News for hackers. Stuff that matters.
Re:Are Emergent Abilities of Large Language Models a Mirage? by p-e-w on Monday May 01, 2023 @01:00AM (#35769251)
> My standing argument is that they wouldn't hallucinate incorrect responses if they understood anything

Alexander Grothendieck, one of the greatest mathematicians of the 20th century, answered "57" when asked in a public lecture to provide an example of a prime number. 57 = 3 * 19 is not a prime number.

According to your argument, this would imply Grothendieck did not understand what a prime number is. Which is laughable.

Re:Are Emergent Abilities of Large Language Models a Mirage? by deafpolygon on Monday May 01, 2023 @01:09AM (#35769304)
Humans are not infallible. It's a mistake to anthromorphize LLM.
Re:Are Emergent Abilities of Large Language Models a Mirage? by SanderNL on Monday May 01, 2023 @02:05AM (#35769532)
LLM wrong => unredeemable and fatally flawed

human wrong => no problem, mistakes happen

Re:Are Emergent Abilities of Large Language Models a Mirage? by lewhoo on Monday May 01, 2023 @04:58AM (#35770498)
Yeah, a little bit. Our biology makes us unpredictable to a certain degree. In simple terms we're allowed to have a bad or a good day. LLM's remain unchangeable and simple mistakes they make can't be explained in the same manner I think.
Re:Are Emergent Abilities of Large Language Models a Mirage? by SanderNL on Monday May 01, 2023 @06:52AM (#35771120)
I get the reflex, but this becomes increasingly untenable as their complexity balloons. (Our "biology" is some sort of code for "it's complex".)
Re:Are Emergent Abilities of Large Language Models a Mirage? by gmerc on Wednesday May 03, 2023 @11:08PM (#35810778)
But LLMs are not unchangable, there are several mechanisms available by now to drive change and learning (especially as context windows improve)

Practically for outcomes it doesn’t matter if that learning is happening in the LLM or on a layer above it.

Re:Are Emergent Abilities of Large Language Models a Mirage? by ChatGTP on Monday May 01, 2023 @08:56AM (#35772144)
I guess now computer systems have personalities, biases, make things up is kind of different? I guess we've come to expect more.
Re:Are Emergent Abilities of Large Language Models a Mirage? by mnl on Monday May 01, 2023 @02:17AM (#35769593)
Let's consider that anecdote. It's obvious that Grothendieck blurted out 57 without testing any kind of divisibility. Not even by 3. If he had applied his primary school mathematics he would have noticed that neither 57 nor 87 are primes for that reason.

But 17, 37, 47, 67 and 97 are primes, so he probably applied a pattern subconsciously because you would never say 27, a drill cube, nor 77, which is divisible by 11 on sight.

I've taught mathematics at assorted levels, I think I can tell the reasoning behind quite a few human mistakes there. I analyse them, they're not usually just whatever, there's a wrong chain of thought behind the interesting ones. That's how I design exercises, tests and generally teach stuff to humans.

So I've been doing this mostly within Skype and the thing doesn't err because it's reasoning wrongly, it errs because it just comes up with stuff, there's nobody home. And that would be my professional assessment of it as a student. And yes, quite often it gives you good answers for the initial base line questions, but if you refine the problem to probe its level of "understanding" it's not difficult to derail it, it elaborates on its previous procedures while it's being steered towards a wall of nonsense.

Re:Are Emergent Abilities of Large Language Models a Mirage? by vixen99 on Monday May 01, 2023 @04:56AM (#35770485)
Good point. Within organisms there are a number of survival mechanisms even in juveniles. Braitenberg hypothesized simple mechanical vehicles that evolved survival tactics. Not understanding the latter will ultimately be a fatal flaw irrespective of any insightful understanding of other problems. If Grothendieck had known he faced execution for failing to give a correct response, he would certainly have survived. I'd be interested to see AI configurations linked to a number of direct externalities (not linked to human derived/directed information) that might then determine their fate.

https://en.wikipedia.org/wiki/Braitenberg_vehicle

Samuel Johnson: “Depend upon it, sir, when a man knows he is to be hanged in a fortnight, it concentrates his mind wonderfully.”

Re:Are Emergent Abilities of Large Language Models a Mirage? by badrabbit on Monday May 01, 2023 @05:54AM (#35770793)
My argument is not that it shouldn't make a mistake but that it cannot recognize it's mistake and correct it because it would have. That mathematician I am sure recognize that 57 isn't a prime number if you ask him again.

Also mathematics is not the right field to compare this to because there are rules. In languages and object recognition/synthesis it is all subjective. Understanding here means understanding context and human-subjective interpretation.

When I say "that sunset is beautiful" you understand what I mean, ML models simply approximate based on what they see other humans do or say. I am not calling the sunset beautiful because other people are, it is my own subjective interpretation.