Google’s most expensive model seems to have crossed an important milestone: beating a 29 -year -old video game.
Last night, Google Sundar Pichai CEO published triumphantly on X, “What finish! Gemini 2.5 Pro has just completed Pokémon Blue!”
To be clear, the Gemini Plays Pokemon Livestream was created by (with his own words) “a 30 -year -old software engineer not affiliated to Google” who goes to Joel Z. But Google managers encouraged the effort.
For example, Logan Kilpatrick, the protagonist of Google to the Studio, published last month that Gemini was “making great progress in completing Pokémon” and had “earned his 5th badge (the next best model has only 3 so far, although with a different harness of the agents)”, leading Pichai to joke, “we are working on the API, Pokémon artificial intelligence 🙂
Why Pokémon? In February, Anthropic highlighted the progress that his Claude Ai models were realizing in “Pokémon Red”, writing that Claude’s “extended thought and training” gives him “a great impulse” on “more unexpected” tasks, such as playing a classic game. (“Pokémon Red” and “Blue” are different versions of a gameboy title released for the first time in 1996 and linked to the longtime Pokémon franchise). There is also a Pokemon Twitch channel by Claude Plays that Joel Z has mentioned as inspiration.
Despite his progress, Claude does not yet seem to have beaten “Pokémon Red”. Does it mean that Gemelli is objectively better in the game? On his contraction page, Joel Z has urged the spectators, “please not consider this a point of reference, however well a LLM can play Pokemon. You cannot really make direct comparisons: Gemini and Claude have different tools and receive different information”.
And both artificial intelligence models need help to play the game: this is where the agents above comes into play, providing the game screenshot models superimposed on further information, allowing the model to decide how to respond (which can involve the call of specialized agents) and therefore by pressing the button that corresponds with the instructions of the AI.
Techcrunch event
Berkeley, ca.
|
June 5th
Book now
Joel Z recognized that there were other “development interventions” to help Gemini complete the game, but insisted on the fact that he is not betraying.
“My interventions improve the overall Gemelli decision -making and reasoning skills,” he says. “I do not give specific suggestions: there are no direct details or instructions for particular challenges such as Mount Moon. The only thing that approaches even is to let Gemelli know that he has to talk to a rocket grunt twice to get the lifting button, which was a bug that was subsequently fixed in Pokemon Yellow.”
In addition, he said, “Gemini Plays Pokémon is still actively developed and the framework continues to evolve”.
