Comment by azan_
2 days ago
It's amazing that there are people who still believe in "it's just autocomplete". It hasn't been true for a long time, but currently it's position that reveals complete lack of awareness how good AI has become. It has solved Erdos problem using novel approach. Constant moving of goalpost is recurring theme when discussing AI, but it's really impossible to move it so far that you can frame this achievement as "impressive autocomplete".
I would draw a distinction here. If its a tool (as the GP proposed), it is just fancy auto complete. If it isn't a tool and is instead solving problems in novel ways, inventing new things, etc then it is intelligence and not a tool.
It can't be both in my opinion. To be a tool it needs to be controllable and predictable, intelligences are neither. See humans, and really all living things, for plenty of examples where they can't be completely controlled or predicted.
I mean, tokens are passed as input to a model, which then outputs the next most-likely token. At the heart of it, that's the technology right? Why is it so silly to call that autocomplete? Because it's capable of impressive things?
I don't actually think its silly to call it auto complete.
Personally I could see it being either one. The LLM companies have drastically underfunded projects for things like interoperability. As long as inference is a black box we don't know whether its text prediction as a fancy tool or if something crazier has emerged that could be considered intelligent, self aware, conscious, etc. The former is easily assumed by the architecture, the latter seem far fetched but we simply can't know.
> Because it's capable of impressive things?
Precisely. Calling it autocomplete when it's capable of completing tasks that have nothing to do with autocomplete is silly. If you want to be consistent with your terminology, you'd have to call any stochastic process "autocomplete". What makes it double silly is that you can't really exclude that human intelligence is a stochastic process.
You're making an assumption that there is no difference between intelligence and auto complete with sufficient resources and learned patterns to complete tasks a human might do.
There may not be a difference there, I don't know but I wouldn't assume that intelligence is nothing more than sufficiently complex auto complete.
2 replies →
> If you want to be consistent with your terminology, you'd have to call any stochastic process "autocomplete".
I'm fine with calling any process "autocomplete" if it takes language as input and returns predicted language as output.
I don't feel any need to broaden the definition to include things that have nothing to do with language.