How Claude 3.7 Plays Pokémon

With David Hershey of Anthropic!

Latent.Space

Mar 4·38 min

Listen to this conversation38 min

The latest release of Claude 3.7 Sonnet and Claude Code was accompanied by a fun side project and a new “benchmark” that has taken the hearts of gamer-AI-engineers everywhere:

This is very reminiscent of NVIDIA’s Voyager project from 2023 showing similar superior progress on Minecraft, though this was more about the agent framework than the model quality:

You can catch Claude continuing to play Pokemon on their live Twitch stream: https://www.twitch.tv/claudeplayspokemon

We caught up with David Hershey, creator of Claude Plays Pokemon, and grilled him on his architecture diagram:

Watch our discussion on YouTube:

Thanks for reading

Stay a little ahead of the curve.

Fresh ideas, original reporting, and deep conversations in your inbox.

Subscribe to Latent Space

Go a little deeper

The wiki

Sourced profiles and separately reviewed explanations connected to this post.

Person

David Hershey

Creator of Claude Plays Pokémon, with a background in machine learning infrastructure and public views on prompting, fine-tuning, and agent reliability.

1 archive post

Grounded writeup

How Claude Plays Pokémon: Tools, Memory, and the Long Game

Pokémon exposes the gap between choosing a goal and reliably acting on it—and makes improvements in long-running agents visible.

2 archive posts

Search Latent Space

How Claude 3.7 Plays Pokémon

Stay a little ahead of the curve.

Go a little deeper

David Hershey

How Claude Plays Pokémon: Tools, Memory, and the Long Game

Join Latent Space.

How Claude 3.7 Plays Pokémon

Stay a little ahead of the curve.

Go a little deeper

David Hershey

How Claude Plays Pokémon: Tools, Memory, and the Long Game

Keep exploring