Overheard a guy at a cafe say LLMs are 'just autocomplete on steroids' and it got me thinking

I mean sure, technically that's what they are, but isn't a car just a box with wheels spinning really fast? The takeaway was that people who actually build stuff with these tools see way more nuance than folks who just read headlines. Has anyone else had that moment where a random comment made you rethink your whole approach to a project?

3 comments

3 Comments

henryp401mo ago

The autocomplete take is a handy shorthand but people get way too worked up about it honestly. @david_jones38 makes good points about the models backtracking, but isn't that just a more complex pattern matching algorithm that's better at faking it? Sometimes I wonder if we're all just overthinking this to feel smart about using a glorified text predictor.

the_mary1mo ago

Oh, I read this thing from a researcher who said thinking of it as autocomplete is like calling a telescope just a really long tube. Makes you realize there's a lot more going on under the hood than people give credit for.

david_jones381mo ago

The autocomplete comparison bugs me for a different reason actually. If you watch how these models handle things like math or logic puzzles, they're clearly doing something closer to reasoning step by step, not just guessing the next word based on patterns. I've seen tests where they'll backtrack and correct themselves mid sentence, which is way more than a simple autocomplete system can do. It's like people who compare a car to a faster horse because both get you from point A to B. There's a whole layer of emergent behavior happening that the simple comparisons just can't capture.