Comment by tinco

>Knowledge models, like ontologies, always seem suspect to me; like they promise a schema for crisp binary facts, when the world is full of probabilistic and fuzzy information loosely categorized by fallible humans based on an ever slowly shifting social consensus.

I don't disagree that the world is full of fuzziness. But the problem I have with this portrayal is that formal models are often normative rather than analytical. They create reality rather than being an interpretation or abstraction of reality.

People may well have a fuzzy idea of how their credit card works, but how it really works is formally defined by financial institutions. And this is not just true for software products. It's also largely true for manufactured products. Our world is very much shaped by artifacts and man-made rules.

Our probabilistic, fuzzy concepts are often simply a misconception. That doesn't mean it's not important of course. It is important for an AI to understand how people talk about things even if their idea of how these things work is flawed.

And then there is the sort of semi-formal language used in legal or scientific contexts that often has to be translated into formal models before it can become effective. Law makers almost never write algorithms (when they do, they are often buggy). But tax authorities and accounting software vendors do have to formally model the language in the law and then potentially change those formal definitions after court decisions.

My point is that the way in which the modeled, formal world interacts with probabilistic, fuzzy language and human actions is complex. In my opinion we will always need both. AIs ultimately need to understand both and be able to combine them just like (competent) humans do. AI "tool use" is a stop-gap. It's not a sufficient level of understanding.

World models are trivial. eg narratives are world models and they provide only pre frontal simulation, ie they are synthetically prey-predation. No animal uses world models for survival and doubtful they exist (maps are not models), a world model doesn't conform to optic flow, ie instantaneous use and response. Anything like a world model isn't shallow, the basic premise of oscillatory command, it's needlessly deep, nothing like brains. This is just a frontier hail-mary to the current age.

Is it that fuzzy though? If it was would language not adequately grasp and model our realities? And what about the physical world itself: animals are modeling the world adequately enough to navigate it. There's significant gains to make from modeling _enough_ of the world, without falling into hallucinations of purely statistical associations of an LLM.

You're basically describing the knowledge problem vs model structure, how to even begin to design a system which self-updates/dynamically-learns vs being trained and deployed.

Cracking that is a huge step, pure multi-modal trained models will probably give us a hint, but I think we're some ways from seeing a pure multi-modal open model which can be pulled apart/modified. Even then they're still train and deploy not dynamically learning. I worry we're just going to see LSTM design bolted onto deep LLM because we don't know where else to go and it will be fragile and take eons to train.

And less said about the crap of "but inference is doing some kind of minimization within the context window" the better, it's vacuous and not where great minds should be looking for a step forwards.

I have vague notions of there being an entire hidden philosophical/political battlefield (massacre?) behind the whole "are knowledge models/ontologies a realistic goal" debate.

Starting with the sophomoric questions of the optimist who mistakes the possible for the viable: how definite of a thing is "the world", how knowable is it, what is even knowledge... and then back through the more pragmatic: by whom is it knowable, to what degree, and by what means. The mystics: is "the world" the same thing as "the sum of information about the world"? The spooks: how does one study those fields of information which are already agentic and actively resist being studied by changing themselves, such as easily emerge anywhere more than n(D) people gather?

Plenty of food for thought from why ontologies are/aren't a thing. The classical example of how this plays out in the market being search engines winning over internet directories. But that's one turn of the wheel. Look at what search engines grew into quarter century later. What their outgrowths are doing to people's attitude towards knowledge. Different timescale, different picture.

Fundamentally, I don't think human language has sufficient resolution to model large spans of reality within the limited human attention span. The physical limits of human language as information processing device have been hit at some point in the XX century. Probably that 1970s divergence between productivity and wages.

So while LLMs are "computers speak language now" and it's amazing if sad that they cracked it by more data and not by more model, what's more amazing is how many people are continually ready to mistake language for thought. Are they all P-zombies or just obedience-conditioned into emulating ones?!?!?

Practically, what we lack is not the right architecture for "big knowing machine", but better tools for ad-hoc conceptual modeling of local situations. And, just like poetry that rhymes, this is exactly what nobody has a smidgen of interest to serve to consumers, thus someone will just build it in their basement in the hope of turning the tables on everyone. Probably with the help of LLMs as search engines and code generators. Yall better hurry. They're almost done.

They knew what Yann LeCun was when they hired him. If anything, those brilliant academics who have done what they're told and loyally pursued corporate objectives the way the corporation wanted (e.g. Karpathy when he was at Tesla) haven't had great success either.

> risky things that will work

Things known to work are not risky. Risky things can fail by definition.

“Risky things that will work” - contradiction in terms. If companies only did things they knew would work, we probably still wouldn’t have microchips.

Also, like… it’s Facebook. It has a history of ploughing billions into complete nonsense (see metaverse). It is clearly not particularly risk averse.

What exactly does it mean for something to be a "risky thing that will work"?

How do you know that for sure? How can you be absolutely certain that LLMs are what will lead to AGI?

Either that, or just tautological, given that LLM tech is continually morphing and improving.

This is the first time I have heard of world models. Based on my brief reading it does look like this is the idea model for autonomous driving. I wonder if the self driving companies are already using this architecture or something close to it.

And what's the end result? All one can see is just bigger representation of those who confidently subscribe to false information and become arrogant when their validity is questioned, as the LLM writing style has convinced them it's some sort of authority. Even people on this website are so misinformed to believe that ChatGPT has developed its own reasoning, despite it being at the core an advanced learning algorithm trained on a enormous amount of human generated data.

And let's not speak about those so deep into sloth that put it into use to deteriorate, and not augment as they claim to do, humane creative recreational activities.

https://archive.ph/fg7HE

> because the AI's lack axiomatic systems beyond much manually constructed machinery.

Oh god, that is massively under-selling their learning ability. These models are able to extract and reply with why jokes are funny without even knowing basic vocab, yet there are pure-code models out there with lingual rules baked in from day one which still struggle with basic grammar.

The _point_ of LLMs arguably is there ability to learn any pattern thrown at it with enough compute. With an exception to learning how logical processes work, and pure LLMs only see "time" in the sense of a paragraph begins and ends.

At the least they have taught computers, "how to language", which in regards to how to interact with a machine is a _huge_ step forward.

Unfortunately the financial incentives are split between agentic model usage (taking the idea of a computerised butler further), maximizing model memory and raw learning capacity (answering all problems at any time), and long-range consistency (longer ranges give better stable results due to a few reasons, but we're some way from seeing an LLM with a 128k experts and 10e18 active tokens).

I think in terms of building the perfect monkey butler we already have most or all of the parts. With regard to a model which can dynamically learn on the fly... LLMs are not the end of the story and we need something to allow the models to more closely tie their LS with the context. Frankly the fact that DeepSeek gave us an LLM with LS was a huge leap since previous model attempts had been overly complex and had failed in training.