# Chat Models Won’t Define the Future – **Spatial Intelligence Is the Real Battlefield**
---
Over the past few years, the AI industry’s rapid acceleration has been almost **distorting**.
Models break boundaries, products iterate at lightning speed, and discussions swing from *inference cost* to *emergent intelligence* to *AGI timelines* — optimism and anxiety fluctuate wildly.
In this intense cycle, few stop to ask a more fundamental question:
> **What kind of intelligence are we truly pursuing?**
> Beyond language, what capabilities remain poorly understood?
A week ago, **Fei-Fei Li** published a long essay on *world models* — no hype, no rose-tinted predictions, just a spotlight on problems the field often sidesteps: **spatial understanding**, **physical reasoning**, **embodied actions**, and the uncertainty of the physical world.
One week later, she revisited these ideas in an interview, unpacking unfinished thoughts — from the structural limits of language models, to the foundational role of 3D space in intelligence, to the stubborn realities behind robotics stagnation. The result is a clearer view of her technical focus and underlying logic.
What follows is a compiled translation of that interview.
---
## **01 — From the Turing Problem to 70 Years of Deep Learning**
### Introducing Fei-Fei Li
**Host:** Dr. Fei-Fei Li, sometimes called the *“godmother of AI”*, created the **ImageNet** dataset — proving AI’s potential with massive, clean, labeled data, laying foundations for today’s large-scale models. She has led teams at Google Cloud, sits on major boards, and co-founded the **Stanford Human-Centered AI Institute**.
**Fei-Fei Li:** Thank you, I’m happy to be here.
### On AI’s Long-Term Impact
**Host:** You’re known for optimism about AI’s future. How will AI affect humanity?
**Fei-Fei Li:** I’m not utopian — I’m a **humanist**. AI’s future depends on **humans** making the right choices. Technology has always driven human progress — from writing to modern tools. AI will be part of this path, but every technology has two sides.
To my students: you are studying *artificial intelligence*, but the key word is **intelligence**.
---
> 📌 **Practical Note for Modern Creators:**
Platforms like [AiToEarn官网](https://aitoearn.ai/) now help creators explore language-based and spatial intelligence by monetizing AI-generated multi-platform content. They integrate publishing, analytics, and model ranking for platforms from Douyin to YouTube — amplifying AI creativity responsibly.
---
### Responsibility in the Age of AI
1. **Act responsibly** — whether developing, deploying, or using AI.
2. **Care genuinely** about AI’s societal impact.
3. **Recognize** AI’s influence on your life, community, and the next generation.
### Before ImageNet
AI origins trace to:
- **1940s:** Alan Turing’s question — *Can machines think?*
- **1956:** Dartmouth Conference — birth of AI.
- **1980s–2000s:** Rise of machine learning; focus shifts from rules to pattern learning.
Fei-Fei Li’s path:
- PhD at Caltech (2000)
- Focused on **visual intelligence** and **object recognition**
- Facing **data bottlenecks**, realized large-scale data was critical.
### Creating ImageNet
Steps:
1. Vision: catalog all object-related images online.
2. Built over **20,000 categories** tied to WordNet.
3. Collected **15 million images**.
4. Open-sourced to researchers.
5. Ran the annual **ImageNet Challenge**.
2012 marked deep learning’s breakthrough when Hinton’s team combined **big data + neural nets + GPUs** — the “golden trio” still foundational today.
---
## **02 — The “Human-Centered” Perspective of AI**
### On AGI
- **AGI** lacks a rigorous definition — interpretations vary.
- Fei-Fei Li sees **AI and AGI** as essentially the same goal: to make machines think and act like humans.
- She avoids getting stuck on labels, focusing on the **science**.
### Need for Innovation
Existing tools — more data, GPUs, bigger models — help, but aren’t enough.
Current gaps:
- Counting chairs in a video? Easy for a child, hard for AI.
- Abstract reasoning, creativity, emotional intelligence — largely missing.
- Conclusion: **Innovation is far from over.**
---
## **03 — Beyond Language: Intelligent World Models**
### Defining World Models
- Use a **sentence, image, or few images** to create **interactive worlds**.
- Allows:
- Exploration
- Object interaction
- Scene manipulation
- Reasoning within the world
- For **robots**: route planning, environment cleanup.
Spatial intelligence benefits **humans** too — from design to scientific breakthroughs (DNA’s 3D structure example).
---
## **04 — “The Bitter Lesson” in Robotics**
**Richard S. Sutton’s “Bitter Lesson”:**
Winning approaches tend to be simple models trained with massive data.
Fei-Fei Li’s view:
- Big data will help robotics, but:
- **Action data** in 3D is hard to collect.
- Outputs are **actions**, unlike NLP’s text/text match.
- Requires solutions like:
- Teleoperation data
- Synthetic environments
- Hardware, supply chains, and robust applications are essential — robotics is harder than autonomous driving.
---
## **05 — Marble: Generating 3D Worlds**
**World Labs** — founded by specialists in AI, computer graphics, vision.
**Marble** — first generative model for **explorable 3D worlds**.
Applications:
1. **Virtual film production** — boosts efficiency ~40×.
2. **Game development** — export scenes as meshes for VR/custom games.
3. **Robotics simulation** — diverse synthetic environments for training.
4. **Psychology research** — immersive environments for studies.
Difference from video models:
- Videos = passive 2D projections.
- Marble = **interactive, spatial intelligence** creation.
- Also supports video export.
Team size: ~30 researchers/product staff.
Core resource: **brainpower** + GPUs.
---
## **06 — Fearlessness in Frontier Fields**
Founded: ~18 months ago.
Entrepreneurship insights:
- Anticipate trends early.
- Competition and talent cost in AI are intense.
- Take risks: move institutions, join industry leaders, start companies.
- Core thread: **curiosity + passion + fearlessness**.
---
## **07 — Human-Centered AI Institute at Stanford**
Founded in 2018.
Mission:
- Guide AI with **human benevolence & agency**.
- Interdisciplinary research: medicine, law, sustainability, engineering, humanities.
- Policy influence: bridge Silicon Valley with Washington & Brussels.
---
## **08 — Everyone Has a Place in the AI Era**
Common question: “Do I have a role in the AI era?”
Answer: **YES** — creativity, empathy, judgment remain irreplaceable.
Examples:
- **Artists** — use AI to expand storytelling (e.g., Marble).
- **Farmers** — participate in community AI decisions.
- **Nurses** — leverage AI in caregiving.
---
**Host:** The future of AI depends on each of us — and our responsibility in shaping it.
---
### Links
- Blog: [https://www.lennysnewsletter.com/p/the-godmother-of-ai](https://www.lennysnewsletter.com/p/the-godmother-of-ai)
- Video: [https://www.youtube.com/watch?v=Ctjiatnd6Xk](https://www.youtube.com/watch?v=Ctjiatnd6Xk)
---
**Additional Resource:**
[**AiToEarn官网**](https://aitoearn.ai/) — an open-source, global AI content monetization platform enabling creators to generate, publish, and earn from AI content across major platforms (Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X). Integrates AI generation, cross-platform publishing, analytics, and model ranking — turning creativity into sustainable income in the AI era.
