Core Summary
Demis Hassabis (co-founder of DeepMind, the company behind AlphaGo and AlphaFold) shares his journey from AI engineer to scientist, his belief that AI is humanity’s ultimate tool for solving big scientific problems, and his vision for responsible AI use. He discusses what AGI (General Artificial Intelligence) really means (we’re not there yet), how AI is transforming fields like biology and physics, and his philosophical take that information might be the fundamental building block of the universe. His goal? Use AI to cure diseases, fight climate change, and unlock the secrets of reality—all while keeping safety and ethics front and center.
1. From Game AI to Life-Saving Science: How AlphaGo & AlphaFold Changed the Game
Hassabis started by building AI for games (like AlphaGo, which beat world champion Lee Sedol in 2016). But he quickly realized AI could do more than win games—it could solve real-world problems.
Take AlphaFold: For 50 years, biologists struggled to predict how proteins (long strings of amino acids) fold into 3D shapes. The shape of a protein determines its function (e.g., fighting viruses or digesting food). Brute-forcing all possible shapes was impossible (a protein has 10^300 possible folds—way more than the number of atoms in the universe!).
AlphaFold didn’t brute-force. It learned patterns from 150,000 known protein structures (collected over 40 years by scientists). Now it can predict a protein’s shape in seconds with near-perfect accuracy. This is a game-changer for drug discovery: scientists can now design drugs that fit exactly into a protein’s shape to treat diseases like cancer or malaria.
The key insight here? AI excels at finding patterns in huge, complex systems—something humans can’t do alone.
2. What Is AGI, and Why Aren’t We There Yet?
AGI is a system that has all the cognitive abilities of a human. Think: it can invent a new scientific theory (like Einstein’s relativity), write a poem, learn to play a new game from scratch, or solve a math problem it’s never seen before.
Right now, AI is “narrow”—it’s great at one thing (like folding proteins or playing Go) but terrible at others (e.g., a top AI math solver might fail a simple arithmetic question if phrased weirdly). Here’s why we’re not at AGI:
- No real creativity: AI can solve existing problems but can’t come up with new, deep questions (like the Riemann Hypothesis, a famous unsolved math problem).
- Inconsistency: It makes silly mistakes humans never would (e.g., mixing up facts or failing basic logic).
- Lack of adaptability: It can’t learn a new skill quickly (unlike a human who can pick up a new game in minutes).
Hassabis says AGI would need to do things like: Given 1910s knowledge, invent general relativity by 1915 (just like Einstein did). We’re far from that.
3. AI as the Ultimate Tool for Big Scientific Problems
Hassabis picks scientific problems for AI based on three rules:
1. Huge search space: The problem has way too many possible solutions (like all possible protein shapes or climate patterns).
2. Enough data: There’s data (or simulators to generate data) to learn patterns from.
3. Clear goal: There’s a way to measure success (e.g., “did the protein fold correctly?” or “did the AI predict the weather accurately?”).
Examples of problems AI is tackling:
- Climate change: AI models can predict extreme weather events or optimize energy grids to reduce carbon emissions.
- Physics: AI helps study quantum gravity (how gravity works at the smallest scales) by processing massive amounts of data from particle accelerators.
- Drug discovery: Beyond AlphaFold, AI is used to find new molecules that target diseases like Alzheimer’s.
The big idea: AI is a supercharged assistant for scientists—it can process data faster and find patterns humans miss, letting us solve problems we’ve been stuck on for decades.
4. The Philosophical Twist: Is Information the Building Block of Reality?
Hassabis has a bold idea: Information (not matter or energy) might be the most fundamental thing in the universe.
What does that mean? Everything in nature—from cells to mountains to stars—has stable patterns. These patterns can be learned by AI (like AlphaFold learned protein shapes). If these patterns are based on information, then understanding information could help us unlock the secrets of reality (like what time is, or how gravity and quantum mechanics fit together).
He uses AI to test this: If AI can simulate complex natural systems (like a cell), that supports the idea that information is the foundation of everything. It’s a big, abstract thought—but AI is giving us a way to explore it.
5. Responsible AI: Using It for Good (and Safely)
Hassabis is obsessed with using AI responsibly. Here’s what that means:
- Follow the scientific method: Test AI rigorously, don’t overpromise, and be honest about its limits.
- Think about long-term impacts: Even if we can’t predict everything, we need to build safety guards (e.g., making sure AI doesn’t harm humans).
- Focus on global good: Use AI to help the poorest people (like curing diseases that affect developing countries) and solve urgent problems (climate change).
In the short term, AI is a tool for scientists. In the long term, it might become a partner—but only if we build it carefully.
Hassabis hopes that by 2050, AI will have helped us cure diseases, fix climate change, and enter a new golden age of scientific discovery—all while being safe and beneficial for everyone.