Kingmakers' Developer Breaks Down OpenAI o1 Making Games

A week ago, OpenAI, the company behind DALL-E and ChatGPT, showcased its new series of AI models called OpenAI o1, which can write code, solve logical puzzles, and do whatever AI can, but better.

If you're already looking for a new profession, preparing for machine domination, relax: we still have a couple of years since the model is still far from being perfect, even in OpenAI's promo videos, which are meant to show its best angles.

Ian Fisch, one of the developers behind Kingmakers, a medieval shooter with "next-gen multi-threaded AI" where you destroy your sword-wielding enemies using a variety of guns, analyzed o1's ability to create video games and concluded that it's not ready to replace human effort yet.

In the original video, OpenAI o1 was asked to write some code for a game called Squirrel Finder. The idea is simple: you play as a koala, and the game spawns strawberries and a squirrel. You need to "catch" the squirrel but avoid strawberries.

Strawberries appear every second, but the squirrel materializes after 3 seconds. Right at the first second, OpenAI's researcher Jason Wei gets the squirrel, thus winning the game without even moving the koala.

This, Fisch notes, happens because the AI counts the time the player spends on the instruction menu, so when the screen is gone, 3 seconds have passed, and the squirrel is already there.

Moreover, the machine lacks awareness of what's fair and fun and can slap strawberries and squirrels right on top of the player from the get-go. If it was made by a human, they would naturally specify that these objects should start at a distance from the koala. 

Fisch concludes that the biggest flaw of OpenAI's video is that the researcher doesn't ask o1 to fix all these issues because, he thinks, it can't fix them. Instead of changing one little thing, AI usually rewrites the whole code, often creating new problems.

Something similar happened with Devin, the "first AI software engineer," which took but ultimately failed an Upwork job.

If you want to see how OpenAI o1 works for yourself, be careful: you might receive a warning message from OpenAI.

Watch Fisch's breakdown here, take comfort in knowing human professionals are still needed, and join our 80 Level Talent platform and our Telegram channel, follow us on Instagram, Twitter, LinkedIn, TikTok, and Reddit, where we share breakdowns, the latest news, awesome artworks, and more.