Nov 25, 2025 · 1h 0m · latent-space
After LLMs: Spatial Intelligence and World Models — Fei-Fei Li & Justin Johnson, World Labs
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
AI pioneers Fei-Fei Li and Justin Johnson discuss the founding of World Labs and their flagship 3D generative platform, Marble, exploring how the frontier of artificial intelligence is moving beyond text-based large language models toward spatial intelligence, physical world models, and interactive 3D environments.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 4.7 Guest teaching 4.4 Guest disagreement 2.1 The hosts pushing back 1.9
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 42:36 Fei-Fei bluntly rejects 'data center full of Einsteins' framing
Fei-Fei immediately and flatly dismisses the popular industry soundbite cited by the host, refusing the premise and insisting on a rigorous evolutionary and psychological definition of intelligence.
Hardest push from the hosts ▶ 45:31 Alessio challenges pure spatial primacy using Newtonian formalismAlessio pushes back against downplaying language by pointing out that language formalizes empirical spatial phenomena into actionable laws like gravity, forcing the guests to address theory building.
Biggest teaching moment ▶ 56:58 Justin educates on transformer architectural foundations as set processorsJustin fundamentally re-educates the host on deep learning theory, correcting the assumption that transformers are sequence models by demonstrating they are permutation-equivariant set models where sequence order is merely injected via embeddings.
The host holds their own ▶ 22:51 Alessio cites Harvard orbital vector paper to challenge LLM world understandingAlessio demonstrates strong domain mastery by introducing an empirical Harvard study showing LLMs predict planetary orbits without capturing underlying physical force vectors.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Origins of World Labs and Founder Backgrounds | 3 | 1 | 0 | 0 | The hosts open with light background questions asking how Fei-Fei Li and Justin Johnson started World Labs. The dynamic is purely conversational, warm, and collegial with no friction or substantive debate. | |
| Compute Scaling, Academic Research, and Future Hardware | 5 | 4 | 2 | 3 | Alessio and Swix probe whether academic open challenges like ImageNet still work under modern compute and commercial pressures. Fei-Fei clarifies that the issue is academic under-resourcing rather than open versus closed research, while Justin pushes back slightly against the 'hardware lottery' assumption by analyzing GPU performance per watt. | |
| The Evolution of Image Captioning and Dense Captioning | 4 | 3 | 1 | 1 | The guests recount the historical lineage of neural image captioning and dense captioning from their early Stanford lab days. The hosts act primarily as engaged facilitators asking technical follow-ups about forward passes and real-time execution. | |
| Pixel Maximalism, Physical Laws, and Machine Understanding | 6 | 5 | 3 | 3 | Alessio brings up a technical paper on LLMs failing to represent orbital force vectors, sparking a debate on whether latent models can truly learn causal physical laws versus pattern matching. Fei-Fei and Justin draw clear philosophical distinctions between statistical pattern fitting and true human-style understanding. | |
| Marble Architecture, Gaussian Splats, and Practical Applications | 5 | 4 | 2 | 2 | The hosts inquire into the fundamental data representations underlying Marble and Gaussian splats, asking why embodied robotics wasn't emphasized. Fei-Fei clarifies that simulation for robotics is already featured prominently in their plans, explaining the bridge between synthetic data generation and embodied learning. | |
| Defining Spatial Intelligence, Embodiment, and Theory Building | 6 | 6 | 3 | 3 | Fei-Fei immediately rejects the common industry framing of 'a data center full of Einsteins', reframing intelligence as multimodal and highlighting how spatial perception took 540 million years of biological evolution. Alessio counters by noting how language formalizes physical laws like Newton's, leading Justin to separate embodied experience from symbolic theory building. | |
| Re-evaluating Model Architectures: Transformers as Set Processors | 4 | 8 | 4 | 1 | When Swix asks if sequence-to-sequence modeling and attention are obsolete for world models, Justin directly corrects the technical premise by explaining that transformers are natively permutation-equivariant set processors rather than sequence models. Swix accepts the clarification as Justin walks through positional embeddings and token-level operations. |
Statements from this episode (15)
Li: World Labs was founded to build spatial intelligence beyond LLMs
“I think around, you know, more than two years ago, for sure, I think both independently, both of us have been looking at the development of the large models and thinking about What's beyond language models and this idea of building world models, spatial intell…”
Johnson: AI compute per model has scaled one million-fold since 2012
“And if you think about, you know, AlexNet required this jump from CPUs to GPUs, but even from AlexNet to today, we're getting about a thousand times more performance per card than we had in AlexNet days. And now it's common to train models, not just on one GPU…”
Johnson: Academic labs can no longer train state-of-the-art AI on few GPUs
“Like five or 10 years ago, you really could train state-of-the-art models in the lab even with just a couple of GPUs. But, you know, because that technology was so successful and scaled up so much, then you can't train state-of-the-art models with a couple of …”
Johnson: Nvidia Blackwell offers roughly same performance per watt as Hopper
“Like, if you look at the numbers, like, even going from Hopper to Blackwell, like, the performance per watt is about the same. They mostly make the number of transistors go up, and they make the chip size go up, and they make the power usage go up. But even fr…”
Li: 3D and 4D spatial structure is fundamentally unlike 1D text
“I do think the architecture of these generative models will share a lot of shareable components, but I think the deeply three D four D spatial world has a level Of structure that is fundamentally different from a purely generative signal that is one dimensiona…”
Johnson: Pixels offer a more lossless world representation than tokenized text
“And then like you actually lose something if you translate to this like purely tokenized representations that we use in LLMs, right? Like you lose the font, you lose the line breaks, you lose sort of the two D arrangement on the page. And for a lot of cases, f…”
Li: Scaling can bridge visual generative AI and physical engineering
“I think this is a matter of scaling data and bettering model. I don't think there's anything fundamental that separates these two.”
Li: Marble is the first public high-fidelity 3D generative world model
“It's the first in-class model in the world that generates three D worlds in this level of fidelity that is in the hands of the public.”
Johnson: Gaussian splats render in real time on nearly any client device
“Gaussian splats are really cool because you can render them in real time really efficiently. So you can render on your iPhone, render, render everything. And that's how we get that sort of precise camera control because The splats can be rendered real time on …”
Li: Marble has real potential to generate synthetic data for robotics
“Marble actually is a, Really potential for helping to generate these synthetic simulated worlds for embodied agent training.”
Li: Evolution spent 540M years on spatial intelligence versus 0.5M on language
“And in nature, you know, it took five hundred, forty million years to optimize perception and spatial intelligence and language in the most Generous estimation of language development is probably half a million years.”
Li: LLMs can predict physical trajectories without abstracting physical laws
“I wouldn't be surprised that given enough celestial movement data, an LLM would actually predict pretty accurate movement trajectories... I wouldn't be surprised, but F equals MA or, you know, action equals reaction. That's just a whole different abstraction l…”
Johnson: Physical theory-building stems from interactive falsification, not model modality
“Because we're constantly interacting with the world, we're constantly having to build theories about what's happening in the world around us, and then falsify or add evidence to those theories. And I think that that kind of process writ large and scaled up is …”