Researcher Eric Jang reconstructs AlphaGo from scratch in a lecture using modern AI tools and frameworks to detail Monte Carlo Tree Search, self-play training, and contrasts with policy-gradient methods in large language models

VIEWS367.7KBOOKMARKS1.8KLIKES2.2KREPLIES41

For the last few months I've been working on a from-scratch implementation of AlphaGo, a 2016 AI breakthrough that inspired me to get into deep learning. My casual understanding of AlphaGo was "search-augmented deep neural networks trained with self-play", but I wanted to go deeper and understand it by creating it.

Frontier deep learning research has always been expensive, but any given capability gets cheaper very quickly. In 2026, you no longer need DeepMind's resources to train a strong Go AI - you can vibe code all of it yourself for just a few thousand dollars of rented compute.

It was a huge honor to be invited to teach this with @dwarkesh_sp on @dwarkeshpodcast

I am an AlphaGo & Go apprentice, not a master, so all factual errors in the podcast are mine.

Web version of tutorial: https://evjang.com/2026/04/28/autogo.html Code: https://github.com/ericjang/autogo Play the go bot here: https://autogo.evjang.com/

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d367.7K2.2K1.8K

RETWEETS271

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d590.5K2.7K2.9K

François Fleuret@francoisfleuret

Awesome.

Seriously, people are harsh with this platform, but if you are careful with whom you follow, it is a constant stream of awesomness.

Eric Jang@ericjang11

For the last few months I've been working on a from-scratch implementation of AlphaGo, a 2016 AI breakthrough that inspired me to get into deep learning. My casual understanding of AlphaGo was "search-augmented deep neural networks trained with self-play", but I wanted to go deeper and understand it by creating it.

Frontier deep learning research has always been expensive, but any given capability gets cheaper very quickly. In 2026, you no longer need DeepMind's resources to train a strong Go AI - you can vibe code all of it yourself for just a few thousand dollars of rented compute.

It was a huge honor to be invited to teach this with @dwarkesh_sp on @dwarkeshpodcast

I am an AlphaGo & Go apprentice, not a master, so all factual errors in the podcast are mine.

Web version of tutorial: https://evjang.com/2026/04/28/autogo.html Code: https://github.com/ericjang/autogo Play the go bot here: https://autogo.evjang.com/

44d150.4K1.2K385

Eric Jang@ericjang11

I'd like to issue a clarification on the section covered around 1:33:44 of the podcast, where @dwarkesh_sp and I discuss multi-step vs. single-step LLM RL. I muddle through some hand-wavy quadratic variance statements around multi-step RL, and make the claim that multi-step RL has higher variance than single-step RL. This is not 100% right, or confusing at best for people following along, so I'd like to issue some clarification. This was my mistake🥹

Full errata writeup here: https://evjang.com/alphago-tutorial/llm_rl_variance.pdf

Eric Jang@ericjang11

For the last few months I've been working on a from-scratch implementation of AlphaGo, a 2016 AI breakthrough that inspired me to get into deep learning. My casual understanding of AlphaGo was "search-augmented deep neural networks trained with self-play", but I wanted to go deeper and understand it by creating it.

Frontier deep learning research has always been expensive, but any given capability gets cheaper very quickly. In 2026, you no longer need DeepMind's resources to train a strong Go AI - you can vibe code all of it yourself for just a few thousand dollars of rented compute.

It was a huge honor to be invited to teach this with @dwarkesh_sp on @dwarkeshpodcast

I am an AlphaGo & Go apprentice, not a master, so all factual errors in the podcast are mine.

Web version of tutorial: https://evjang.com/2026/04/28/autogo.html Code: https://github.com/ericjang/autogo Play the go bot here: https://autogo.evjang.com/

39d71.7K460362

λux@novasarc01

really loving this new blackboard-style explanation format in dwarkesh’s podcast!

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d45.4K371112

Steve Hou@stevehou

"how to build AlphaGo from scratch, but with modern AI tools."

Wow I was just thinking exactly this when I was re-listening to the "Infinity Machine" audiobook when I was going over the part where David Silver built AlphaGo Zero.

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d8.9K2739

Dwarkesh Patel@dwarkesh_sp

@ericjang11 Watch with video if you can: https://youtu.be/X_ZVSPcZhtw

And check out the flashcards I wrote to retain the insights: https://flashcards.dwarkesh.com/eric-jang/

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d16.6K5330

Eric Jang@ericjang11

For the last few months I've been working on a from-scratch implementation of AlphaGo, a 2016 AI breakthrough that inspired me to get into deep learning. My casual understanding of AlphaGo was "search-augmented deep neural networks trained with self-play", but I wanted to go deeper and understand it by creating it.

Frontier deep learning research has always been expensive, but any given capability gets cheaper very quickly. In 2026, you no longer need DeepMind's resources to train a strong Go AI - you can vibe code all of it yourself for just a few thousand dollars of rented compute.

It was a huge honor to be invited to teach this with @dwarkesh_sp on @dwarkeshpodcast

I am an AlphaGo & Go apprentice, not a master, so all factual errors in the podcast are mine. For example, I mistakenly say 361 several times when it should be 381 🤦‍♂️

Web version of tutorial: https://evjang.com/2026/04/28/autogo.html Code: https://github.com/ericjang/autogo Play the go bot here: https://autogo.evjang.com/

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d7.5K4827

Dwarkesh Patel@dwarkesh_sp

See the full lecture here:

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

41d9.6K2513

Eric Jang@ericjang11

PS i'm working on some RL envs for automated research based on this. reach out if you're interested

Eric Jang@ericjang11

For the last few months I've been working on a from-scratch implementation of AlphaGo, a 2016 AI breakthrough that inspired me to get into deep learning. My casual understanding of AlphaGo was "search-augmented deep neural networks trained with self-play", but I wanted to go deeper and understand it by creating it.

Frontier deep learning research has always been expensive, but any given capability gets cheaper very quickly. In 2026, you no longer need DeepMind's resources to train a strong Go AI - you can vibe code all of it yourself for just a few thousand dollars of rented compute.

It was a huge honor to be invited to teach this with @dwarkesh_sp on @dwarkeshpodcast

I am an AlphaGo & Go apprentice, not a master, so all factual errors in the podcast are mine.

Web version of tutorial: https://evjang.com/2026/04/28/autogo.html Code: https://github.com/ericjang/autogo Play the go bot here: https://autogo.evjang.com/

44d4.3K3710

will brown@willccbb

@dwarkesh_sp @ericjang11 was glossed over but the UCB proof is really simple + kinda cute. after K samples of an action, you have some variance of your mean estimate wrt true mean, which goes down exp in K -> high-conf window

regret guarantee = the true mean is in the window for each, w/ prob 1 - eps

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d1.6K154

Delip Rao e/σ@deliprao

@dwarkesh_sp @ericjang11 You are winning with this format. Excellent videos!

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d1.2K131

kache@yacineMTB

@dwarkesh_sp @ericjang11 Thanks for doing these man

Dwarkesh Patel@dwarkesh_sp

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

45d1.1K180

Yassine Yousfi@yassineyousfi_

@ericjang11 "Podcasts with code." This raised the bar of technical podcasts. Eric and Dwarkesh! I stopped mid-episode to open the code and follow through while reading the relevant parts of the code! Super cool!

44d36851