Live data from Hacker News

You are the scariest monster in the woods

jamie.ideasasylum.com

351–360 of 405 posts

Re: You are the scariest monster in the woods

#351

Earlier quoted context omitted.

No, they don't - you can update context, make it a sliding window, create a sort of register and train it on maintaining stateful variables, or various other hacks, but outside of actively managing the context, there is no state. You can't just leave training mode on, which is the only way LLMs can currently have persisted state in the context of what's being discussed. The context is the percept, the model is engram…

I agree what people probably actually want is continual training, I disagree continual training is the only way to get persistent state. The GP is (explicitly) talking about long term memory alone and in the examples. If you have an e.g. 10 trillion token context then you have long term memory, which can give the ability and enable long term goals and affect actions over tasks as listed, even without continual traini…

A giant, even infinite, context cannot overcome the fundamental limitations a model has - the limitations in processing come from the "shape" of the weights in latent space, not from the contextual navigation through latent space through inference using the context.

The easiest way to understand the problem is like this: If a model has a mode collapse, like only displaying watch and clock faces with the hands displaying 10:10, you can sometimes use prompt engineering to get an occasional output that shows some other specified time, but 99% of the time, it's going to be accompanied by weird artifacts, distortions, and abject failures to align with whatever the appropriate output might be.

All of a model's knowledge is encoded in the weights. All of the weights are interconnected, with links between concepts and hierarchies and sequences and processes embedded within - there are concepts related to clocks and watches that are accurate, yet when a prompt causes the navigation through the distorted, "mode collapsed" region of latent space, it fundamentally distorts and corrupts the following output. In an RL context, you quickly get a doom cycle, with the output getting worse, faster and faster.

Let's say you use CFG or a painstakingly handcrafted LORA and you precisely modify the weights that deal with a known mode collapse - your model now can display all times, 10:10 , 3:15, 5:00, etc - the secondary networks that depended on the corrupted / collapsed values now "corrected" by your modification are now skewed, with chaotic and complex downstream consequences.

You absolutely, 100% need realtime learning to update the engrams in harmony with the percepts, at the scale of the entire model - the more sparse and hierarchical and symbol-like the internal representation, the less difficulty it will be to maintain updates, but with these massive multibillion parameter models, even simple updates are going to be spread between tens or hundreds of millions of parameters across dozens of layers.

Long contexts are great and you can make up for some of the shortcomings caused by the lack of realtime, online learning, but static engrams have consequences beyond simply managing something like an episodic memory. Fundamental knowledge representation has to be dynamic, contextual, allow for counterfactuals, and meet these requirements without being brittle or subject to mode collapse.

There is only one way to get that sort of persisted memory, and that's through continuous learning. There's a lot of progress in that realm over the last 2 years, but nobody has it cracked yet.

That might be the underlying function of consciousness, by the way - a meta-model that processes all the things that the model is "experiencing" and that it "knows" through each step, that comes about through a need for stabilizing the continuous learning function. Changes at that level propagate out through the entirety of the network, Subjective experience might be an epiphenomenological consequence of that meta-model.

It might not be necessary, which would be nice if we could verify - purely functional, non-subjective AI vs suffering AI would be a good thing to get right.

At any rate, static model weights create problems that cannot be solved with long, or even infinite, contexts, even with recursion in the context stream, complex registers, or any manipulation of that level of inputs. The actual weights have to be dynamic and adaptive in an intelligent way.

Re: You are the scariest monster in the woods

#352

I struggle to understand how people can just dismiss the possibility of artificial intelligence. Human cognition was basically bruteforced by evolution-- why would it be impossible to achieve the exact same result in silicon, especially after we already demonstrated some parts of those results (e.g. use of language) that critically set us apart from other animals? I'm not buying the whole "AI has no agency" line eith…

The worst thing Star Trek did was convince a generation of kids anything is possible. Just because you imagine a thing doesn’t make it real or even capable of being real. I can say “leprechaun” and most people will get the same set of images in their head. They aren’t real. They aren’t going to be real. You imagined them.

But somehow you don't hate on Steamboat Willie? I grew up believing a mouse could operate a ship, and a coyote could survive impacts from anvils.

Re: You are the scariest monster in the woods

#353

I struggle to understand how people can just dismiss the possibility of artificial intelligence. Human cognition was basically bruteforced by evolution-- why would it be impossible to achieve the exact same result in silicon, especially after we already demonstrated some parts of those results (e.g. use of language) that critically set us apart from other animals? I'm not buying the whole "AI has no agency" line eith…

> I'm not buying the whole "AI has no agency" line either; this might be true for now, but this is already being circumvented with current LLMs (by giving them web access etc). They don't have agency, because they don't have persistent state. They're like a function that you can query and get an answer. During that answer the LLM has state, but once it's done the state is gone. Humans (and other "agents") have persis…

There are two separate issues that folks are getting tripped up on here. The first is that the most powerful AI systems do not do offline learning. There are a bunch of hard problems here: e.g. known unsupervised learning techniques have been far less successful, and inference only approaches cost effectiveness by decoupling from training. It seems plausible that we will solve some of these, though I don't know about others.

The way I have been thinking about the other bit is is that LLMs are functionally pretty similar to the linguistic parts of a brain attached to a brain stem (the harness is the brain stem). They don't have long-term memory, the capacity for inspiration, theory of mind, prioritization, etc because they just don't have analogues of the parts of the brain that do those things. We have a good sense of how to make some of those (e.g. vision), but not all.

The common ground here is that some fundamental research needs to happen. We need to solve all of these problems for AI to become independently dangerous. On the other hand, it's proving mildly dangerous in human hands right now - this is the immediate threat.

Re: You are the scariest monster in the woods

#354
post #272

Earlier quoted context omitted.

No, it is not accurate at all. There are some people that do all of these sure, however the vast majority of people live pretty ordinary lives where they do very little of what is described. I actually think it is very intellectually lazy to be this cynical.

It's absolutely true, but most of the destruction is abstracted away so modern humans don't have to experience it directly or even really think about it. A staggering number of creatures are killed by normal people driving to normal things every year. Many more are killed by all of the resource extraction needed to supply our normal lives. Not to mention the vast numbers of wildlife that sees their habitat destroyed…

Sorry but we weren't talking about Animal Welfare arguments. It was clearly in the scope of how people treat each other. Philosophical discussions similar to Janist/Vegan style argument are like well outside of the scope that being discussed.

As for hunting. I don't see anything wrong with hunting. I don't see anything wrong with eating meat.

As someone that has lived the vast majority of their life in the countryside, I also have little time for animal welfare arguments of the sort you are making.

> But given our history I'm also pretty terrified of what happens if and when we run into a smarter and more powerful alien species or AI. If history is any guide, it won't go well for us.

This is all sci-fiction nonsense. If we had any sort of aliens contact they wouldn't be many of them, or it would most likely be a probe like we send out probes to other planets. As for the super intelligence, the AI has an off switch.

Re: You are the scariest monster in the woods

#355

Earlier quoted context omitted.

We built something we don’t understand by trial and error. Evolution took a few billion years getting to intelligence, so I guess we’re a few sprints away at least

Evolution took millions of years to reach the minimal thing that can read, write and solve leetcode problems, so by similar logic we're at least a few million years in...

Except it learned by direct examples and none of these models can solve even basic logic that isn’t in training data somehow.

Us having the ability doesn’t mean we’re close to reproducing ourselves, either

Re: You are the scariest monster in the woods

#356

Earlier quoted context omitted.

But that's the point: there isn't anything substantially different within the scope of computation. If you are given a set of LEGOs and all you can do is snap the pieces together, then there's nothing other than snapping pieces together that you can do. Adding more of the same LEGOs bricks to the set doesn't change the game. It only changes how large the structures you build can be, but scale isn't some kind of magic…

I think it is well established that scale can transcend limits. Look at insect colonies, animals, or any complex system and you will find it is made out of much simpler components.

Scale qua scale does not transcend the limits of the activity occurring in the system. Meaning, each ant in an ant colony has a set of fixed powers to do X, Y, and Z. Adding more ants doesn't change what an individual ant is capable of per se. Now, you might say that while an individual ant exercising its power X cannot lift a tree branch, a colony of a thousand, with each ant exercising its power X, might. But this doesn't transcend the limitation of the individual ant - each ant is still only exercising some power X - nor is the aggregate manifesting some new kind if power. It is only multiplying the exercise of X.

In the case of computers, a program is a set of formal rules that take some sequence of uninterpreted symbols and maps them to some other sequence of uninterpreted symbols. Adding more rules and more symbols scales the program - and it is the only scaling you can perform - but you don't somehow magically create interpreted symbols. How could it? This is magical thinking.

Re: You are the scariest monster in the woods

#357

Earlier quoted context omitted.

I don't know what else I can write without repeating myself. I already wrote that even under the assumption that intelligence is a purely biological phenomenon, it does not follow that computation can produce intelligence. This isn't a matter of probabilities. We know what computation is, because we defined it as such and such. We know at least some essential features of intelligence (chiefly, intentionality). It is…

What. Are you saying that "intentionality", whatever you mean by it, can't be implemented by a computational process? Never-ever? Never-ever-ever?

If you don't know what intentionality is, a commonplace term in the philosophy of mind, then I suggest you take some time to better acquaint yourself with the subject matter. This is probably the major blow against any kind of AI fantasy, and you don't do yourself any favors by wading in the shallows and confidently holding to misinformed opinions.

So, one last time. A computer program is a set of formal rules that takes one sequence of uninterpreted symbols and produces another sequence of uninterpreted symbols. Adding more rules and more uninterpreted symbols doesn't magically cause those symbols to be interpreted, and it cannot, by definition.

Re: You are the scariest monster in the woods

#358

I struggle to understand how people can just dismiss the possibility of artificial intelligence. Human cognition was basically bruteforced by evolution-- why would it be impossible to achieve the exact same result in silicon, especially after we already demonstrated some parts of those results (e.g. use of language) that critically set us apart from other animals? I'm not buying the whole "AI has no agency" line eith…

When evolution designed the brain it didn't optimize for it being realistic about things. It optimized for reproductive success, better achieved by thinking you're the greatest and no other folk, let alone some mere machine could do what you do.

Re: You are the scariest monster in the woods

#359
>.. if you equip humans with a hammer, or sword, or rifle, or AI then you’ve just made the scariest monster in the woods (that’s you) even more terrifying.

I have some optimism that humans with AI may be less scary than without. One of the biggest problem is once humans get in power, having said they'll be nice they go a bit nuts like Putin or quite a few others I'm sure you can think of. If instead we had things run by open source AI everyone could see what it was thinking and it wouldn't be so able to act nice and then flip to trying to be dictator and launch wars. So far the LLMs have tended towards bland niceness which is kind of what you want in government.

Re: You are the scariest monster in the woods

#360

Earlier quoted context omitted.

What. Are you saying that "intentionality", whatever you mean by it, can't be implemented by a computational process? Never-ever? Never-ever-ever?

If you don't know what intentionality is, a commonplace term in the philosophy of mind, then I suggest you take some time to better acquaint yourself with the subject matter. This is probably the major blow against any kind of AI fantasy, and you don't do yourself any favors by wading in the shallows and confidently holding to misinformed opinions. So, one last time. A computer program is a set of formal rules that t…

I don't think that "philosophy of mind" contains anything useful for AI development. Most of what I've come across in the field was worthless wordcel drivel, far divorced from reality and any practical applications.
Post reply on HN