 really can produce useful code snippets and coherent essays and passable poetry. The critique isn’t fake either — ask it a question that requires actual world-model reasoning rather than pattern matching, and it’ll confidently generate nonsense with the same fluency it uses for the correct answer. There’s no internal signal of “I’m not sure” that shows up differently from “I’m sure.” That’s a real limitation, not a rhetorical trick by skeptics.

Why the swing happens so fast

Part of it is just how demos travel. A cherry-picked output spreads in an afternoon; the failure cases take longer to collect and don’t compress into a single tweet as well. So the hype front-runs the correction by default, every time, for every model, not just this one.

Part of it is that “understanding” is doing a lot of work in these arguments, and nobody agrees what it means. If you define understanding behaviorally — does it produce the output a person who understood would produce — GPT-3 clears the bar constantly. If you define it as having some internal representation that tracks truth and revises itself when it’s wrong, it clearly doesn’t, and there’s no architecture here that would let it.

My take, for what it’s worth: the useful move is to stop asking whether GPT-3 “really” understands and start cataloging what tasks it’s reliable for and what tasks it confidently fails at. That’s a more boring question than the philosophical one, but it’s the one that actually tells you whether to ship a product on top of this thing. Expect this exact cycle — dazzling demo, viral skepticism, quiet settling into “it’s a tool with a specific shape of failure” — to repeat with every model that comes after this one. The hype cycle isn’t a bug in how we talk about AI. It’s just how a new capability gets stress-tested in public before anyone writes the boring documentation.