Introducing ChatGPT

OpenAI dropped a new model this week called ChatGPT, and the pitch is simple: it talks like a conversation partner, not a search box. You ask something, it answers, you push back, and it actually remembers what you just said instead of starting from zero every time.

That sounds small until you've used the older generation of chatbots, which mostly just autocompleted text and had the memory of a goldfish. ChatGPT can handle a followup question without you re-explaining the whole context. It'll also cop to being wrong when you point out a mistake, rather than doubling down, and it's built to flag premises that are just false instead of playing along.

The dialogue format is the real trick here. OpenAI trained it specifically to hold a back-and-forth, which means it can reject requests it decides are inappropriate mid-conversation, not just at the first prompt. That's a meaningfully different design goal than a model built to finish a paragraph convincingly.

None of this makes the model infallible. It can still be confidently wrong, and admitting mistakes after the fact isn't the same as not making them. But the shift from one-shot text completion to something that behaves like an ongoing exchange is the part worth paying attention to, because it changes how people will actually use these tools day to day.