Hacker News Reader

Comment by z3c0 | Hacker News Reader

I too agree with Chomsky. It's concerning to see people piling onto GPT/LLMs as an extractive tool - something it does okay at, but in the end, something it's not made for. So many companies rushing to use it as such is going to result in some colossal problems for some. The potential for completely false assertions is not something one should be in a hurry to productionize.

There's already a myriad of extractive NLP solutions available that organizations could be utilizing (like any simpler BERT-based model.) Unfortunately, they don't come with the veneer of misleading ethos that GPT/LLMs pretend to have.

Personally, I see some potential for customer-facing uses, but that's because customers like novelties.

View on HN

Do you think AGI will be achieved if this trend continues?

Not really, no, but that's not to say that LLMs couldn't be a useful shell through which an AGI could interact with the rest of the world. But to herald LLMs as the coming of AGI would be to put the cart before the horse.

A rudimentary intelligence capable of basic object-recognition and permanence* would be an actual step towards AGI over LLMs, in my professional (but still fallible) opinion.

Something as basic as an insect, or hell, even protozoa, is all we would need to say that we're on the way. Once we have that, tying it with an LLM to give it denotative capabilities would be the boom we're all hoping for/terrified of.

[*] this is key, but we don't need to codify this so much as find a way for the object-recognition to build up to it. Even children take a while to develop object permanence (hence peek-a-boo being fun for them - you really are disappearing.)

I really don’t understand why it isn’t plausible that LLMs are weak AGI. If we’d been wise enough to create benchmarks for them a decade ago, I’m sure they would have exceed our expectations.

The generality of the intelligence is unquestionable, at least in a language domain. Discounting it as AGI because a fly is better at being a fly seems besides the point.

The generality of its intelligence within text is still spurious at best. Context-switching is practically nonexistent, and whenever it appears to do so, it's only because your prompt pushed it towards statistically appearing to switch context. And frankly, computers have been exceeding our benchmarks for some time - Turing mistakenly thought that appearing human was a sufficient marker of intelligence, but chatbots have been passing the Turing Test for about a decade now. Saying it would surpass our benchmarks only illustrates how little we understand intelligence, not that LLMs are anything close to intelligence.

Edit: that final point is why I propose an in-to-out approach over an out-to-in approach (which is what LLMs are). Let the intelligence make itself, instead of us trying to make intelligence.

“ not that LLMs are anything close to intelligence.”

Well, that’s ridiculous out of hand. What makes you say something like that? If they feel intelligent, pass intelligence benchmarks and are designed to be intelligent— why not ascribe to them intelligence? Is it because they aren’t wet?

I'll turn that right back around and ask if you considered any of the chatbots of the last decade to be "intelligent". They "felt" intelligent enough to pass Turing Tests. Nonetheless, nobody was saying it was ridiculous to conclude that they were only superficial representations of intelligent behavior.

I don't call any of the ML models that I create "intelligent" - I call them probabilistic. Feigning intelligence with probabilistic methods is nothing new (eg ELIZA). If you can scientifically demonstrate that feigning intelligence is equivocal to de facto intelligence, with something more than "feelings", I'll be happy to read your write-up.

I don’t understand. Who is “feigning?” Why is that the basis? I’m saying, let’s judge intelligence based on benchmarks.

No problem with probabilistic methods. Human intelligence is also probabilistic (ie, predictive coding). Maybe the biggest question: do you think these will be exponentially more powerful in 5 years, learning from failures, self-improving and generally economically “a very big deal” due to human replacement? Because that’s what AGI is supposed to do, rapidly—and that’s why it is worth acknowledging the possibility.

From my perspective, if it barks like a dog, walks like a dog… fulfills all the “dog” benchmarks… even if it came about through unusual means, you’d better consider the possibility.

All these experts Pooh-poohing the possibility of AGI because “I’ve worked with LLMs” is really unconvincing. The debate isn’t balanced because most professionals wouldn’t want to claim AGI for professional reasons.

> most professionals wouldn’t want to claim AGI for professional reasons

Most professionals in my field would probably consider this a career-defining moment that would land one in the halls of the greats. I don't really follow this thinking.

And I get the "looks like a duck, quacks like a duck" argument, but as I've repeatedly stated, that's not good enough.

Frankly, I don't buy the "predictive coding" theory - at least, not to the extent that I believe the brain is only probabilistically modeling the phenomena it experiences. This ignores the glaring fact that our senses in-of-themselves exhibit some astounding capabilities right out-of-the-box, and predictive coding has no ground to stand on without the cross-referential analysis of the senses. Human babies are developing their intelligence long before any heavy stimuli. Hearing is the most active before birth, eyesight - our most rich sense - doesn't kick in until after birth. Nonetheless, babies come out of the womb with intelligent reflexes, such as rooting. That bundle of rudimentary intelligence is the basis upon which predictive coding begins - a functional default. An LLM is a set of statically-coded weights, incapable of context switching or object permanence. A child may take a while to understand the difference between "you" and "me", but they still can intuit "thing A" and "thing B". LLMs cannot - at least not without being paired with another model of some sort. And that model will be what tips the scale, not scaling up the features of what's merely a beautiful implementation of the logit function.

Ok, so you don’t buy predictive coding for humans — that helps explain your position.

You also see that some basic production rules (for basic reflexes) could mean the difference between AGI chatGPT and chatGPT today. Because of how it would help chatGPT orient towards the world, I suppose?

I buy that critique because I do view a cybernetic loop as essential to intelligence — there must be some goalstate (however vague) to be achieved successfully.

But, have you seen many claim that chatGPT is AGI? My issue is not with your argument (it’s good!) but the lack of acknowledgment for the reasonability of the position that chatGPT can be called AGI. Most people DO buy predictive coding. And many would view the missing “reflexes” in chatGPT just a matter of good UI/UX (for instance). Yet, I haven’t seen anyone stand up and systematically argue that chatGPT is AGI. (But I’m not on Twitter, so…)

I've considered that senses/reflexes may be supplemented with well-defined components feeding various signals together, but this is more of a challenge than it may seem at first. Wiring components of varying signal types together in a way that they can meaningfully be fed into the same vectorized input stream is no small feat (ie sound, images). We really underestimate the wonder of our nervous system. However, if that can be found out, I'm sure AGI would be right around the corner. Sometimes, I wonder if vectorized formats are even the answer at all. "Life is in analogue" after all.

As for the assertions being made - at first, no, but once the Bing Search fiasco kicked off, I've seen more and more (here on HN) making the claim that ChatGPT and its kin are the beginning. Ultimately, it doesn't irk me - once we hit the same plateau as usual, the hype will boil down - but I do feel a bit concern at what may come of poorly-wielded LLMs before the industry wises up to their risks. A lot of people take ideas they see on HN back to their jobs, so we have a bit of responsibility to manage such hype here.

> manage such hype

With hype I think of blockchain (wherein I never understood why removing people from decision-making through automated contracts was worth so much) or the metaverse (where I never understood how they envisioned a world with tech strapped to their face).

The public communication about AGI isn’t exactly a hype cycle. People should, instead, be expecting a fairly dramatic change in many areas of life over the next 3-5 years.

Because they don't do anything but generate text when prompted, while humans are walking around doing things. I would expect at least two things from an AGI, weak or strong:

1. It has agency. 2. It can move about and manipulate the world in a general as opposed to predefined sense.

I'm not so sure about this. Text is the universal interface to almost anything imaginable. You can easily embed ChatGPT into a robot, describe the robot text-based API to it and give it a task you like - and it will walk around doing things. Will it do a good job, that's a whole another question - in it's current state it may easily hallucinate something disastrous, but that's because the text-generating "intelligence" is weak (in some sense, we cannot describe it precisely, we don't know what it lacks - abstraction, reflection capabilities or something else)

Reply on news.ycombinator.com