Based on Eric Jang's experience, large language models like Claude Opus 4.6 and 4.7 are effective at hyperparameter optimization and experiment execution but are not yet capable of high-level strategi...

Eric Jang's project to rebuild AlphaGo was funded by a $10,000 donation from Prime Intellect, of which approximately $7,000 was spent on research and model serving.

DeepMind's institutional experience in solving games like Go and StarCraft likely provided a positive transfer of research skills to their subsequent work on large language models.

Systems like AlphaStar and OpenAI's Dota bot used an algorithm called Neural Fictitious Self-Play (NFSP) instead of Monte Carlo Tree Search.

The 2021 paper by Andy Jones on scaling laws for board games also showed that it is possible to predict the amount of compute required to solve a larger version of a board game.

Most Go practitioners today train against the AI model Katago.

All Go AIs are trained against and resolve games using the Tromp-Taylor rules because they are completely unambiguous for computers.

The naive search tree for a game of Go has a complexity on the order of 361 to the power of 300, which is larger than the number of atoms in the universe.

Eric Jang found that Claude 4.6 was able to generate a reasonable data structure for a Monte Carlo Tree Search implementation.

The Katago paper found that aggregating global features throughout the network was useful for giving the model a global sense of the board state.

A Go-playing neural network trained on expert human data, without any search, can become a very strong player that beats most humans by simply taking the highest-probability move from its policy netwo...

Katago proposed an architecture that can be trained on both 9x9 and 19x19 boards, enabling effective transfer learning from the smaller board to the larger one, particularly for the value head.

The Monte Carlo Tree Search (MCTS) algorithm provides a low-variance learning signal for every action, in contrast to naive reinforcement learning methods which suffer from high variance.

Current large language model reinforcement learning (LLM RL) typically treats an entire generated sequence as a single action, which is a major contributor to the high variance of the learning signal.

Q-learning propagates value estimates backward over trajectories an agent has already visited, whereas Monte Carlo Tree Search plans forward over trajectories the agent has not yet been to.

The compute required to be the first to achieve a research breakthrough is always much larger than the compute it takes for others to catch up and replicate the result.

Off-policy training can harm performance if the replay buffer contains too many states that the current policy would never visit, causing the model to waste capacity on irrelevant states.

Many algorithmic improvements that act as "compute multipliers" may not stack effectively because they can have correlated benefits or become redundant as hardware performance increases.

To learn tabula rasa like AlphaZero or Katago, one can start by playing tens of thousands of random games on a smaller 9x9 board to generate enough data to train an initial, reasonably accurate value ...

Neural Fictitious Self-Play (NFSP) works by training a "best response" policy against a fixed opponent using a model-free reinforcement learning algorithm.