The three-way AI talent war between OpenAI, Meta, and Elon Musk’s SpaceXAI has mostly played out through billion-dollar hires and poached researchers. This week, it spilled over into something considerably weirder: a full browser-based video game, built by a SpaceXAI employee, that turns a rival into a boss fight.
How the beef started
The back-and-forth had been simmering for a while, but escalated when an OpenAI’s Tibo Sottiaux registered the domain poteto.com, which redirects to ChatGPT, taking a jab at a SpaceXAI engineer who goes by “Poteto.” Meta’s chief AI officer Alexandr Wang then got involved, registering thibaultsottiaux.com, which redirects to Muse, the AI assistant Meta launched this year through its Superintelligence Labs.
Poteto’s response was to build an actual game.
Tibo.dog: a boss fight built in four days
Poteto has now launched tibo.dog, a browser-playable survival game reportedly built in four days using Grok and the AI coding tool Cursor. In the game, players control a character modeled on Tibo, fighting off waves of enemies represented as users demanding usage-limit resets, before facing a final boss themed around Muse and modeled on Alexandr Wang, whose hiring by Meta in a deal worth roughly $14.3 billion made him one of the AI industry’s highest-profile executives this year.
According to Poteto’s own account of the build, posted on X, the project only cost $5.99 (the price of the tibo.dog domain) but involved some fairly serious engineering for a four-day side project. Poteto says the game racked up roughly 1,150 pull requests over those four days, with every bug that turned up in the code becoming a new lint rule so the same mistake couldn’t be made twice. The simulation is fully deterministic, meaning every run can be replayed exactly from its inputs, which Poteto says is also how the game checks leaderboard submissions for cheating, by replaying the run server-side before accepting a high score. Performance-critical code gets its own file extension with a lint rule banning memory allocation inside it, so the game loop avoids garbage collection stutters, and the whole thing was reportedly stress-tested by bots playing it headless to tune difficulty against target survival rates. There are no audio files: the music and sound effects are synthesized live in the browser by a custom Game Boy-style sound chip layered with a synthwave track, written as text directly in the code, and the soundtrack is built to transition seamlessly between biomes without breaking the melody. The pixel art works the same way, with every sprite defined in code from a shared palette and packed into an atlas at build time. Poteto also says the game supports 10 languages, generates a dynamic shareable image for every run based on that player’s final stats, runs at 60 frames per second with 500 enemies on screen even on a mid-range phone, is playable at sizes from a small embedded X card up to full screen, and ships at around 124 KB gzipped, a budget the build reportedly fails if it’s exceeded. Code comments, notably, are banned by a lint rule: if something needs explaining, the fix is a better name or a test.
The game has been circulating widely on X, with many in the AI community calling it one of the sharper pieces of corporate-rivalry trolling the industry has produced, given that it was reportedly built almost entirely with AI coding tools in under a week.
Part of a bigger pattern
The stunt is the latest flashpoint in a rivalry that has defined AI hiring all year. Meta’s push to build out its Superintelligence Labs under Alexandr Wang, following its multi-billion-dollar move on Scale AI, directly targeted researchers at OpenAI and elsewhere, and the friction between the companies has shown up repeatedly in public jabs between employees. Elon Musk’s AI unit, meanwhile, was folded into SpaceX and rebranded SpaceXAI earlier this year, putting Poteto’s game inside the same corporate umbrella as SpaceX and Musk’s broader AI ambitions.
Whether or not tibo.dog settles anything, it’s a reminder that the people writing the model cards and training runs at these companies are also, apparently, very online.