Anthropic is training its AI by having it play Pokemon on Twitch

When it comes to Pokémon, Anthropic‘s Claude AI has gotta scrape ’em all.

The generative AI bot has joined in on the decade-old phenomenon known as Twitch Plays Pokémon. On a channel called Claude Plays Pokémon, the latest version of Anthropic’s large language model is exploring the original Game Boy version of Pokémon Red.

Twitch Plays Pokémon began in 2014 and quickly became a viral phenomenon, reaching millions of viewers. By inviting scores of players to choose controller inputs via chat, the original Twitch Plays Pokémon stream created an entertaining brand of chaos that enchanted fans and spawned numerous memes.

After recently celebrating its tenth birthday, Twitch Plays Pokémon is still going strong. The game has been played by a long list of quirky entities, ranging from someone’s pet fish to the mathematical constant pi.

Claude is the latest unusual player to pick up a Game Boy and take on the role of a Pokémon trainer. Previous iterations of the LLM have stalled before reaching the end of the game, but Claude 3.7 Sonnet hasn’t gotten stuck yet — and it’s explaining the reasoning behind its inputs as it goes.

The result is an unusual form of AI training that doubles as entertainment. Viewers on platforms like Twitch and YouTube have shown that they have a big appetite for deep reinforcement learning videos that show how AIs become smarter. One YouTube clip in that genre has more than 10 million views.

For Anthropic, Claude Plays Pokémon has multiple uses. It’s a tool for deep reinforcement learning, but it’s also a more ethical form of research for a company that has been accused of slipshod AI training methods. It was one of the companies that fed its LLMs the unauthorized content in The Pile, and it recently settled a lawsuit from music publishers who accused it of scraping lyrics without permission.

With Pokémon Red serving as Claude’s new training material, the generative AI innovation will develop an interesting set of skills. Claude may still struggle to walk across a room in Professor Oak’s lab, but it has the game’s combat system down pat. Watch out, Ash Ketchum: Claude might put you out of a job.