GPT-4 Destroys Everyone on Its Way in Doom
It has some ADHD problems, though.
Image credit: Adrian de Wynter
By now, you must know that anything can run Doom, but the tech world has a new fascination: making AI play video games. Of course, this old classic is the prime specimen for such experiments.
As spotted by The Register, researcher Adrian de Wynter from the University of York managed to get GPT-4 to play Doom with a few instructions and a textual description generated by it from screenshots, and his findings are fascinating.
As you might know, GPT-4 can't play the game, but its variant, GPT-4V, which recognizes images, can and doesn't even need training. As mentioned before, the LLM takes screenshots and provides descriptions and instructions to successfully beat Doom. It works but is pretty slow. More importantly, it seems to suffer from the same ADHD problem I do: when it doesn't see the enemy, the zombie stops existing.
"For example, it would be very common for the model to see a zombie on the screen, and start firing at it until it hit it (or died)," said de Wynter. "Now, this is AI written to work with 1993 hardware, so I'm going to guess it doesn't have a super deep decision tree. So the zombie shoots at you and then starts running around the room.
"What's the issue here? Well, first that the zombie goes out of view. Worse, it is still alive and will whack you at some point. So you gotta go after it, right? After all, in Doom, it's whack or be whacked."
So it operates on the "out of sight, out of mind" principle and just keeps walking, which is not ideal. However, when the AI does see a zombie, it goes all out. It's actually pretty concerning how bloodthirsty the model is: it eliminates everyone in its sight without any second thoughts.
While GPT-4's reasoning is not very good and it experiences hallucinations, or provides wrong information, it can play Doom:
"We find that GPT-4 can play the game to a passable degree: it is able to manipulate doors, combat enemies, and perform pathing. More complex prompting strategies involving multiple model calls provide better results."
De Wynter could make the AI destroy enemies quite easily, which should "raise some alarms around possible misuse." What if it is programmed to eliminate fleshier, more real targets? Hopefully, hopping behind its back will work then as well.
Sounds scary, but if you're still interested in AI playing games, check out OpenAI's neural network trying its hand at Minecraft. Now imagine if we had a machine companion to play together – it might be more possible than you think.
Read de Wynter's research here and join our 80 Level Talent platform and our Telegram channel, follow us on Instagram, Twitter, and LinkedIn, where we share breakdowns, the latest news, awesome artworks, and more.
Keep reading
You may find these articles interesting
