From Words to Worlds

Why AI Needs Spatial Intelligence to Move Beyond Language

AI excels at talking, but struggles to truly understand the world.

Recently, Google unveiled Gemini 3 Pro, sparking a wave of online buzz. The questions came quickly: Does it have more parameters? Can it handle longer context windows? Are we getting closer to AGI (Artificial General Intelligence)?

Renowned computer scientist Fei-Fei Li — U.S. National Academy of Engineering member and Stanford professor — offers a reality check.

On November 10, she published a detailed article arguing:

> Bigger models and better algorithms aren’t enough. Without world understanding, AI will never reach true intelligence.

---

The Problem with Today’s Large Language Models

Like Well-Read Scholars Who’ve Never Gone Outside

Think of ChatGPT, Gemini, DeepSeek, or Doubao — all powered by Large Language Models (LLMs).

LLMs predict the next word in a sequence.

Example: You say “床前明月光”; the model guesses the next phrase is “疑是地上霜.”

This word-prediction ability, trained on massive text datasets, enables LLMs to:

• Pass professional exams
• Solve complex math problems

But they falter at simple, physical reasoning:

• “How far is that car from the tree?”
• “Will this box fit in the trunk?”

Sometimes, they make absurd predictions — e.g., assuming a cup will float upward.

LLMs may know formulas, but they lack physical common sense. Fei-Fei Li calls them “wordsmiths in the dark.”

---

Why Hallucinations Happen

LLMs rely on statistical patterns in text, not real-world experience.

They could claim “The sun rises from the west” because grammar and probability suggest it — even if physics says it’s impossible.

They’ve read thousands of books, but never touched the world outside.

---

Language Can Fabricate — The Physical World Doesn’t Lie

Enter Spatial Intelligence

Fei-Fei Li believes AI must develop spatial intelligence — the ability to understand and interact with the physical world.

Example: Drinking coffee

• Vision: Judging cup-to-mouth distance
• Motor control: Adjusting grip based on weight
• Touch: Avoiding burns
• Balance & motion: Keeping cup level

This process uses perception, imagination, and action — not verbal commands.

Spatial intelligence is key because true intelligence requires:

• Prediction
• Action
• Goal achievement — in changing, uncertain environments.

---

Learning Through Interaction

• Babies: Push over blocks → hear crash → learn cause & effect.
• Scientists: Watson and Crick built physical DNA models to reveal the double helix — insight wasn’t in the “words” but in spatial arrangement.

---

From Predicting Words to Predicting Frames

Fei-Fei Li advocates shifting AI from predicting the next word → to predicting the next frame of reality.

Example: Letting go of a glass cup

• Human mind predicts: fall → impact → shatter
• Without reading about it, you know what happens.

Key Difference:

• Word prediction = grammatical logic
• Frame prediction = physical logic

This requires world models — spaces with consistent gravity, light, occlusion, and physics.

---

Challenges in Building World Models

Fei-Fei Li identifies two key obstacles:

• Finding the formula: LLMs succeed with “predict next word” simplicity; can we find an equally elegant equation for spatial intelligence?
• Finding the data: Requires massive spatial datasets; extracting 3D info from 2D video is an active research area.

---

Marble: A Glimpse of Spatial AI

Fei-Fei Li’s World Labs developed Marble — input text or a photo, get an explorable 3D space.

Testing it: Uploading a single image → Marble inferred chairs, desks, and room layout (still rough, but promising).

---

Potential Impact of Spatial Intelligence

• Robots at Home:
• Avoid fragile vases
• Dry wet floors before walking
• Assist elderly care
• Industrial Applications:
• Controllable video generation for ads & film
• Virtual production efficiency (Sony partner reported 40× gains with Marble)
• Consumer Products:
• Interactive interior design
• 3D memory albums
• VR therapy for phobias
• Synthetic Data Markets:
• “Textbooks” for robots: domain-specific task data

---

AI Monetization in the Spatial Era

As spatial AI matures, tools that combine creation + cross-platform publishing will become essential.

Example: AiToEarn官网 offers:

• AI content generation
• Distribution across platforms like Douyin, Bilibili, Xiaohongshu, Instagram, YouTube, LinkedIn
• Analytics and AI model ranking (AI模型排名)

This could help spatial AI creators share interactive environments instantly and monetize their innovations globally.

---

Final Takeaways

Why AI still makes simple mistakes:

It “thinks” in statistical patterns, not through cause-and-effect in reality.

Fei-Fei Li’s proposal:

Shift from predicting text → to predicting reality via spatial intelligence & world models.

Possible outcomes:

• Household robots with genuine awareness
• AI scientists finding new laws of nature
• Fully controllable, physically consistent creative tools

Current status:

• Marble is early-stage
• Formula for world models unknown
• Spatial datasets scarce

But now, the path to true intelligence looks clearer.

---

References

• From Words to Worlds: Spatial Intelligence is AI’s Next Frontier
• Google Developer Guide: Introduction to Large Language Models

---

If you'd like, I can also create a concise 10–15 bullet executive summary of this rewritten piece so it’s easier for leaders or investors to digest in under 2 minutes. Would you like me to prepare that?