Eric Jang – Building AlphaGo from scratch

Implementing AlphaGo from scratch reveals the mechanics of AI search and reasoning, specifically how Monte Carlo Tree Search (MCTS) combined with neural networks makes intractable game-tree searches computationally feasible. By using a policy network to guide move selection and a value network to evaluate board states, AI systems effectively "amortize" deep search, achieving superhuman performance without exhaustive computation. This process demonstrates how neural networks compress complex simulation tasks into efficient forward passes. The discussion highlights the shift from model-free reinforcement learning to search-based methods, the role of self-play in bootstrapping performance, and the potential for automated research agents to accelerate scientific discovery by iteratively refining experimental hypotheses. These insights underscore the power of combining search heuristics with deep learning to solve problems previously considered beyond the reach of current computational limits.