Topics

Latest

AI

Amazon

Article image

Image Credits:picture alliance / Getty Images

Apps

Biotech & Health

clime

Article image

Image Credits:picture alliance / Getty Images

Cloud Computing

Commerce Department

Crypto

Enterprise

EVs

Fintech

Fundraising

Gadgets

stake

Google

Government & Policy

Hardware

Instagram

Layoffs

Media & Entertainment

Meta

Microsoft

privateness

Robotics

Security

societal

Space

Startups

TikTok

Transportation

speculation

More from TechCrunch

Events

Startup Battlefield

StrictlyVC

Podcasts

Videos

Partner Content

TechCrunch Brand Studio

Crunchboard

touch Us

Google’smost expensive AI modelseems to have crossed a major milestone : Beating a 29 - year - sometime video game .

Last dark , Google CEO Sundar Pichaiposted triumphantly on X , “ What a finish ! Gemini 2.5 Pro just completed Pokémon Blue ! ”

To be clear , theGemini Plays Pokemon livestreamwas created by ( in his own word ) “ a 30 year old software locomotive engineer unaffiliated with Google ” who go bad byJoel Z. But Google executives have been cheering the exploit on .

For example , Logan Kilpatrick , the product conduct for Google AI Studio , posted last monththat Gemini was “ making great progress at discharge Pokémon ” and had “ earned its 5th badge ( next upright mannequin only has 3 so far , though with a different agentive role harness ) , ” leading Pichai tojoke , “ We are working on API , Artificial Pokémon Intelligence :) ”

Why Pokémon ? Back in February , Anthropic highlight progressthat its Claude AI simulation were making in “ Pokémon Red , ” compose that Claude ’s “ extended thinking and agentive role education ” cave in it “ a major boost ” on “ more unexpected ” tasks , like playing a classical game . ( “ Pokémon Red ” and “ Blue ” are different versions ofa GameBoy titlefirst release in 1996 and draw to the long - draw Pokémon enfranchisement ) . There ’s evena Claude Plays Pokemon Twitch channelthat Joel Z cited as an inspiration .

Despite its onward motion , Claude does not appear to have beat “ Pokémon Red ” yet . Does that think Gemini is objectively good at the secret plan ? On his Twitch page , Joel Z urged viewers , “ Please do n’t consider this a benchmark for how well an LLM can bet Pokemon . You ca n’t really make verbatim equivalence — Gemini and Claude have unlike puppet and receive different information . ”

And both AI models necessitate help to play the game — that ’s wherethe aforementioned federal agent harnessescome in , providing the model with plot screenshots overlaid with additional information , allow the model to adjudicate how to respond ( which may demand calling specialized agent ) , and then pressing the button that match with the AI ’s instruction .

Join us at TechCrunch Sessions: AI

Exhibit at TechCrunch Sessions: AI

Joel Z admit that there were other “ dev interventions ” to help Gemini fill out the game , but take a firm stand that it ’s not cheating .

“ My interventions improve Gemini ’s overall determination - making and abstract thought power , ” he says . “ I do n’t give specific hint — there are no walkthroughs or direct instructions for particular challenge like Mt. Moon . The only thing that hail even close is letting Gemini cognize that it needs to talk to a Rocket Grunt twice to obtain the Lift Key , which was a hemipterous insect that was by and by fixed in Pokemon Yellow . ”

Plus , he say , “ Gemini roleplay Pokémon is still actively being developed , and the framework continues to evolve . ”