Bloomberg Originals · Published 2026-08-19

What's Beyond ChatGPT? The 'Godmother of AI' Has a Plan | The Circuit

privateReview: unreviewedConfidence: high
world modelsspatial intelligenceimagenetroboticsartificial intelligenceworld-modelsspatial-intelligenceembodied-ai3d-generationdataset-scalingai-policy
Open on YouTube ↗

Summary

Overview

  • Speaker: Emily Chang and Fei-Fei Li
  • Channel: Bloomberg Originals
  • Main topic: World models and spatial intelligence in AI beyond large language models
  • Purpose: To explore Fei-Fei Li's career, her philosophy on human-centric AI, and World Labs' mission to develop world models and spatial intelligence. Emily Chang interviews Fei-Fei Li, co-founder and CEO of World Labs and often called the 'Godmother of AI' for her pioneering work on ImageNet. They discuss her journey from a curious child in China and student at Princeton to leading the AI revolution, and her latest venture into world models and spatial intelligence. Li argues that while the rest of the AI industry focuses heavily on large language models (LLMs) like ChatGPT, the next frontier is building AI that can understand, predict, and interact with the physical world through spatial intelligence.

Topic Map

Introduction to Fei-Fei Li and World Labs

  • Explanation: Introduction of Fei-Fei Li in San Francisco, discussing her background, her company World Labs, and her relentless advocacy for AI that serves people.
  • Key claims:
    • Leaders of AI talk too much about replacing humanity rather than empowering it.
    • World Labs is building AI that predicts what happens next in the real world.
  • Examples:
    • Walking through San Francisco parks and offices.
  • Terminology:
    • World Models
    • Spatial Intelligence
  • Why it matters: Establishes the core thesis that spatial intelligence is the next major leap in AI beyond text-based LLMs.

The ImageNet Breakthrough

  • Explanation: Recounting how Fei-Fei Li built ImageNet, a massive visual library of 14 million pictures, which sparked the modern AI revolution.
  • Key claims:
    • Algorithms needed data to get smarter.
    • Before ImageNet, nobody was paying attention to data.
  • Examples:
    • ImageNet dataset categorization.
    • AlexNet combining data, neural networks, and GPUs.
  • Terminology:
    • ImageNet
    • Computer Vision
    • Neural Networks
    • GPUs
  • Why it matters: ImageNet established the data-driven foundation of modern deep learning.

Personal Background and Scientific Journey

  • Explanation: Fei-Fei Li reflects on growing up in China during the 80s and 90s, immigrating to the US as a teenager, running a family dry-cleaning business while studying physics at Princeton, and earning her PhD in computer science at Caltech.
  • Key claims:
    • Growing up in nature and science shaped her identity.
    • Immigrating as a teenager was tough, shifting culture, language, and life.
  • Examples:
    • Childhood photos in China.
    • Graduation and family dry-cleaning business.
  • Terminology:
    • Physics
    • Computer Science
    • Immigration
  • Why it matters: Provides context on how her lifelong curiosity and resilience shaped her scientific career.

What Are World Models and Spatial Intelligence?

  • Explanation: Defining world models and spatial intelligence, contrasting them with large language models.
  • Key claims:
    • World models aim to predict physical reality, not just the next word in a sentence.
    • Spatial intelligence involves rendering, simulation, and planning.
  • Examples:
    • Sora video generation.
    • Robotics interacting with physical rooms and objects.
  • Terminology:
    • World Models
    • Spatial Intelligence
    • LLMs
    • Rendering
    • Simulation
    • Planning
  • Why it matters: World models represent the necessary step for AI to navigate and operate in the physical world alongside humans.

Business, Regulation, and AI Policy

  • Explanation: Discussing the billion-dollar funding of World Labs, competition with tech giants, the role of regulation, and maintaining a positive societal vision for AI.
  • Key claims:
    • World Labs is building technology and focusing on scientific discovery.
    • Regulation should be rooted in science, not science fiction.
    • The US must maintain a positive path for AI to remain a global leader.
  • Examples:
    • Time Magazine cover discussion.
    • Meetings with US presidents and UN speeches.
  • Terminology:
    • Venture Capital
    • Regulation
    • AGI
    • Ecosystem
  • Why it matters: Highlights the delicate balance between rapid technological innovation, commercial competition, and ethical governance.

Key Points

Data-driven AI revolution started with ImageNet

  • Explanation: Fei-Fei Li realized that teaching machines to see required massive datasets, leading to the creation of ImageNet in 2006.
  • Evidence: 14 million pictures categorized across over 21,000 categories used in the 2012 ImageNet competition.
  • Practical implication: Proved that data scale combined with neural networks and GPUs is the recipe for modern AI breakthroughs.

Moving from LLMs to World Models

  • Explanation: While LLMs process text and predict words, world models focus on spatial intelligence, understanding physics, and predicting physical reality.
  • Evidence: World Labs' Marble platform allows users to generate and explore 3D worlds from text or visual prompts.
  • Practical implication:

Balancing commercial competition with human-centric values

  • Explanation: World Labs competes against deep-pocketed tech giants by focusing on spatial intelligence and empowering human society.
  • Evidence: World Labs raised $1 billion while maintaining a small, highly collaborative team of ~50 people.
  • Practical implication: Small teams can drive foundational paradigm shifts in AI if focused on the right core scientific challenges.

Frameworks, Models & Processes

Spatial Intelligence Framework

  • How it works: Breaks down spatial intelligence into three fundamental functions that allow machines to understand and operate in physical environments.
  • Components:
    • Rendering
    • Simulation
    • Planning
  • When to use: When developing AI systems for robotics, autonomous vehicles, virtual production, and physical world interaction.

Examples & Case Studies

Creation and launch of ImageNet in 2006-2012

  • Illustrates: The shift in AI research from purely algorithmic focus to data-driven deep learning.
  • Lesson: High-quality, large-scale data is essential for training smarter AI models.

World Labs launching Marble platform in 2024

  • Illustrates: Generative 3D world models created from single visual or text prompts.
  • Lesson: AI can go beyond text generation into interactive 3D spatial environments for gaming, film, and robotics.

Actionable Takeaways

  • Immediate:
    • Spatial intelligence is the next major frontier in AI beyond text-based LLMs.
    • World models integrate rendering, simulation, and planning to navigate physical reality.
  • Strategic:
    • AI policy and regulation must be grounded in scientific reality rather than sci-fi fears.
    • Human capital and STEM education are vital to ensuring AI benefits society.
  • Questions to investigate:
    • How quickly will world models integrate with robotics for physical task execution?
    • How will startups compete with tech giants in the resource-intensive world model race?

Claims Worth Verifying

  • ImageNet is a visual library of 14 million pictures in over 21,000 categories. (historical fact)
  • World Labs raised $1 billion in funding. (business fact)

Notable Quotes

"I feel the leaders of AI are not talking enough about empowering humanity. They're talking too much about replacing humanity." (at 0:49) "What's it like running your own shop?" (at 1:42) "I feel like a tiger mom." (at 1:44) "There's nothing artificial about artificial intelligence. It's inspired by people, it's created by people, and most importantly, it has an impact on people." (at 29:58)

Compressed Summary

  • Fei-Fei Li created ImageNet, sparking the modern AI revolution through data-driven computer vision.
  • World Labs focuses on world models and spatial intelligence to predict and navigate the physical world.
  • Spatial intelligence consists of rendering, simulation, and planning.
  • AI regulation and policy must remain grounded in science to support a healthy ecosystem and human capital.
  • Keywords: world models, spatial intelligence, imagenet, robotics, artificial intelligence
  • Core insight: Moving beyond large language models, the next era of AI requires spatial intelligence and world models to understand and interact with the physical world.

Core insights

6
Architecturemedium noveltymoderate evidence

Spatial intelligence decomposes into three distinct functional capabilities — rendering, simulation, and planning — rather than being treated as one monolithic 'understand 3D' model capability. Each function implies a different computational contract: rendering produces observable geometry, simulation produces forward physical consequences, planning produces action sequences. Treating them as separable subsystems is the architectural claim worth testing.

Why it matters

If rendering/simulation/planning are separable, an embodied agent stack can swap vendors or models per function and can be evaluated per function, instead of relying on one opaque end-to-end model whose failures are unattributable. It also suggests where latency budgets belong: planning loops are latency-critical, rendering may be amortizable/cacheable.

Generalization

Complex perception-to-action capabilities are more debuggable and substitutable when decomposed into generation, prediction, and decision layers with distinct interfaces than when fused into a single model.

Spatial intelligence involves rendering, simulation, and planning.
Open source video
World models integrate rendering, simulation, and planning to navigate physical reality.
Open source video
Mental Modelmedium noveltystrong evidence

The ImageNet story is presented as a paradigm inversion: the field's bottleneck moved from algorithms to data. The recorded mechanism is explicit — algorithms did not improve until a 14M-image / 21K-category dataset existed, and the breakthrough (AlexNet) came from combining that data with neural networks and GPUs.

Why it matters

For new capability axes with no consolidated dataset (3D scenes, physical interaction traces), the highest-leverage engineering investment may be dataset-and-benchmark construction, not model architecture — the same ordering that preceded the vision boom. This reframes what a 'pre-training' team should build first.

Generalization

When a modality lacks a canonical large-scale labeled corpus, corpus construction is a precondition for algorithmic progress rather than a downstream task.

Algorithms needed data to get smarter.
Open source video
Before ImageNet, nobody was paying attention to data.
Open source video
14 million pictures categorized across over 21,000 categories used in the 2012 ImageNet competition.
Open source video
Proved that data scale combined with neural networks and GPUs is the recipe for modern AI breakthroughs.
Open source video
Tradeoffmedium noveltyweak evidence

World models are framed as a different prediction target than LLMs: predicting physical reality (what happens next in a scene) versus predicting the next word. The summary cites Sora video generation as an example in the same breath as robotics interacting with rooms and objects — which conflates photorealistic generation with interactive, physically consistent state.

Why it matters

A model can render a plausible future frame without maintaining a consistent, controllable latent state. If teams accept generative video quality as evidence of world modeling, they will ship systems that look right and fail on counterfactual or intervention-based tasks central to planning. It argues for evaluation that probes intervention and state consistency, not just visual fidelity.

Generalization

Fidelity metrics on generated output are weak proxies for a model's usable internal state; interactive/controllability tests are the discriminating benchmark.

World models aim to predict physical reality, not just the next word in a sentence.
Open source video
Sora video generation.
Open source video
Mechanismmedium noveltymoderate evidence

Marble is described as a prompt-to-explorable-3D-world platform: text or a single visual prompt yields a navigable 3D environment. This makes world generation an environment-authoring primitive, not just a media output — it turns prompts into simulatable spaces usable for training or testing embodied policies.

Why it matters

If a text or image prompt can mint a walkable environment, environment supply stops being the bottleneck for embodied agent training and simulation-based evaluation. That changes the harness design: environments become cheap, versioned, and disposable artifacts rather than hand-built assets.

Generalization

Generative artifacts become infrastructure when they can serve as runtime environments for other systems, not merely as content for humans.

World Labs' Marble platform allows users to generate and explore 3D worlds from text or visual prompts.
Open source video
AI can go beyond text generation into interactive 3D spatial environments for gaming, film, and robotics.
Open source video
Empirical Resultmedium noveltymoderate evidence

World Labs reportedly raised $1 billion while operating with roughly 50 people, deliberately pursuing a foundational scientific bet rather than headcount scaling to match incumbents. The summary frames the leverage as focus on the right core scientific challenge rather than resource parity.

Why it matters

This is a concrete counter-example to the assumption that frontier AI requires massive engineering orgs. It suggests capital allocation toward compute and a small set of scientific bets can substitute for organizational scale — relevant when deciding whether to staff a research direction broadly or keep it narrow and dense.

Generalization

In frontier research bets, capital and headcount are partially substitutable; a narrow scientific thesis can substitute for broad organizational scale.

World Labs raised $1 billion while maintaining a small, highly collaborative team of ~50 people.
Open source video
Small teams can drive foundational paradigm shifts in AI if focused on the right core scientific challenges.
Open source video
Predictionlow noveltyweak evidence

The stated application surfaces for spatial intelligence are robotics, autonomous vehicles, virtual production, and physical world interaction — a set defined by requiring closed-loop action in space rather than by text throughput.

Why it matters

It identifies which downstream systems would need a spatial layer first and therefore where integration work and tooling demand will appear: simulators, action interfaces, and safety envelopes for physical actuation rather than prompt/response surfaces.

Generalization

Emerging model classes are best scoped by the loop they close (perception-action) rather than by the modality they consume.

When developing AI systems for robotics, autonomous vehicles, virtual production, and physical world interaction.
Open source video
Spatial intelligence is the next major frontier in AI beyond text-based LLMs.
Open source video

Deep dives

5

Separability of rendering, simulation, and planning in a world-model stack

Research question

If rendering, simulation, and planning are separable subsystems, which of the three can be implemented by non-neural or externally constrained components (e.g., a physics engine) without degrading counterfactual task performance, and does decomposition improve failure attribution?

Why

Determines whether an embodied agent stack can swap or constrain vendors per function, and where latency budgets and evaluation suites must be attached, instead of relying on one opaque end-to-end model whose failures cannot be localized.

Spatial intelligence involves rendering, simulation, and planning.
Open source video
World models integrate rendering, simulation, and planning to navigate physical reality.
Open source video
Source video

What a spatial-intelligence benchmark should measure, built the way ImageNet was

Research question

Can a large-scale, taxonomy-structured evaluation corpus for spatial intelligence be constructed with the same corpus-first ordering as ImageNet, and does it separate world-model capability from multimodal LLM capability?

Why

If corpus construction gates algorithmic progress, the highest-leverage investment for a 3D/physical capability axis is dataset-and-benchmark construction, not architecture. This reframes what a pre-training team builds first and what a benchmark must contain.

Algorithms needed data to get smarter.
Open source video
Before ImageNet, nobody was paying attention to data.
Open source video
14 million pictures categorized across over 21,000 categories used in the 2012 ImageNet competition.
Open source video
Source video

Fidelity versus controllable state consistency in world models

Research question

Do models that maximize visual fidelity of predicted futures also maintain a consistent, controllable latent state under intervention, or do the two metrics decouple?

Why

A model can render a plausible future frame without a consistent, controllable state. If teams accept generative video quality as evidence of world modeling, they will ship systems that look right and fail the counterfactual and intervention-based tasks central to planning.

World models aim to predict physical reality, not just the next word in a sentence.
Open source video
Sora video generation.
Open source video
Source video

Sim-to-real transfer from prompt-generated 3D environments

Research question

Do policies trained inside prompt-generated explorable 3D worlds transfer to hand-authored simulators or physical tasks at a success rate comparable to policies trained natively in those environments?

Why

If prompt-minted environments are usable training substrates, environment supply stops being the bottleneck for embodied agent training and simulation-based evaluation, making environments cheap, versioned, and disposable artifacts.

World Labs' Marble platform allows users to generate and explore 3D worlds from text or visual prompts.
Open source video
AI can go beyond text generation into interactive 3D spatial environments for gaming, film, and robotics.
Open source video
Source video

Capital and headcount as substitutable inputs in frontier research bets

Research question

For a single narrow scientific thesis, how much capability per dollar does a ~50-person, capital-concentrated lab achieve relative to a broad generalist lab, and which resource (data, compute, or talent) is actually binding?

Why

Concrete counter-example to the assumption that frontier AI requires massive engineering orgs; directly informs whether to staff a research direction broadly or keep it narrow and dense, and where capital should be allocated.

World Labs raised $1 billion while maintaining a small, highly collaborative team of ~50 people.
Open source video
Small teams can drive foundational paradigm shifts in AI if focused on the right core scientific challenges.
Open source video
Source video

Article ideas

4

Fidelity Is Not World Modeling: Why Generative Video Is a Weak Proxy for Usable Internal State

Photorealistic future-frame generation is routinely accepted as evidence of world modeling, but visual plausibility is uncorrelated with controllability and intervention consistency; teams that adopt fidelity as their acceptance metric will ship systems that fail exactly the planning tasks the spatial layer was supposed to unlock.

Angle

Take the strongest pro-generative-video claim at face value, then specify the failure modes (counterfactual queries, intervention response, novel viewpoints) and the evaluation harnesses that would expose them.

Source video

The Dataset Comes First, Again: Sequencing Investment for Modalities That Have No Canonical Corpus

The ImageNet ordering — corpus and benchmark before architecture — is the correct sequencing for 3D scenes and physical interaction traces, so budget and headcount should shift toward data pipelines, labeling taxonomies, and evaluation suites before model work is staffed.

Angle

Argue against the prestige ordering that treats architecture as the primary research bottleneck, and translate the corpus-first principle into concrete team-shape and budget consequences.

Source video

Decompose the Embodied Stack: Rendering, Simulation, and Planning as Separable Subsystems

Treating spatial intelligence as one monolithic 'understand 3D' capability produces unattributable failures; separating generation, forward prediction, and decision layers with distinct interfaces makes failures localizable, vendors substitutable, and latency budgets assignable per function.

Angle

Convert a three-word capability list into an architectural contract: what each subsystem owns, what crosses its boundary, and which subsystem can be replaced by a non-neural component.

Source video

Fifty People and a Billion Dollars: When a Narrow Scientific Thesis Substitutes for Organizational Scale

Capital concentration on one scientific bet can substitute for headcount scaling up to a point, but concentrating a frontier bet in a small team with no redundancy moves the risk from coordination overhead to single-thesis failure, and the article identifies which resources are actually binding before either model should be copied.

Angle

Steelman both the small-team leverage claim and the giant-competitor risk, then specify the conditions under which the concentrated model is a defensible strategy rather than survivorship bias.

Source video

Project ideas

4

DecouplingBench: Fidelity vs. Intervention Consistency

beyond-evals

Rank correlation between visual-fidelity scores and intervention-consistency scores on generated 3D environments is low or negative; fidelity-maximizing systems do not maintain controllable state.

Proof of concept

Generate a fixed set of explorable environments from the same prompts, score each on visual fidelity, then apply scripted interventions (move an object, query a counterfactual, re-enter from a novel viewpoint) and measure state divergence from the expected consequence.

Measurement

Spearman correlation between fidelity rank and intervention-consistency rank across >=30 environments, plus per-prompt divergence magnitudes.

Source video

Hybrid Spatial Stack Ablation

new

A stack using learned rendering plus an explicit physics engine for simulation plus a learned planner outperforms an end-to-end learned world model on counterfactual tasks at lower compute cost.

Proof of concept

Implement three variants of the same task (end-to-end learned, learned rendering + physics simulation, fully learned) behind one interface, then run an identical counterfactual evaluation suite against each.

Measurement

Counterfactual task success rate and compute cost per solved task for each variant; report the pairwise deltas.

Source video

Prompt-to-Environment Transfer Harness

movement-lab

Policies trained inside prompt-generated explorable 3D worlds transfer to hand-authored simulators with a success-rate delta below a fixed threshold versus policies trained natively in those simulators.

Proof of concept

Mint environments from text/image prompts, train a navigation/manipulation policy in them, then evaluate the same policy in matched hand-authored simulators and compare against a native-trained control policy.

Measurement

Transfer success-rate delta between generated-environment-trained and natively-trained policies, plus environment-generation cost per training episode.

Source video

Deterministic Environment Registry

gatehouse

Nondeterminism in generated environments produces evaluation variance larger than typical policy-improvement deltas, and seeded, versioned environment snapshots reduce that variance below the policy delta threshold.

Proof of concept

Run the same evaluation suite repeatedly against unversioned prompt-generated environments, record score spread, then re-run against frozen, seeded, registry-versioned snapshots of the same environments.

Measurement

Evaluation score variance before versus after versioning, compared against the median measured policy-improvement delta.

Source video

Architectural implications

3

A capability stack organized around text prediction is being reframed around a spatial/world-model layer with distinct sub-functions.

Before

Capability stack centered on a language model plus text context, tools, and retrieval.

After

An additional world-model layer owning rendering, simulation, and planning, sitting beneath or beside the language layer for embodied tasks.

Consequence

New interface boundaries and separate runtimes/state stores for spatial state; language models become planners/callers rather than the whole system.

Source video

Dataset and benchmark construction is treated as the enabling artifact rather than a downstream convenience.

Before

Model architecture as the primary research bottleneck and the object of prestige.

After

Corpus and benchmark construction as first-class infrastructure that gates algorithmic progress.

Consequence

Budget and headcount shift toward data pipelines, labeling taxonomies, and evaluation suites; model work is sequenced after corpus availability.

Source video

Generative output becomes a runtime environment other systems act inside.

Before

Generated media treated as an endpoint artifact consumed by humans.

After

Generated 3D worlds treated as instantiated environments for training, testing, and agent execution.

Consequence

Need for versioning, reproducibility, and determinism of generated environments if they are used as evaluation or training substrates.

Source video

Tradeoffs and failure modes

3

Generative video fidelity vs. interactive physical consistency

Benefit

Video generation (e.g., Sora) provides a visually convincing demonstration of predicted futures and requires no explicit physics engine.

Cost or risk

Visual plausibility does not guarantee a consistent controllable state, so systems may look correct while failing intervention, counterfactual, and planning tasks that robotics requires.

World models aim to predict physical reality, not just the next word in a sentence.
Open source video
Source video

Capital concentration with a small team

Benefit

~50 people with $1B can stay focused on one scientific thesis and avoid coordination overhead.

Cost or risk

Competing against deep-pocketed tech giants on a resource-intensive frontier while keeping the team small concentrates risk in a single thesis with no redundancy.

World Labs competes against deep-pocketed tech giants by focusing on spatial intelligence
Open source video
Source video

Policy framing of AI risk

Benefit

Grounding regulation in demonstrated science rather than speculative scenarios avoids rules aimed at fictional capabilities.

Cost or risk

A science-only framing can lag behind capabilities that have no deployed evidence yet, leaving governance reactive rather than anticipatory.

Regulation should be rooted in science, not science fiction.
Open source video
Source video

Open questions

4

How quickly will world models integrate with robotics for physical task execution?

Why unresolved

The summary asserts the frontier and lists robotics as an application surface but offers no timeline, integration mechanism, or measured transfer from generated worlds to physical tasks.

Research direction

Build a sim-to-real harness using generated 3D environments and measure policy transfer success rate against hand-authored simulators.

Source video

How will startups compete with tech giants in the resource-intensive world model race?

Why unresolved

The summary cites $1B raised against ~50 people but does not specify whether the advantage is data, compute, talent, or a narrower scientific scope.

Research direction

Compare cost-per-capability-axis for focused world-model efforts versus broad generalist labs; identify which resources are actually binding.

Source video

If rendering, simulation, and planning are separable, which functions can be non-neural or externally constrained?

Why unresolved

The summary names the three components but does not state whether any is implemented by a physics engine, a learned model, or a hybrid.

Research direction

Prototype hybrid stacks where rendering is learned but simulation is a physics engine, and ablate failure modes on counterfactual tasks.

Source video

What does a benchmark for spatial intelligence measure, and can it be built the way ImageNet was?

Why unresolved

The summary documents ImageNet's scale and category structure but provides no analogous dataset or metric structure for spatial intelligence.

Research direction

Define evaluation suites probing state consistency, intervention response, and physical plausibility under novel viewpoints rather than pixel fidelity.

Source video

Key claims

7
predictionVerification needed

Spatial intelligence is the next major frontier in AI beyond text-based LLMs.

Evidence

Spatial intelligence is the next major frontier in AI beyond text-based LLMs.

Question

What measurable capability does spatial intelligence provide that current multimodal LLMs cannot, and is there a benchmark that separates the two?

Source video
causalVerification needed

World models integrate rendering, simulation, and planning to navigate physical reality.

Evidence

World models integrate rendering, simulation, and planning to navigate physical reality.

Question

Are all three components required for useful world models, or can subsets suffice for specific embodied tasks?

Source video
factualVerification needed

The 2012 ImageNet competition used 14 million pictures across more than 21,000 categories.

Evidence

14 million pictures categorized across over 21,000 categories used in the 2012 ImageNet competition.

Question

What were the exact image and category counts of ImageNet at the time of the 2012 competition?

Source video
causalVerification needed

Data scale combined with neural networks and GPUs is the recipe for modern AI breakthroughs.

Evidence

Proved that data scale combined with neural networks and GPUs is the recipe for modern AI breakthroughs.

Question

Which of the three ingredients (data scale, architecture, compute) was the binding constraint, and does the ordering generalize to non-vision modalities?

Source video
factualVerification needed

World Labs' Marble platform generates explorable 3D worlds from text or visual prompts.

Evidence

World Labs' Marble platform allows users to generate and explore 3D worlds from text or visual prompts.

Question

What are Marble's consistency, resolution, and controllability limits, and can generated worlds be used as training environments?

Source video
factualVerification needed

World Labs raised $1 billion while maintaining a team of roughly 50 people.

Evidence

World Labs raised $1 billion while maintaining a small, highly collaborative team of ~50 people.

Question

What is the verified funding total and headcount, and how is the capital allocated between compute and personnel?

Source video
causalVerification needed

Small teams can drive foundational paradigm shifts in AI if focused on the right core scientific challenges.

Evidence

Small teams can drive foundational paradigm shifts in AI if focused on the right core scientific challenges.

Question

What historical counter-examples exist, and how often has a small team produced a paradigm shift versus incremental model improvement?

Source video

Connections

4