Intern recalls building alphago on its tenth anniversary

Ten years after Google DeepMind's AlphaGo defeated Go champion Lee Sedol, Chris Maddison reflects on his role as an intern in developing the groundbreaking AI. The 2016 victory in Seoul marked a pivotal moment in artificial intelligence, demonstrating neural networks' potential to surpass human intuition in complex games. Maddison, now a professor at the University of Toronto, highlights the enduring technological principles behind AlphaGo that influence modern systems like large language models.

In March 2016, Google DeepMind's AlphaGo faced off against Lee Sedol, the world's top Go player, in a five-match series in Seoul, South Korea. The AI won 4-1, shocking observers with its intuitive play. As Sergey Brin noted at the time, "AlphaGo actually does have an intuition. It makes beautiful moves. It even creates more beautiful moves than most of us could think of." Lee Sedol later said he was "in shock."

Chris Maddison, then a master's student, joined the project as an intern in the summer of 2014 after Ilya Sutskever persuaded him with an argument linking Go expertise to neural network capabilities in half a second—comparable to a visual cortex forward pass, as proven in ImageNet. Working with Aja Huang and David Silver, Maddison built neural networks trained on expert games to predict the next move. This simple approach succeeded where others failed; by summer's end, his networks defeated Thore Graepel, a DeepMind researcher and decent Go player.

Go's complexity, with 10^171 possible positions—far exceeding the 10^80 atoms in the observable universe—made it a formidable challenge. AlphaGo advanced by playing millions of games against itself, discovering strategies beyond human play, as Pushmeet Kohli at Google DeepMind explained: "By learning through these games, it could discover new knowledge and could go beyond human-level players."

Maddison left the team before the matches to pursue his PhD but consulted remotely. In Seoul, the atmosphere was intense; crowds lined sidewalks watching the games on large screens, with hundreds of millions in China tuning in. He recalled Aja Huang describing Lee Sedol as "one stone from God," underscoring the gap they bridged.

AlphaGo's legacy endures. Noam Brown at OpenAI stated, "AlphaGo definitively showed that neural nets can do pattern recognition better than humans. They can essentially have intuition that surpasses humans." Its method—pretraining on vast data like Go games or internet text, followed by reinforcement learning to align with goals—mirrors large language models. Successors include AlphaFold, which earned a Nobel Prize in chemistry for protein prediction, and AlphaProof, achieving gold-medal performance at the International Mathematical Olympiad.

Yet challenges remain: neural networks are black boxes, as seen in AlphaGo's unexplained move 37, initially puzzling spectators. Progress hinges on abundant data and clear reward signals, particularly in fields like mathematics and programming. Maddison expressed sympathy for Lee Sedol, who apologized to humanity after the loss and couldn't traditionally review the match with the AI. Still, he sees AI enhancing human appreciation of games like Go and chess, preserving their cultural purpose beyond mere victory.

Awọn iroyin ti o ni ibatan

Former Go champion Lee Sedol will meet artificial intelligence again on Monday, nearly a decade after his landmark match against Google's AlphaGo. This time, the encounter emphasizes collaboration over competition. Lee will work with an agentic AI to create a Go game model.

Ti AI ṣe iroyin

Google DeepMind's AlphaFold artificial intelligence system has marked its fifth anniversary, continuing to evolve after revolutionizing biology and chemistry. The project earned the Nobel Prize in Chemistry last year for its groundbreaking contributions. WIRED recently discussed its trajectory with DeepMind's Pushmeet Kohli.

At the AGI-Next summit in Beijing, Alibaba AI scientist Lin Junyang warned that China has less than a 20% chance of exceeding the US in artificial intelligence over the next 3 to 5 years due to resource limits. He pointed out the gap, with US firms like OpenAI pouring massive computational resources into next-generation research while China is stretched thin just meeting daily demands.

Ti AI ṣe iroyin

A study applying Chile's university entrance exam, PAES 2026, to AI models shows several systems scoring high enough for selective programs like Medicine and Civil Engineering. Google's Gemini led with averages near 950 points, outperforming rivals like ChatGPT. The experiment underscores AI progress and raises questions about standardized testing efficacy.

 

 

 

Ojú-ìwé yìí nlo kuki

A nlo kuki fun itupalẹ lati mu ilọsiwaju wa. Ka ìlànà àṣírí wa fun alaye siwaju sii.
Kọ