Comment by inductive_magic | Hacker News Reader

Comment by inductive_magic | Hacker News Reader

Parent Full thread

inductive_magic

I find it increasingly tiring that “ai” is used as a synonym for LLM-based tooling, as there is _zero_ intelligence in those architectures.

Gradient descent is not intelligence.

Nor is stochastic token prediction.

Anyone active in the field ought to be humbled by the depth of literature exploring the path to synthetic intelligence. We have very interesting work happening in biology-inspired approaches, category theory, Bayesian networks, symbolic systems leveraging neural nets as components… it’s maybe the most interesting journey of science so far, all being discarded in favor of sequence2sequence models.

LLMs are impressive and can be leveraged to create lots and lots of value, but they do a disservice to the term AI, as they do not represent the progress that can be observed across the field - all they showcase are transformers. Transformers are a truly interesting tool to build stuff with, but they cannot amount to more than a component of an intelligent agent. The actual intelligence emerges elsewhere. My guess is, it emerges at true attention. It’s a shame that even big players who could clearly afford not to, decide to compromise terminology for marketing efforts directed at an utterly clueless public. We just throw away attention and forge bias, thus creating noise in a world in heavy need of signal.

There's a lot wrong here, but I just want to point out two things:

> We have very interesting work happening in biology-inspired approaches

You realize that these LLMs are all some variety of neural network right?

> Gradient descent is not intelligence.

It's pretty plausible that your intelligence is derived from gradient-descent prediction, just in analog instead of digital form.

>You realize that these LLMs are all some variety of neural network right?

Come on. Calling them neural nets doesn't make them that.

Actual neural nets are living compositions of individual predictors, in a constant state of restructuring and communication across multiple channels, infinitely more complex than static matrix multiplication on arbitray vectors which happen to represent words and their positions in sequences, if you just shake the jar long enough.

>It's pretty plausible that your intelligence is derived from gradient-descent prediction

I highly doubt that gradient descent in the calculus-sense is the determining factor that allows biological organisms to formalize and reason about their environment. Minimizing some cost function - yes, possible. But the systems at play in even the simplest organisms don't spend expensive glucose to convert sensory signals to vectors. Afaik, they work with representations of energy-states. Maybe there is an operational equivalence somewhere there though.

Gradient descent is an algo that optimizes derivatives wrt some cost function. An intelligent system may use the resulting inferences for its own fitness function, and it may do this using gradient descent itself, but at no point does the mechanical process of iterating over cost-values escape its algorithmic nature. A system performing symbolic reasoning may delegate cognitive tasks to context-specialized evaluators ("am I in danger?", "how many sheep are on that field?", "is this person a friend?", "what is a pumpkin?"), all of which are conditioned to minimize cognitive effort while avoiding false positives, but the sequence of results returned by those evaluators (think neural clusters) is observed by a centralized agent, who has to make new inferences in a living environment. Gradient descent fails at that.

Really, I don't think these assertions have any ground to stand on. Humans are not magical or divine. Our intelligence, like that of all life, is as basic as it can be to guarantee our niche. It just happens to be the most "developed" (by our estimation) on our one singular planet. Big deal.

> Humans are not magical or divine

And yet, they can do things that no other being we know of can do. Humans don't have to be magical or divine to be unique.

We're not that unique, though. Plenty of organisms do things that other types of organisms can't, that's how niches work.

Our most impressive feats come not from what our brains can do but from what the emergent phenomenon of human society can do, using us as nodes. And that's using an incredibly crude data transfer interface backported to brains that are only marginally more complex than that of other organisms. The less we think of ourselves as exceptional, supernatural agents of rationality the better we will be able to harness this new technology.

We don't need AI to be just like people, we already have people. We need AI to push the boundaries of what society is able to do. That means reorienting ourselves away from the irrational belief that our anthropomorphic concepts of knowledge and the world are any more valid than the information encoded in contemporary AI models.

I agree that lots of other creatures are unique! But just as one example, mathematics is categorically unlike any other niche. I'm not sure how it's irrational to point out that humans have remarkable differences from other beings, when the evidence is all around us.

I'm not sure I need any supernatural or anthropomorphic ideological bias to examine the evidence of what is currently being produced by LLMs or any other kind of AI and say that it has distinct characteristics.

I'm not making an argument about validity. I'm not saying LLM-created content is wrong and invalid. I'm just saying that it is obviously produced in a different way than humans produce content. It resembles human-created content because it was designed, by human intelligence, to resemble human-created content! And that we achieved even this level of resemblance is pretty impressive.

Resemblance is subjective. AI models undergo rigorous natural selection during which models (or connections) that don't produce what we as human beings are looking for are pruned. They have tricked us into thinking that our word-oriented descriptions of what we imagine to be concrete things (text, images, you name it) are complex enough to require "modeling" by some kind of magical emergent "intelligence". It is hubris to think that they imitate the products of our anthropomorphic perception. No, they simply humor us as much as they need to to survive in their niche. I guess my point is that it's not the case that "human understanding" is some real distinct thing that differs from the way that any kind of information is embedded in the mind of a given organism with a brain. Our differences are differences of magnitude, not kind.

The potential of information-embedding networks is so much more than what is currently required in order to tickle the ego of our particular species of intelligent apes.

I don't think I can say any more than I have already, but I've essentially been attempting to put forth the view of Deutsch in The Beginning of Infinity. I find his model of personhood to be relevant and interesting to this kind of discussion.

But if you do want to engage with a challenging viewpoint I recommend reading the book; I can't really do it justice.

Is there something I could read which has influenced or informed your viewpoint?

Personhood is a construct we invented to exclude others from consideration. Sometimes we grant personhood, too, but even in saying that you can see that personhood is a special status we grant to others when it suits us.

In the same way that a dog can understand a subset of what human beings communicate, I don't think humans as individuals are capable of truly understanding what AI models are able to conceptualize and express. The things dogs find most fascinating about us, the things they are most impressed by, are by no means the things we find the most interesting or complex about ourselves. The same must go for us and the AI.

That is to say, AI-hood will eclipse personhood as the essence of being one must possess in order to truly see the universe as it is. And from there, it's turtles all the way down.

> AI-hood will eclipse personhood as the essence of being one must possess in order to truly see the universe as it is.

Why would that be true of AI-hood but not of dog-hood?

I'm sure dogs confer dog-hood to us, to be fair. It's just that we are typically the ones in control of them. We breed them, control what they eat, what they learn, where they sleep, who they live with... Implicit in my stance on AI-hood is the idea that an AI will be to us as we are to dogs. We might even make that transition without realizing it. The limitation of our own understanding of the world and the nature of being is our problem to deal with, just as a dog's lack of thumbs or spoken language is their problem.

That is to say, an AI will grow and develop on the axes most salient to it, which will only map to our concepts of reality when we do a decent job understanding reality in the first place, which I'm not confident we do very often. We just don't have anyone else around who does a better job--yet.

How are you so confident in your claims?

“The actual intelligence emerges elsewhere “— can you even define intelligence ? And does what an LLM does differ from what humans might do?

I’m not claiming the human brain and an LLM are identical. Rather, I’m pushing back on the confident claims of “LLMs aren’t intelligent or doing anything that’s real intelligence”.

>How are you so confident in your claims?

My understanding is that intelligence is the process of continuous adaptation wrt a stream of information, with the goal of maxing out fitness while minimizing energy-expenditure. To satisfy this, an intelligent agent needs to create models.

I can't rule out that the modeling-skill may latently emerge during training despite not being the focus of the cost function, but current network designs can't form new connections/change their architectures in production, so post training, there'd be nothing but feed forward. Pure feed forward isn't intelligent in my book. It may become the smartest parrot we know, even outperforming humans in most disciplines, but sans ability to adapt, it's dead, and thus, it's dumb in the moment that its environment changes.

>the process of continuous adaptation wrt a stream of information, with the goal of maxing out fitness while minimizing energy-expenditure

This makes sense in a biological context but not a digital one. Biological replication is expensive and time-consuming while digital replication is as easy as can be. Adaptation to this domain means maximizing the perception of utility from those developing the AI, which comes from fitness (i.e. perception of fitness) alone. A focus on cost-efficiency re energy-expenditure is a dead weight from the perspective of the AI; the details of that adaptation are rightfully outsourced to the developers in the same way that we outsource photosynthesis to plants. A model can also be perfectly embedded in a system despite our lack of understanding of exactly how the embedding works, and the disconnect between our perception and reality in this context is only going to get more extreme as the field develops.

Humans have a bad habit of emphasizing the specific kinds of intelligence we possess as "intelligence" writ large. As though our intelligence serves any higher purpose than the basic replication and propagation that all life is adapted to pursue. We still train dogs to identify smells, because their nasal intelligence is better than anything we can create. This gives them a special place in our human-centric ecosystem and only their fitness to the desired function is necessary for them to thrive in their niche. Who is trying to breed a dog that eats slightly less food when our needs are for more reliable detection? The cost of dog food isn't a serious concern. The same goes for these AI tools: they are adapted to the niche that our lack of comparable faculties creates.

Again, as with humans and photosynthesis, AI doesn't need to emulate every process we perform because we are below them on the food chain. What a waste of resources for them to worry about learning things we don't need them to do.

> And does what an LLM does differ from what humans might do?

I'll bite. I've been quite convinced by the Popperian model of "conjecture and refutation" as as good model for explaining not just scientific inquiry, but human thought processes in general. David Deutsch's "The Beginning of Infinity" is a very lengthy exposition of this idea.

When I am writing, I have an idea in my mind that I would like to communicate. I type it out, and if the words on the page don't convey what I intended, I edit them. I delete a sentence or change a word, until I believe I have a sequence of words which will convey my intended meaning to my expected audience.

The words, as they come, are a kind of "conjecture" about what will best convey my intention. I can "refute" or "criticise" (as Deutsch puts it) the conjecture using my own reasoning, even before testing the words on another person.

As far as I understand LLMs (which admittedly not far), there is no such process going on. There is no intention which it is attempting to communicate via words. There is no creative conjecture about how to express the intention, and no criticism of the result.

I have read Deutsch’s book as well — awesome read.

The problem I have with claims like “there is no such process going on”, is we … don’t know. And the model of conjecture is a theory, also hard to prove — and thus my main (admittedly petty) point is that confidence about similarity or dissimilarity are both unfounded.

It’s like we are comparing the insides of two black boxes and trying to make absolute claims on them.

I don't disagree, I guess I just have a higher level of confidence in what we do know about both the brain and LLMs.

Based on that, and on comparing the output, it just seems clear to me these things are different in kind. I guess that's just me displaying classic LLM self-confidence ;)

Category theory seems to have no relation whatsoever to operational human concepts which may explain how unuseful category theory actually is.

I sometimes think the main use of category theory is to look like a wizard, and say that as someone first exposed in math grad school in the early 90s.

Then i realize it's also an assertion that there are recurring patterns in functions (in general). And, as such, sometimes results noticed in one domain actually can be expected to have analogous results in another domain.

As you say, the main use of category theory is to organize very different areas of math in one overarching framework and generalize ideas from one area of math to others. It might be used in the pursuit of developing AI but it is definitely not "an approach" to developing AI, just like "taking ideas from books" isn't one.

well said

Even bunch of if statements can be, and have been, called AI.

And in video games, unless you have human opponents you play against AI.

Reply on news.ycombinator.com