1. Introduction: The Conversation on Creativity 🎨
The video opens with a discussion on creativity. The host emphasizes that "models need to be creative and learn how to invent things," identifying creativity as a core requirement for AGI (Artificial General Intelligence).
"I think models need to be creative and learn how to invent things. That's precisely the condition for AGI."
2. AlphaGo and the Meaning of 'Move 37' ♟️
Hassabis brings up rewatching the AlphaGo documentary, and mentions the moment the algorithm played a creative move — 'Move 37'.
"The algorithm played a creative move. That was Move 37."
He emphasizes this as a historic moment in which AI invented a new strategy no human had ever conceived.
"What's really interesting is that the algorithm was already creative several years ago."
3. The Absence of Creativity in LLMs 🤖
The host points out that we have yet to see this kind of creativity in LLMs (Large Language Models).
"These tools are genuinely impressive, but the biggest disappointment is that they're ultimately bounded by their training data. They can combine things they already know, but they can't create something truly new."
Hassabis explains why AlphaGo's creativity was special, and why LLMs have not yet reached that level.
4. Three Levels of Creativity 🧠
Hassabis breaks creativity down into three levels.
1) Ordinary Creativity: Interpolation
- Averaging over things already seen.
- Example: "If you ask a model that has seen many cat photos to generate a new cat photo, it produces an average of what it has seen."
- Hard to call this genuine creativity — it's the lowest level.
"Technically it's a new cat, but in practice it's just an average of what it's seen. That's not very creative."
2) AlphaGo's Creativity: Extrapolation
- Creating new strategies no human has ever tried.
- 'Move 37' is the prime example.
- A move that completely reinvented Go as humans had played it for thousands of years.
"What AlphaGo demonstrated was extrapolation. It created a new strategy that no human had ever tried before. That was Move 37."
- A capability that could also be enormously useful for advancing science.
3) Human Creativity: Invention
- The ability to create an entirely new game or concept.
- Example: "Inventing a beautiful, elegant game whose rules you can learn in five minutes but that takes a lifetime to master."
"If you asked a system to create a game as beautiful as Go — one whose rules you can learn in five minutes but that takes a lifetime to master — today's AI cannot do that."
- The objective function is too abstract and vague for current AI to handle.
5. Why LLMs Don't Produce a 'Move 37' 🤔
Hassabis explains why LLMs can't produce creative results the way AlphaGo did.
"The reason people are disappointed is that you don't see anything like 'Move 37' in today's LLMs."
Structural Differences Between AlphaGo and LLMs
- AlphaGo consists of two parts: a 'model' and a 'search'.
- The model alone can reach master- or grandmaster-level play, but cannot produce a world champion or innovative moves.
- Search is what allows the model to venture beyond what it knows and explore new territory.
"If you run only the model in AlphaGo or AlphaZero without search, it just plays the most plausible move based on pattern matching. That gets you to master or grandmaster level, but not world champion, and not a breakthrough move like Move 37."
- Search is precisely what lets the model move beyond its training data into a new tree of knowledge.
"Search is what allows the model to venture into a new tree of knowledge beyond what it knows — and that's why novel ideas like 'Move 37' emerge."
LLMs Have No Search
- Current LLMs have no search, so they cannot go beyond combinations of training data.
- For genuine creativity, LLMs also need search or agent-based systems.
"In language models, you'd need to search new parts or compositions of the world model, and that's far more complex — which is why we haven't seen it yet."
6. Looking Ahead: The Promise of Agent-Based Systems 🚀
Hassabis anticipates that once agent-based systems arrive, LLMs too will be capable of creative results like 'Move 37'.
"I think the agent-based systems coming soon will be able to do something like 'Move 37.'"
7. Closing: Are We Expecting Too Much from AI? 🤷♂️
Finally, the host closes by raising the question of whether we are setting our expectations for AI too high.
Key Concepts Summary
- Creativity
- Interpolation
- Extrapolation
- Invention
- AlphaGo
- Move 37
- Search
- Agent-based systems
- LLM (Large Language Model)
- Objective function
- Training data
Notable Quotes
"The algorithm played a creative move. That was Move 37."
"These tools are genuinely impressive, but the biggest disappointment is that they're ultimately bounded by their training data."
"Technically it's a new cat, but in practice it's just an average of what it's seen. That's not very creative."
"What AlphaGo demonstrated was extrapolation. It created a new strategy that no human had ever tried before. That was Move 37."
"If you asked a system to create a game as beautiful as Go — one whose rules you can learn in five minutes but that takes a lifetime to master — today's AI cannot do that."
"If you run only the model in AlphaGo or AlphaZero without search, it just plays the most plausible move based on pattern matching. That gets you to master or grandmaster level, but not world champion, and not a breakthrough move like Move 37."
"Search is what allows the model to venture into a new tree of knowledge beyond what it knows — and that's why novel ideas like 'Move 37' emerge."
"I think the agent-based systems coming soon will be able to do something like 'Move 37.'"
This video offers a deep and illuminating look at the nature of AI creativity and the possibilities ahead. In particular, it clearly explains why 'Move 37' matters so much in the history of AI — and what LLMs would need in order to reach that level. 😊