Hacker News Reader

Comment by crop_rotation | Hacker News Reader

This seems to miss the forest for the trees. Whether LLMs will have an impact similar to the industrial revolution will not depend on whether they pass some arbitrary threshold where everyone is convinced that it is an AGI and understands what words mean. It will depend on the utility they provide. And the utility is there right now. GPT4 is so immensely useful for so many things. At even a reasonable pace of improvement, it is hard to see why LLMs would not be able to do more and more things.

Morever and slightly off topic, most humans also don't care what words mean or what numbers mean in a philosophical sense. If you talk of "The Axiom of Choice" in a big company software meeting, people will ask if it is the new ice cream flavour in the cafeteria. That doesn't prevent people from getting value out of both words and numbers.

View on HN

The problem is that most tasks require general intelligence at some point. So the question is more: given that in the vast majority of cases, weak-AI cannot go from "0 to 100%" how much should we care?

The whole "the sky is falling" schtick i think requires weak-AI systems to be getting very near 100%s reliably across industries. Anything even 90% and we're talking more about assisting and empowering users than replacing them.

> This seems to miss the forest for the trees ... most humans also don't care what words mean

You miss the point, because you are reasoning around the different perspective of "hav[ing] an impact similar to the industrial revolution", so you take the semantic "payload" as if it were "just an expression".

The point is, that it produces statements, and in general we want true and founded output from an entity producing statements - "intelligence" is the condition enabling that.

So, when the machine says

  "Napoleon was addicted to cotton candy"

, we "must know" that the machine has verified it, like a "good" statement producer would (as opposed to all the other reasons to produce that statement);

and when the machine says

  "The current war will be won by the white faction"

, we must know that it effectively reasoned on the relevant conditions and factors, on par with the best Intelligence Analysts (as opposed to all the other reasons to produce that statement).

Maybe a weird question, by why does everyone assume that we will continue to have improvements and GPT-5,6,7,8 will surely "get there"?

Isn't it possible that we've almost reached the limit of this technology and it won't improve much for a very long time? Like self driving cars.

Self driving cars have a very different problem. Regardless of the improvements and sufficiency they reach, their errors can be fatal. One news headline about a self driving car crash will ensure that the technology doesn't get any approvals.

ChatGPT for all it's flaws can't jump out of the computer and harm the user. It can harm users in others ways, but those are less known or understood by the general public, and thus it is much easier to take to general availability.

I agree on the limit part that it might be possible the technology has reached a limit, however, it is still better to think about what happens if it hasn't and the rate of improvement continues.

I doubt it. I would expect that if GPT's "knowledge" reaches some limit, OpenAI could just hire subject matter experts to write books for the model to read, and then use those to (effectively) down-rank material that differs substantially.

If we're talking about the modeling itself, I think we're a long long ways from any limits. There's room for improvements such as cheap personalized results, faster hardware with more memory for lower costs, bot-to-bot communication allowing my instance to interface with yours to come up with mutual understanding, etc. Pretty much all of the components for this exist already in some form, they just need to be optimized and implemented on a large scale and made available to us humans.

I found this Veritasium video interesting, about a company that is paring down models so they fit on analog chips, so the chips can be used in self-contained devices. It also appears that quantum computers may be even better for matrix math that transformers rely on. If all these hardware advances pan out...

https://www.youtube.com/watch?v=GVsUOuSjvcg

https://phys.org/news/2023-01-analog-quantum-previously-unso...

It takes months to years to write a well researched and in-depth book and subject matter experts are not necessarily good authors. Beyond that, each book only moves the needle a little bit as well because of the size of the training corpus. If there's going to be improvement from that path, it will be at best slow and incremental.

You're right, maybe a whole book is unrealistic. I guess it's more likely that they'd have someone make small corrections, like if enough people flag a common result as wrong or poor.

Reply on news.ycombinator.com