Re:GPT-4 Can't Reason
by throwanem on Tuesday August 08, 2023 @11:46AM
(#37050789)
There are humans who can't add two four-digit numbers without scratch paper, too. So what?
[ Parent
| on HN ]
Re:GPT-4 Can't Reason
by bruce343434 on Tuesday August 08, 2023 @11:49AM
(#37050868)
So what not? This is still incredibly impressive technology.
Yes, and it's incumbent upon us all not to let first impressions run away with us to the extent of making unsupportable claims, even by implication.
Re:GPT-4 Can't Reason
by bruce343434 on Tuesday August 08, 2023 @06:12PM
(#37056153)
I'm sick and tired of the generally dismissive attitude towards these LLMs. They're no messiah, but they're for sure not useless.
I think you must have meant to reply to someone else, since I haven't written off LLMs as useless here or indeed, to the best of my recollection, anywhere.
That said, I suspect dismissive attitudes toward LLMs would be less common if messianic claims on their behalf were likewise.
Re:GPT-4 Can't Reason
by jovial_cavalier on Tuesday August 08, 2023 @01:05PM
(#37052228)
I think GP's point is that you can reach something approximating reasoning by brute force.
In the case of your example, you don't need to encode an algorithm to do 4-digit addition, just remember that 1000+1000=2000, 1000+1001=2001, etc.
Think about how dramatic a change it was when computers started being able to efficiently add those numbers and got better than them. Now no matter what the human's capacity a huge number of them have a computer in their pocket that can fill in the gaps. And that change was wildly disruptive. Same as when machines started to do tasks that humans couldn't. Any time a relatively accessible technology overtakes humans in a thing that used to be human-only and is important to civilization it's been highly disruptive. If gpt's reasoning is good enough to replace like, the lowest level of phone tech support or the like, that's already jobs that'll be lost.
That "if" is doing a lot of legwork here. Most notably, it's concealing an implication that the LLM may be reasoning in this example, and not merely fitting tokens from its input into a statistical reconstruction of the most basic and thus likely most commonly encountered proof in elementary symbolic logic.
That said, you're not wrong that this seems "good enough" to be wildly disruptive. Part of that disruption, which is already ongoing, is the development of a whole new class of exploits. I expect once companies start firing phone support representatives in favor of LLMs, and similar, this process will accelerate in fascinating ways.
Re:GPT-4 Can't Reason
by throwawaymaths on Tuesday August 08, 2023 @06:22PM
(#37056269)
So what? In your context, you would say "computers can math better than humans". I think that's uncontroversial.
Crazy thing is, your difficulty with this syllogism proves my point about the median human.
Whatever their capacity for reason or lack thereof, I'll give LLMs this: I've never yet seen one resort to ad hominem for the lack of any more substantive counterargument.
Re:GPT-4 Can't Reason
by throwawaymaths on Tuesday August 08, 2023 @10:11PM
(#37057951)
The counterargument is substantive, and you're committing fallacy of fallacies anyways.