What happened with eric jang says ai agents slash alphago's cost to thousands?

Eric Jang rebuilt AlphaGo using LLM assistants, cutting its cost from millions to a few thousand dollars.

What happened with eric jang says ai agents slash alphago's cost to thousands?

Steve Lee argues software is shifting from Bitcoin protocol engineering to consumer utility using agents.

AI & TECH

Eric Jang says AI agents slash AlphaGo's cost to thousands

Sunday, May 17, 2026 · from 2 podcasts

2 SOURCESPresidio Bitcoin Jam Dwarkesh Podcast

Presidio Bitcoin Jam Dwarkesh Podcast

Eric Jang rebuilt AlphaGo using LLM assistants, cutting its cost from millions to a few thousand dollars.
Steve Lee argues software is shifting from Bitcoin protocol engineering to consumer utility using agents.

The barrier to creating AI agents just collapsed. Eric Jang rebuilt AlphaGo from scratch on a sabbatical, leveraging David Wu’s KataGo and LLM-assisted coding. What once required a Google-sized fleet of PhDs and millions in compute is now within reach of a single engineer.

Steve Lee argues software development is undergoing a parallel shift. On Presidio Bitcoin Jam, he pointed to Matt Belez’s Babel Agent - a tool that live-transcribes podcasts using LLMs funded by Lightning payments - as evidence of “vibe coding.” These aren’t traditional Bitcoin apps; they are useful tools that use Bitcoin because it’s the only viable rail for global micro-payments.

“The infrastructure has matured to the point where the ‘idea guy’ is now a software developer. The focus is shifting from protocol engineering to consumer utility.”
- Steve Lee, Presidio Bitcoin Jam

Jang speculates neural networks compressing NP-hard problems like protein folding hint our computational hardness classifications might be incomplete. He suggests AlphaGo later versions dropped random playouts, but preventing resignation in 10% of games ensures late-game value accuracy.

Steve Lee sees this convergence enabling the first Bitcoin apps to generate $1 million in revenue by the end of 2026. Spiral’s Project Loupe is an open-source AI security scanner for Bitcoin repos that uses LLMs to find vulnerabilities, filters results, and funds token costs with Bitcoin. The era of niche, PhD-heavy research departments is ending.

Models Startups AI Infrastructure

Steve Lee Spiral

Source Intelligence

- Deep dive into what was said in the episodes

Presidio Bitcoin Jam

Bitcoin Core v31 Release, Project Loupe Launches, Lightning Network's Future • May 15

Steve believes the first app generating $1M in Bitcoin will launch by year-end 2026, citing mature developer tools and needing more people to take swings.
LDK Server is a new binary daemon from Spiral that simplifies running a Lightning node or LSP, offering features like splicing and Bolt 12 ahead of LND.

Also from this episode: (13)

AI & Tech (2)

Matt Belez built Babbel Agent, a tool that live translates podcast audio into any language using an LLM funded by Lightning payments.
Spiral's Project Loop is an open-source AI security scanner for Bitcoin repos that uses LLMs to find vulnerabilities, filters results, and funds the token costs.

Protocol (11)

Steve pitches vibe coding or hypercoding to people between jobs as a better path than just studying Bitcoin.
Steve recounts a Slack message from a Block colleague wanting to license a PBJ clip for paid advertising, suggesting content licensing could help fund community treasuries.
Bitcoin Core v31 includes embedded ASMAP to diversify peer connections across Autonomous Systems, defending against eclipse attacks by ensuring nodes don't connect only to one cloud provider.
A new privacy feature in Core v31 creates ephemeral connections to Tor or I2P peers to broadcast transactions, obscuring the origin IP address from surveillance companies.
Cluster mempool in Core v31 groups related transactions for fee optimization and improves Lightning security, addressing historical ancestor limit issues for layer-two protocols.
Steve argues mining pools don't optimize fee selection algorithms because block rewards dominate revenue; fee optimization becomes relevant only in 10-20 years.
BNOC (Bitcoin Network Operations Center) provides public network data tools, including an OFAC censorship detector to identify if mining pools omit high-fee transactions.
Steve advocates for fusing Bitcoin payments with Loop to create a sustainable model where the community pays for scans, moving away from central Block funding.
Steve positions LDK as the purest public good Lightning implementation due to its open contribution model, public communication, and lack of commercial revenue pressure.
DK argues Nostr keys could improve PGP-based trust ceremonies for Bitcoin Core binaries, enabling a decentralized web of trust for software verification.
Cash App is the largest user of the Bitcoin blockchain, with Miles Suter stating it accounted for 4-8% of network activity at the 2023 Bitcoin Conference.

Protocol Mining Lightning Agents BTC Markets

Dwarkesh Podcast

Eric Jang – Building AlphaGo from scratch • May 15

Eric Jang emphasizes the core breakthrough of AlphaGo: a 10-layer neural network amortizes an intractable search problem like Go, compressing deep simulations into a single forward pass.
David Wu’s Cadego achieved a 40x compute reduction to train a strong Go bot, showing how LLM coding now lets individuals replicate what required DeepMind's team and millions of dollars.
Go scoring uses Tromp-Taylor rules for AI training because they are unambiguous for computers, unlike Chinese or Japanese scoring which relies on human consensus.
Monte Carlo Tree Search (MCTS) uses a four-step process: selection via PUCT criteria, expansion of new nodes, evaluation with the value network, and backup to update action values.
AlphaGo's training distills the MCTS-improved policy back into the neural network, letting the network start from a stronger baseline so search needs less work.
Jang finds Resnets outperform Transformers for Go in low-budget experiments because convolutions provide a useful local inductive bias; transformers need more data for global context.
Training starts with supervised learning on expert human games to get good initial value estimates, especially for late-game states, before moving to full self-play.
Jang links MCTS policy improvement to the DAGGER algorithm in robotics: both relabel suboptimal actions in a trajectory with better ones provided by a teacher.
Neural Fictitious Self-Play, used in Alpha Star and OpenAI's Dota, is an alternative when tree search is impossible; it trains best-response policies against fixed opponents and distills them.
Model-free RL like PPO has high gradient variance because it must solve credit assignment across long sequences; MCTS avoids this by providing a better label for every action directly.
Jang speculates that neural networks compressing NP-hard problems like protein folding hints our computational hardness classifications might be incomplete; they solve macroscopic structures not worst-case exact solutions.
AlphaGo Lee used a mix of neural network value estimates and random playouts to the end for grounding; later versions dropped playouts, but Jang suggests preventing resignation in 10% of games ensures late-game value accuracy.

Models Reasoning Startups AI Infrastructure

Source Intelligence

- Deep dive into what was said in the episodes

Presidio Bitcoin Jam

Bitcoin Core v31 Release, Project Loupe Launches, Lightning Network's Future • May 15

Steve believes the first app generating $1M in Bitcoin will launch by year-end 2026, citing mature developer tools and needing more people to take swings.
LDK Server is a new binary daemon from Spiral that simplifies running a Lightning node or LSP, offering features like splicing and Bolt 12 ahead of LND.

Also from this episode: (13)

AI & Tech (2)

Matt Belez built Babbel Agent, a tool that live translates podcast audio into any language using an LLM funded by Lightning payments.
Spiral's Project Loop is an open-source AI security scanner for Bitcoin repos that uses LLMs to find vulnerabilities, filters results, and funds the token costs.

Protocol (11)

Steve pitches vibe coding or hypercoding to people between jobs as a better path than just studying Bitcoin.
Steve recounts a Slack message from a Block colleague wanting to license a PBJ clip for paid advertising, suggesting content licensing could help fund community treasuries.
Bitcoin Core v31 includes embedded ASMAP to diversify peer connections across Autonomous Systems, defending against eclipse attacks by ensuring nodes don't connect only to one cloud provider.
A new privacy feature in Core v31 creates ephemeral connections to Tor or I2P peers to broadcast transactions, obscuring the origin IP address from surveillance companies.
Cluster mempool in Core v31 groups related transactions for fee optimization and improves Lightning security, addressing historical ancestor limit issues for layer-two protocols.
Steve argues mining pools don't optimize fee selection algorithms because block rewards dominate revenue; fee optimization becomes relevant only in 10-20 years.
BNOC (Bitcoin Network Operations Center) provides public network data tools, including an OFAC censorship detector to identify if mining pools omit high-fee transactions.
Steve advocates for fusing Bitcoin payments with Loop to create a sustainable model where the community pays for scans, moving away from central Block funding.
Steve positions LDK as the purest public good Lightning implementation due to its open contribution model, public communication, and lack of commercial revenue pressure.
DK argues Nostr keys could improve PGP-based trust ceremonies for Bitcoin Core binaries, enabling a decentralized web of trust for software verification.
Cash App is the largest user of the Bitcoin blockchain, with Miles Suter stating it accounted for 4-8% of network activity at the 2023 Bitcoin Conference.

Protocol Mining Lightning Agents BTC Markets

Dwarkesh Podcast

Eric Jang – Building AlphaGo from scratch • May 15

Eric Jang emphasizes the core breakthrough of AlphaGo: a 10-layer neural network amortizes an intractable search problem like Go, compressing deep simulations into a single forward pass.
David Wu’s Cadego achieved a 40x compute reduction to train a strong Go bot, showing how LLM coding now lets individuals replicate what required DeepMind's team and millions of dollars.
Go scoring uses Tromp-Taylor rules for AI training because they are unambiguous for computers, unlike Chinese or Japanese scoring which relies on human consensus.
Monte Carlo Tree Search (MCTS) uses a four-step process: selection via PUCT criteria, expansion of new nodes, evaluation with the value network, and backup to update action values.
AlphaGo's training distills the MCTS-improved policy back into the neural network, letting the network start from a stronger baseline so search needs less work.
Jang finds Resnets outperform Transformers for Go in low-budget experiments because convolutions provide a useful local inductive bias; transformers need more data for global context.
Training starts with supervised learning on expert human games to get good initial value estimates, especially for late-game states, before moving to full self-play.
Jang links MCTS policy improvement to the DAGGER algorithm in robotics: both relabel suboptimal actions in a trajectory with better ones provided by a teacher.
Neural Fictitious Self-Play, used in Alpha Star and OpenAI's Dota, is an alternative when tree search is impossible; it trains best-response policies against fixed opponents and distills them.
Model-free RL like PPO has high gradient variance because it must solve credit assignment across long sequences; MCTS avoids this by providing a better label for every action directly.
Jang speculates that neural networks compressing NP-hard problems like protein folding hints our computational hardness classifications might be incomplete; they solve macroscopic structures not worst-case exact solutions.
AlphaGo Lee used a mix of neural network value estimates and random playouts to the end for grounding; later versions dropped playouts, but Jang suggests preventing resignation in 10% of games ensures late-game value accuracy.

Models Reasoning Startups AI Infrastructure

Eric Jang says AI agents slash AlphaGo's cost to thousands

Source Intelligence

Related Stories

Eric Jang says AI agents slash AlphaGo's cost to thousands

Source Intelligence

Related Stories