Reporting by XinZhiyuan
Edit: David
DeepMind's chief research scientist David Silver's star temperament contrasts sharply with his quiet, humble personality. The reinforcement learning bull who lets AI teach themselves to play Go games believes that the consequences of human intelligence may be more terrible than the threat of AI.
According to David Silver, DeepMind's principal research scientist and professor of computer science at University College London, games are key to sparking creativity, especially for AI.
Silver competed in the National Scrabble Tournament as a child, after which he continued to study at Cambridge and co-founded a video game company.
Later, after earning his Ph.D. in artificial intelligence, he led the DeepMind team to develop AlphaGo, the first AI program to beat a world champion in a Go game.
For Silver, games are experimental grounds for understanding how humans and artificial brains learn on their own to achieve goals.
For Go AI AlphaGo, the learning of providing programs with decisions about humans in different positions and then having the AI imitate them is known as "supervised learning." Implementing AI programs learns through self-play is called "reinforcement learning."
Then, in the crucial game of AlphaGo and the world championship, Silver suddenly thought: Perhaps, the decisions of machines should not be influenced by humans at all.
The idea eventually became the beginning of AlphaGo Zero, the successor to AlphaGo, who, without accepting human knowledge at all, relied solely on reinforcement learning to learn how to play Go well from the rules of the game.
Later, AlphaGo Zero played 100 games with AlphaGo, sealing his "Big Brother".
In April 2020, David Silver won the 2019 ACM Computing Award for "Groundbreaking Advances in Computer Games".
At the ceremony, then-ACM President Cherri M. Pancake said:
"Few other researchers have produced so many exciting results in the field of AI as David Silver. His insights into deep reinforcement learning have spread from games to a variety of applications, including improving grid efficiency, reducing power consumption in data centers, and planning the trajectory of space probes for the European Space Agency."
Silver is also the Royal Society Fellow and the first scientist in the field of AI to win the Mensa Foundation Award for Best Scientific Discovery.
Silver's stardom contrasts with his quiet, humble nature. In this interview, Silver talks about gaming, the meaning of creativity, and the potential of AI to avoid climate change, pathogen infections, mass poverty, and environmental disasters.
Here's what the interviews are:
Did you play games differently from other children when you were a kid?
I was in the National School Scrabble Competition at the time, and I remember some interesting moments.
Once, at the end of the final match, I asked my opponent, "Are you sure you want to spell that word?" Why not spell a higher-scoring word?" He changed his spelling and won the race and the championship, which makes me very happy.
I'm more fascinated by the meaning of a better game than winning a game.
How did you turn your love of gaming into a real job?
Later, I was introduced to chess and met Demis Hassabis (co-founder of DeepMind). At that time, he was the strongest young chess player of his age in the world.
When he didn't have enough pocket money, he would come to my town, compete, win £50 and go home. Later, we met in Cambridge and together we founded a game company, Elixir, and now we're back together at DeepMind.
What aspects of problem solving has this fascination with games taught you?
On the one hand, we believe we have a special ability to be called "creativity", which AI algorithms do not have. This is actually a fallacy.
Now, there are hints of creativity in AI. In the second game of AlphaGo against Lee Sedol in 2016, AlphaGo played the 37th hand in the black, and the Go community thought it was full of creativity, even beyond the understanding of human professional chess players at that time. This shows that AI has indeed tried something new and extraordinary.
But can humans apply this broad creativity to anything, not just in games?
The whole process of trial and error learning, such as trying to think of a way by itself, or letting the AI think of its own way, how to solve the problem, is a creative process. AI didn't know anything from the start. Then you'll discover something new, a creative leap, a new model, or a new idea that helps to achieve your goals better than before.
Now you've got new ways to play games, solve puzzles, or interact with people. This process is actually the accumulation of thousands of small discoveries one after another. This is the essence of "creativity".
If AI algorithms are not creative, they will be in trouble. AI needs to be able to experiment with new ideas for itself — ideas that humans don't tell them. This should be the direction of future research to continue to push the system of "new ideas" through self-experimentation.
Many people believe that computers can only play Go at the level of human amateurs. Have you ever doubted your ability to improve?
When I arrived in South Korea for the 2016 AlphaGo competition, I saw rows of cameras, and I heard that there were more than 200 million people watching online, and I thought, "Wait, can this really work?"
Rival Lee Se-chi is a talented world champion who will try everything to bring AI programs into strange situations that wouldn't normally happen.
I feel fortunate that we withstood that test. I later asked myself, "Can you take a step back and get back to the basics and understand what it means for a system to really learn for itself?" To find something purer, we gave up human knowledge in AI models and developed AlphaZero.
For thousands of years, humans have developed well-known strategies for Go. What did you think when AlphaZero quickly discovered and rejected these favorable strategies?
We set the original AlphaGo in the wrong position (compared to the human pro). We believe that if we can find a new version that allows the AI to find the right place on its own, it will be successful. At first, we made great progress, but then it didn't seem to work. We don't think the AI has found the right 20 to 30 lot positions on its own.
Fan Li, a pro we worked with, spent hours studying these moves. In the end, he said that it was the pros who were wrong, and AlphaZero was right. The AI found a solution and re-evaluated the chess pieces that were initially classified as "wrong." I realized that we have the ability to overturn what humanity thinks of as standard knowledge.
After that, you're in charge of developing AlphaStar and letting the AI play StarCraft 2. Why jump from Go to video games?
Go is a relatively small field. To start with Go, expanding into the human brain's range of abilities requires a lot of steps. We try a lot of areas of more complex dimensions, those that humans do well, but AI doesn't do well.
Going from AlphaGo to AlphaStar is actually a natural development. Like humans, AI systems can only see a portion of the map. It's not like playing Go or chess, where you can see your opponent and all your pieces. The game can only see information near the control object, and you have to scout to get the information. This is closer to what happens in the real world.
What is the end goal?
I think the capabilities of AI agents are as broad as the human brain. Although we don't yet know how to fully realize the function of the brain, there is evidence in the human brain.
Is the human brain completely replicated? Do you really think this is realistic?
I don't believe in magical, mysterious explanations of the brain.
In a way, the human brain is an algorithm that accepts input and produces output in a powerful and universal way. Our ability to understand and build AI is limited, but this understanding is growing rapidly. Today, we have AI that can crack narrow areas like Go, as well as models that can understand and produce natural language.
So, do you think there is no upper limit to the ability of human AI?
Now we're just starting to hit the road. Imagine what we would be like if we had gone through another 4 billion years of evolution. Maybe we'll have more complex intelligence and can do better. AI is a bit like this, there are no limits to this process because the world is infinitely complex in nature.
So, will there be a ceiling? At some point, physical limits do exist, so they are not without limits. Eventually, you'll use up all the energy and all the atoms in the universe to build a computing device. But relative to the present, it can actually be regarded as infinite. Beyond human intelligence, this scope is very broad.
Stephen Hawking worries that machine intelligence will have dire unintended consequences. Do you have similar concerns?
I'm more worried about the unintended consequences of human intelligence than this, such as climate change, pathogens, mass poverty, and environmental disasters.
The pursuit of AI should lead to new technologies, deeper understanding, and smarter decision-making. AI could one day be our best tool to avoid such disasters. However, we should proceed with caution and make clear rules against unacceptable AI applications, such as the development of autonomous weapons.
Now, you've dealt with these huge challenges through the success of game AI, but have you ever been disappointed?
Well, supervised learning has had a huge mainstream impact. Most of the large apps from Google use supervised learning somewhere in the system.
One of my disappointments at the moment is that we haven't found this level of impact on self-learning systems through reinforcement learning. In the future, I would like to see self-learning systems that can interact with people in the virtual world in a way that truly achieves our goals. For example, a digital assistant who learns for himself the best way to achieve a goal. That would be a beautiful achievement.
Do you have a personal goal for your job?
In an AlphaGo match against Lee Sedol, I walked outside the playing field and found a Go player crying. I thought I was upset, but in reality he wasn't.
In this field where he's fully committed, AlphaGo is playing chess that he never realized before. This made him feel a deep sense of beauty.
At my level of Go, it's not enough to fully appreciate this. But we should strive to build similar intelligence wherever we can feel it.
I think AI intelligence should be developed this way, not because of what AI does or how much AI helps us, but because intelligence itself is a beautiful thing.
https://thebulletin.org/2022/01/deepminds-david-silver-on-games-beauty-and-ais-potential-to-avert-human-made-disasters/