Despite being a big deal in the AI world, Google’s Gemini AI turns out to be hilariously bad at playing Pokémon Blue. The AI spent over 800 hours just to beat the Elite Four, often ‘panicking’ when its Pokémon were low on health or Power Points, making some baffling moves.
This comes from a Twitch channel called Gemini_Plays_Pokemon, created by Joel Zhang, an engineer not affiliated with Google. The DeepMind report on the Gemini 2.X model family highlights this quirky example, where the AI tries to run away or makes poor decisions when the situation gets tough, something they dubbed “Agent Panic.”
Imagine your Pokémon’s health bar is dangerously low, and instead of strategizing, the AI forgets to use its tools and freaks out. It even got stuck fixating on a fake item called “Tea,” which isn’t in the original game but in a remake. Because of all this, the first playthrough took over 813 hours. After some tweaks, the second run was faster but still a whopping 406.5 hours.
For comparison, Pokémon Blue’s main story usually takes about 26 hours to complete. Yeah, Gemini’s performance is pretty abysmal for a game over 25 years old. It’s funny, though, to watch an AI struggle with a childhood classic.
While the report includes some cool scatter graphs showing the AI’s slow progress, I’m not sold on using AI playing games as a serious benchmark. Honestly, thinking about AI chewing through games like this feels a bit bleak. Plus, the report’s use of words like “panic” really tries to humanize the AI, but remember, it doesn’t feel anything; it’s just patterns and data.
Seeing Gemini’s botched attempts at Pokémon is a neat curiosity, but I wouldn’t start cheering for it just yet. What do you think? Is it funny to watch AI struggle at games, or kinda sad? Drop your thoughts in the comments below!
