Why GPT-4 is More Than Just Fancy Auto-Complete

8 May 2023 · 1 min

Watch on YouTube

0:50

Dharmesh Shah argues GPT-4 is a reasoning engine, not just autocomplete. Given facts, it logically generates answers rather than predicting the next word. Sam Altman calls it a reasoning engine, and the gap between statistical prediction and what it actually does is bigger than people think.

Chapters

Is this just really fancy autocomplete or is there something more to it? On some spectrum, almost everything that you've ever experienced is fancy autocomplete, right? Like that's— I think the reason we kind of fall into this trap is it's a gross oversimplification of what's actually happening there, right? So GPT-3 and now 4 is a reasoning engine. And Sam Altman has talked about this. It's not a knowledge base. What they've built is a reasoning engine that says, given this set of facts that it knows about the world, How can it try to logically come up with something that answers the question? So yes, at some root level, it's like autosuggest, but, and I'm not gonna suggest that it has consciousness, it's thinking, but we're kind of headed down that path. It's like it's able to do things that are not explainable by a simple probabilistic model of autosuggesting next character, next word, next token, next sentence, right? Like it's gone well beyond that.