A portrait of Lee Sedol sat on the desk. Not as motivation in any conventional sense — more as a fixed point, a measure of how far the work still had to go.
That detail comes from Chris Maddison, now a professor of artificial intelligence at the University of Toronto, who in the summer of 2014 joined Google Brain as an intern with a specific task: build the neural networks that would eventually become AlphaGo. He was a master’s student at the time.
The project started with a phone call, or rather an argument. Ilya Sutskever — who later co-founded OpenAI — contacted Maddison with a deceptively simple line of reasoning. If an expert Go player can identify the best move in half a second, Sutskever told him, then a neural network should be able to approximate that decision. The logic rested on what was already known about visual processing: half a second is roughly the time the human visual cortex needs for a single forward pass, and researchers had already demonstrated, through the ImageNet competition, that neural networks could approximate that kind of processing reliably. Maddison says he bought the argument.
When he arrived at DeepMind, a small team was already in place: Aja Huang and David Silver had started working on Go. Maddison’s job was to build the networks around that foundation. Many early approaches failed. Eventually, he says, frustration led him to the simplest possible idea: train a neural network to predict the next move an expert would make, using a large collection of expert games as training data. It worked.
One stone from God
By the end of that summer, Maddison’s networks were tested in a small internal match against Thore Graepel, a DeepMind researcher who considered himself a capable Go player. The networks won. That result, according to Maddison, was what convinced the company to commit real resources to the project.
Progress came steadily. Each new network played slightly better than the last. Maddison would turn to Huang — an actual Go player — and ask how close they were getting to Sedol’s level. The answer was always the same. “Chris, you don’t understand,” Huang told him. “Lee Sedol is one stone from God.”
Maddison left before the project reached its conclusion. David Silver asked him to stay and drive the work forward. He declined, choosing to return to his PhD. He describes it, in the interview, as “maybe one of the stupider decisions I made.” He stayed loosely connected as a consultant. He notes, with some pride, that it took the team time to surpass his original networks.
A match on a big screen
He was present in Seoul in March 2016 when AlphaGo faced Sedol in the five-match series that was broadcast to an audience that, by some accounts, numbered in the hundreds of millions. From his hotel room, high enough to see a major city intersection, Maddison looked out and saw a large screen — his words evoke something close to Times Square — showing the match live. People lined the sidewalks watching.
The final system that played Sedol was, Maddison says, “the product of a big engineering effort and a big team.” The intern’s networks were the floor it was built from.
Photo by Pixabay
This article is a curated summary based on third-party sources. Source: Read the original article
