It's hard to describe, but it's felt like LLMs have completely sucked the entire energy out of computer vision. Like... I know CVPR still happens and there's great research that comes out of it, but almost every single job posting in ML is about LLMs to do this and that to the detriment of computer vision.
I tried LLM's for geolocation recently and it is both amazing how good they are at recognizing patterns and how terrible they are with recognizing and utilizing basic spatial relationships.
I appreciate the video and generally agree with Fei-Fei but I think it almost understates how different the problem of reasoning about the physical world actually is.
Most dynamics of the physical world are sparse, non-linear systems at every level of resolution. Most ways of constructing accurate models mathematically don’t actually work. LLMs, for better or worse, are pretty classic (in an algorithmic information theory sense) sequential induction problems. We’ve known for well over a decade that you cannot cram real-world spatial dynamics into those models. It is a clear impedance mismatch.
There are a bunch of fundamental computer science problems that stand in the way, which I was schooled on in 2006 from the brightest minds in the field. For example, how do you represent arbitrary spatial relationships on computers in a general and scalable way? There are no solutions in the public data structures and algorithms literature. We know that universal solutions can’t exist and that all practical solutions require exotic high-dimensionality computational constructs that human brains will struggle to reason about. This has been the status quo since the 1980s. This particular set of problems is hard for a reason.
I vigorously agree that the ability to reason about spatiotemporal dynamics is critical to general AI. But the computer science required is so different from classical AI research that I don’t expect any pure AI researcher to bridge that gap. The other aspect is that this area of research became highly developed over two decades but is not in the public literature.
One of the big questions I have had since they announced the company, is who on their team is an expert in the dark state-of-the-art computer science with respect to working around these particular problems? They risk running straight into the same deep, layered theory walls that almost everyone else has run into. I can’t identify anyone on the team that is an expert in a relevant area of computer science theory, which makes me skeptical to some extent. It is a nice idea but I don’t get the sense they understand the true nature of the problem.
ah but neural networks are universal function approximators! (proceeds to ignore the size of network needed for an adequate approximation and/or how much data would be required to train it)
makes sense - humans have evolved a lot of wetware dedicated to 3D processing from stereo 2D.
I've made some progress on a PoC in 3D reconstruction - detecting planes, edges, pipes from pointclouds from lidar scans, eg : https://youtu.be/-o58qe8egS4 .. and am bootstrapping with in-house gigs as I build out the product.
Essentially it breaks down to a ton of matmulls, and I use a lot of tricks from pre-LLM ML .. this is a domain that perfectly fits RL.
The investors Ive talked to seem to understand that scan-to-cad is a real problem with a viable market - automating 5Bn / yr of manual click-labor. But they want to see traction in the form of early sales of the MVP, which is understandable, especially in the current regime of high interest rates.
Ive not been able to get across to potential investors the vast implications for robotics, AI, AR, VR, VFX that having better / faster / realtime 3D reconstruction will bring. Its great that someone of the caliber of Fei-Fei Li is talking about it.
Robots that interact in the real world will need to make a 3D model in realtime and likely share it efficiently with comrades.
While a gaussian splat model is more efficient than a pointcloud, a model which recognizes a wall as a quad plane is much more efficient still, and needed for realtime communication. There is the old idea that compression is equivalent to AI.
What is stopping us from having a google street-view v3.0 in which I can zoom right into and walk around a shopping mall, or train station or public building ? Our browsers can do this now, essentially rendering quake like 3D environments - the problem is with turning a scan into a lightweight 3D model.
Photogrammetry, where you have hundreds of photos and reconstruct the 3D scene, uses a lot of compute, and the colmap / Structure-from-Motion algorithm predates newer ML approaches and is ripe for a better RL algorithm imo. Ive done experiments where you can manually model a 3D scene from well positioned 360 panorama photos of a building, picking corners, following the outline of walls to make a floorplan etc ... this should be amenable to an RL algorithm. Most 360 panorama photo tours have enough overlap to reconstruct the scene reasonably well.
I have no doubt that we are on the brink of a massive improvement in 3D processing. Its clearly solvable with the ML/RL approaches we currently have .. we dont need AGI. My problem is getting funding to work on it fulltime, equivalently talking an investor into taking that bet :)
I've always wondered how spatial reasoning appears to be operating quite differently from other cognitive abilities, with significant individual variations. Some people effortlessly parallel park while others struggle with these tasks despite excelling at other forms of pattern recognition. What was particularly intriguing for me is that some people with aphantasia have no difficulty with spatial reasoning tasks, so spatial reasoning may be distinct from reasoning based on internal visualization.
I mean just walking down the street or through a supermarket, it seems to me like 95% of people have no spatial awareness at all. Walking forward while looking to the side or backward.
Either that or they're perfectly capable, they just don't care.
Most of our spatial intelligence is innate, developed through evolution. We're born with a basic sense of gravity and the ability to track objects. When we learn to drive a car, we simply reassign these built-in skills to a new context
we're actually working on a practical implementation of aspects of what Fei-Fei describes - although with a more narrow focus on optimizing operations in the physical space (mining, energy, defense etc) https://hivekit.io/about/our-vision/
We've been working on this challenge in the satellite domain with https://earthgpt.app. It’s a subset of what Fei-Fei is describing, but comes with its own unique issues like handling multi-resolution sensors and imagery with hundreds of spectral bands. Think of it as computer vision, but in n-dimensions.
Happy to answer questions if you're curious. PS. still in early beta, so please be gentle!
An immaterial side note: funny how obsessed she seems to be with her age. She said once that people in the audience could be half or even third of her age.
Given that she's 49, is it really typical that 16-year olds attend these fireside YC chats?
Intelligence is not only embodied (it needs a body), it is also embedded in the environment (it needs the environment). If you want an intelligence in your computer, you need an environment in your computer first, as the substrate from which the intelligence will evolve. The more accurate the environment the better the intelligence that will be obtained.
The universe is able to create intelligence and we are proof. Thus, if you want to create intelligence, you have to find a way of efficiently simulate our reality at the desired level of detail.
Currently we don’t know such efficient algorithm, but one way could be finally harnessing Quantum Computing to hack the universe itself, cheat and be able to simulate our environment efficiently without even knowing the algorithm behind Quantum Physics.
How can I be sure that spatial intelligence AIs will not be just intricate sensoring that ultimately fails to demonstrate actual intelligence?
> "trilobite"
The trilobite ancestor had a nervous system before it had an eye. It was able to make decisions and interact with the environment before the ability to see or speak a language.
It feels to me like this basic step is still missing. We haven't even crossed the first AI frontier yet.
Great talk. Dr. Li has a way of cutting through the hype and getting to the fundamental challenges that is really refreshing. Her point about spatial intelligence being the next frontier after language really resonates.
I'm particularly hung up on the data problem she touched on (41 min). She rightly points out that unlike language, where we could bootstrap LLMs with the vast, pre-existing corpus of the internet, there's no equivalent "internet of 3D space." She mentions a "hybrid approach" for World Labs, and that's where the real engineering challenge seems to lie.
My mind immediately goes to the trade-offs. If you lean heavily on synthetic data, you're in a constant battle with the "sim-to-real" gap. It works for narrow domains, but for a general "world model," the physics, lighting, and material properties have to be perfect, which is a monumental task. If you lean on real-world capture (e.g., massive-scale photogrammetry, NeRFs, etc.), the MLOps and data pipeline challenges seem staggering. We're not just talking text files; we're talking about petabytes of structured, multi-sensor data that needs to be processed, aligned, and labeled. It feels like an entirely new class of data infrastructure problem.
Her hiring philosophy of "intellectual fearlessness" (31 min) makes a lot of sense in this context. You'd need a team that's not intimidated by the fact that the foundational dataset for their entire field doesn't even exist yet. They have to build the oil refinery while also figuring out where to drill for oil.
It's exciting to see a team with this much deep learning and computer vision firepower aimed at such a foundational problem. It pulls the conversation away from just optimizing existing architectures and towards creating entirely new categories. It leaves me wondering: what does the "AlexNet moment" for spatial intelligence even look like? Is it a novel model architecture, or is the true breakthrough a new form of data representation that makes this problem tractable at scale?
She says "there is no language in nature" which does not seem accurate. Even though she might mean something else or a particular form of language but even then, bees and birds still use sound and something similar to language.
Is it just me?
for e.g. the form of communication used by bees is very well known now, it involves not just spatial movements but also "buzzing" which is totally similar tot he sounds we make, they just lack vocal cords.
I’m surprised no one compares her work more directly with Elon Musk’s efforts in autonomous vehicles and robotics — seems like there’s overlap. Anyone know more about how their approaches differ?
28 comments
[ 2.6 ms ] story [ 44.3 ms ] threadMost dynamics of the physical world are sparse, non-linear systems at every level of resolution. Most ways of constructing accurate models mathematically don’t actually work. LLMs, for better or worse, are pretty classic (in an algorithmic information theory sense) sequential induction problems. We’ve known for well over a decade that you cannot cram real-world spatial dynamics into those models. It is a clear impedance mismatch.
There are a bunch of fundamental computer science problems that stand in the way, which I was schooled on in 2006 from the brightest minds in the field. For example, how do you represent arbitrary spatial relationships on computers in a general and scalable way? There are no solutions in the public data structures and algorithms literature. We know that universal solutions can’t exist and that all practical solutions require exotic high-dimensionality computational constructs that human brains will struggle to reason about. This has been the status quo since the 1980s. This particular set of problems is hard for a reason.
I vigorously agree that the ability to reason about spatiotemporal dynamics is critical to general AI. But the computer science required is so different from classical AI research that I don’t expect any pure AI researcher to bridge that gap. The other aspect is that this area of research became highly developed over two decades but is not in the public literature.
One of the big questions I have had since they announced the company, is who on their team is an expert in the dark state-of-the-art computer science with respect to working around these particular problems? They risk running straight into the same deep, layered theory walls that almost everyone else has run into. I can’t identify anyone on the team that is an expert in a relevant area of computer science theory, which makes me skeptical to some extent. It is a nice idea but I don’t get the sense they understand the true nature of the problem.
Nonetheless, I agree that it is important!
I've made some progress on a PoC in 3D reconstruction - detecting planes, edges, pipes from pointclouds from lidar scans, eg : https://youtu.be/-o58qe8egS4 .. and am bootstrapping with in-house gigs as I build out the product.
Essentially it breaks down to a ton of matmulls, and I use a lot of tricks from pre-LLM ML .. this is a domain that perfectly fits RL.
The investors Ive talked to seem to understand that scan-to-cad is a real problem with a viable market - automating 5Bn / yr of manual click-labor. But they want to see traction in the form of early sales of the MVP, which is understandable, especially in the current regime of high interest rates.
Ive not been able to get across to potential investors the vast implications for robotics, AI, AR, VR, VFX that having better / faster / realtime 3D reconstruction will bring. Its great that someone of the caliber of Fei-Fei Li is talking about it.
Robots that interact in the real world will need to make a 3D model in realtime and likely share it efficiently with comrades.
While a gaussian splat model is more efficient than a pointcloud, a model which recognizes a wall as a quad plane is much more efficient still, and needed for realtime communication. There is the old idea that compression is equivalent to AI.
What is stopping us from having a google street-view v3.0 in which I can zoom right into and walk around a shopping mall, or train station or public building ? Our browsers can do this now, essentially rendering quake like 3D environments - the problem is with turning a scan into a lightweight 3D model.
Photogrammetry, where you have hundreds of photos and reconstruct the 3D scene, uses a lot of compute, and the colmap / Structure-from-Motion algorithm predates newer ML approaches and is ripe for a better RL algorithm imo. Ive done experiments where you can manually model a 3D scene from well positioned 360 panorama photos of a building, picking corners, following the outline of walls to make a floorplan etc ... this should be amenable to an RL algorithm. Most 360 panorama photo tours have enough overlap to reconstruct the scene reasonably well.
I have no doubt that we are on the brink of a massive improvement in 3D processing. Its clearly solvable with the ML/RL approaches we currently have .. we dont need AGI. My problem is getting funding to work on it fulltime, equivalently talking an investor into taking that bet :)
Either that or they're perfectly capable, they just don't care.
Happy to answer questions if you're curious. PS. still in early beta, so please be gentle!
https://community.openai.com/t/time-awareness-in-ai-why-temp...
https://boraerbasoglu.medium.com/the-impact-of-ais-lack-of-t...
[1] https://arxiv.org/abs/2507.01955
Once that happens it’s all over.
Here's an on the fly video I made (no retakes) of Claude generating a Godot scene file.
https://youtu.be/2gARJpDG7Jo?si=W4rlISO-J4EPJYyG
https://spatialwebfoundation.org/
> "trilobite"
The trilobite ancestor had a nervous system before it had an eye. It was able to make decisions and interact with the environment before the ability to see or speak a language.
It feels to me like this basic step is still missing. We haven't even crossed the first AI frontier yet.
Enough said.
I'm particularly hung up on the data problem she touched on (41 min). She rightly points out that unlike language, where we could bootstrap LLMs with the vast, pre-existing corpus of the internet, there's no equivalent "internet of 3D space." She mentions a "hybrid approach" for World Labs, and that's where the real engineering challenge seems to lie.
My mind immediately goes to the trade-offs. If you lean heavily on synthetic data, you're in a constant battle with the "sim-to-real" gap. It works for narrow domains, but for a general "world model," the physics, lighting, and material properties have to be perfect, which is a monumental task. If you lean on real-world capture (e.g., massive-scale photogrammetry, NeRFs, etc.), the MLOps and data pipeline challenges seem staggering. We're not just talking text files; we're talking about petabytes of structured, multi-sensor data that needs to be processed, aligned, and labeled. It feels like an entirely new class of data infrastructure problem.
Her hiring philosophy of "intellectual fearlessness" (31 min) makes a lot of sense in this context. You'd need a team that's not intimidated by the fact that the foundational dataset for their entire field doesn't even exist yet. They have to build the oil refinery while also figuring out where to drill for oil.
It's exciting to see a team with this much deep learning and computer vision firepower aimed at such a foundational problem. It pulls the conversation away from just optimizing existing architectures and towards creating entirely new categories. It leaves me wondering: what does the "AlexNet moment" for spatial intelligence even look like? Is it a novel model architecture, or is the true breakthrough a new form of data representation that makes this problem tractable at scale?
Is it just me?
for e.g. the form of communication used by bees is very well known now, it involves not just spatial movements but also "buzzing" which is totally similar tot he sounds we make, they just lack vocal cords.
https://www.noemamag.com/how-to-speak-honeybee/
https://videotobe.com/play/youtube/_PioN-CpOP0