Why Fei-Fei Li and Yann LeCun Are Both Betting on "World Models"
entropytown.com
entropytown.com
Biology does not see 3-D, we integrate two 2-D inputs and make a 2.5-D from them in varying degrees of resolution. Within the brain our mapping systems and sensory/emotional/memory architectures range from no D (they are affinities of exploding areas) to 2-D topology, which integrate from whole brain to the finer aspects of mapping in what we understand so far are unique combinations of allo and egocentric scales of space.
Also as thought and words (and symbols, representations, metaphors, add anything arbitrary here in our weak externals) are divergent and unrelated, the manifest idea these approaches link them as a route to an unsupportable and additonal layer of "world modelling" is inferior innovation (if can even be called innovation). These are all steps back from the integrations we see biointelligence operating with. This is synthetically deciding that intelligence can be simplified with limited general processes and then form fitting them into a binary code to mimic it. It's really poor ideation.
None of this requires models (maps are not models), and the entire idea that a model is being used as a gateway to intelligence as redundant and oxymoronic as in any "world model". In essence, this is the last range of disability to achieve intelligence from binary, which in itself is a poor form to interpolate the oscillatory dynamical nature of consciousness/intelligence. It was very premature to develop this phase of code like this.
To follow completely in your footsteps would require recreating the totality of physical evolution that led to intelligence within a fraction of the time. This isn't feasible to an investor expecting returns within a reasonable timeline.
Getting to this point leaped from text generation to image generation to video generation, and now these three approaches are experimenting on what the next step should be. These three approaches did not come from an isolated vacuum and are the result of iterative progress.
The general idea now is to take the video model and give it 3D spatial capabilities to better model the implicitly symbolic and virtual worlds it is depicting and reason what would happen next. Fei-Fei Li wants to produce a 3D scene asset. DeepMind wants to simulate what can happen in that 3D scene. Yann LeCun wants to expand upon the symbolic reasoning by adding another layer of intelligence.
Traditional AI balk at the lack of inherent agentic purpose and goals, but LLMs separately evolved from pattern matching and statical analysis of the digital output of human labor.
A number of people in the LLM field have accepted that recreating the animal brain is not the point. Instead they work on a unique digital intelligence as if it was select fragments of the human brain existing in a digital world, informed by neuroscience research.
I think LLMs may not reason like a human with skin in the game, but humans are rather flawed in making sense of the nihilistic stage of history we find ourselves in. It is difficult for at least half of humanity on an IQ level to reason with what we have created. I think there is a case for a separate digital intelligence to analyze and make sense of the digital world which only seems to merge further with reality. Maybe this is a transhumanist singularity not in a technological term, but in terms of human idealism in creating values for ourselves.
They’re irrelevant in terms of any idea of intelligence as intel is built upon topologies.
Intelligence is tied to development where functional outcomes are by genetic tinkering with environments for flexibility. These gaslighted trivial models are function aimed, so they have zero ability to even mimic what intel is.
Face it this is a massive washout. It has no ambition. It lacks even a weak definition of the models name “spatial intelligence” and it lacks one because there is no such thing.
Fundamentally world models do not exist and are oxymoronic.
The real concern of having no biomechanical skin in the game is lacking sensory input that could ground it within our reality. All input into LLMs are based on digital output of human labor, which are ultimately symbolic representations filtered through our brain and its ideas of reality. However, this may not be too different from how our real human brains work.
There has been a philosophical dilemma over how real consciousness can be as if it is imagined by our brains since our brains provide convincing hallucination of what seems like real sensory input or even free will. That is to say that humans at a philosophical level live in their brains interpreting a fragment of reality based upon how it interprets sensory input.
Now the LLM as a brain cuts out an entire step of agentic sensory input and they exist wholly as the result of our ideas.
There were never approximations merely false models.
The field is trapped in bad definitions and decisions
“We refute (based on empirical evidence) claims that humans use linguistic representations to think.” Ev Fedorenko Language Lab MIT
https://pmc.ncbi.nlm.nih.gov/articles/PMC4874898/
And btw, that’s not a blanket statement that’s an empirical statement that wipes away quite a bit of LLM relevance. I’d say it destroys the approach.
Do the research. And an apology is in good order.
This paper does little to dismiss LLMs. LLMs can use a different medium than text and that would not take away from its underlying mathematical models based on neuroscience. LLMs only understand language representations implicitly through statistical analysis, and that may instead show a commonality with how the human brain thinks as written in this paper.
I will not apologize for how you keep pushing an agenda despite how poorly supported it is. I have tried to be intellectually honest about the state of the industry and its flaws. I would implore you to instead do the research about LLMs so you can better refine your critique of them.
https://docs.google.com/document/d/1cXtU97SCjxaHCrf8UVeQGYaj...
There is a baseline expectation of how quotes and citations are supposed to work within Western intellectual circles. The fact that you do not know them and refuse to accept it means either you are not familiar with Western academia or you are an intellectually dishonest Internet troll or an LLM bot.
Spatial reasoning and world models are a research topic because elements of them were found in video and agentic models, and investors want to further refine either of them.
I do not have the time to read through this entire Google doc, but from what I have skimmed, I can see that the most substantial critiques are from academia being honest of the current state of AI and its limitations. That is fine.
However, the opening paragraphs aren't impressive. Language is arbitrary, yes, but they must also be intelligible by other humans. It is like a canvas to pattern match and create all sorts of inductive reasonings. There isn't much to explain how pattern matching math would be inherently incapable of pattern matching the written language. This reads like a basic understanding of postmodernist philosophy as if it is proof of math becoming a failure when applied to a socially constructed reality. However, philosophy and other social sciences do not surrender and give up as if their fields are fundamentally flawed. They make do and continue matching patterns to make observations of social reality.
The burden is ultimately on you to prove that the limitations of current AI/LLM cannot be overcome or that there is something that cannot make world models or spatial reasoning possible. Simply having a mountain of text to read is not an argument. There has to be some summary or point that can be used at the thrust of your position. As they say, brevity is the soul of wit.
For example, "LKNs Gaussian frontier" is another random phrase you have pulled out as if you are an LLM hallucinating something.
The bubble is in whether the investors will get their return on time. This is orthogonal to the underlying technology. Investor interests are not in progressing technology but to get a profit. Hope you enjoy the ride too because this is going to affect all of us.
Laughing all the way at the AI clown show.
Is "task demand" what the LLM would expect to do in order to respond to the user prompt? It seems incredulous that semantics would only exist here. As I have mentioned before, semantics is already embedded in the input and output for the LLM to implicitly discover and model and reason with.
https://arxiv.org/html/2507.05448v1 This paper is an interesting overview of semantics in LLMs. Here's an interesting quote, "Whether these LLMs exhibit semantic capabilities, is explored through the classical semantic theory which goes back to Frege and Russell. We show that the answer depends on how meaning is defined by Frege and Russell (and which interpretation one follows). If meaning is solely based on reference, that is, some referential capability, LLM-generated representations are meaningless, because the text-based LLMs representation do not directly refer to the world unless the reference is somehow indirectly induced. Also the notion of reference hinges on the notion of truth; ultimately it is the truth that determines the reference. If meaning however is associated with another kind of meaning such as Frege’s sense in addition to reference, it can be argued that LLM representations can carry that kind of semantics."
As for reference-based meaning reliant on truth, this was mentioned earlier in the paper, "An alternative to addressing the limitations of existing text-based models is the development of multimodal models, i.e., DL models that integrate various modalities such as text, vision, and potentially other modalities via sensory data. Large multimodal models (LMMs) could then ground their linguistic and semantic representations in non-textual representations like corresponding images or sensor data, akin to what has been termed sensorimotor grounding (Harnad, 1990). Note however, that such models would still not have direct access to the world but a mediated access: an access that is fundamentally task- driven and representational moreover. Also, as we will argue, the issue is rather that we need to ground sentences, rather than specific representations, because it is sentences that may refer to truth. But attaining the truth is not an easy task; ultimately, future LMMs will face the same difficulties as we do in determining truth."
In other words, this is the approach Fei-Fei Li and other multimodal models are using to create the world model.