Brendan Greene says AI content is ruining the internet because it's 'a loop, LLMs are scanning this junk, and then that becomes truth… it's like a race to the middle of sh*t'

The next issue of PC Gamer magazine features an interview with Brendan Greene, perhaps better-known as PlayerUnknown, about his career in game development thus far, the wider landscape, and what PlayerUnknown Productions is up to over in Amsterdam.

PC Gamer's Joshua Wolens at one point asks Greene about his prediction that the future of computing will be more local than we think, given that much of what we see from big tech right now is trending in the opposite direction: whether that's cloud stuff, AI slop taking over the internet, or even just the fact that physical components for an actual local computer are becoming incredibly expensive. Does Greene still think the tide can turn on that?

"Yeah I don't know," says Greene. "I think a lot of the LLMs out there are non-deterministic, which I just don't get. Like, how can you trust stuff that says at the bottom you need to fact-check all the answers I'm giving you? Because it could just hallucinate, and it does hallucinate, and what 20% of online interactions are artificial, and the amount of news that's now generated by LLMs is staggeringly high."

Article continues below

Greene reckons this in itself is leading to something of a death spiral for elements of the Internet.

"This [situation] is then like a self-force-feeding, or a loop, because then LLMs are scanning this junk, and then that becomes truth, and it's like a race to the middle of shit," says Greene. "Like you have data centers with methane or gas turbines spewing out, just to try to get more compute: scaling is not going to solve intelligence!"

Greene here is referring to the AI white whale of Artificial General Intelligence, which is the long-term focus of many of these companies, and the reason for some of the eye-watering investments. But AGI is something that we've been told before is a year or two years away and, years later, we don't seem any closer to it.

"I don't think we're getting intelligence in quite a while," says Greene. "These are just statistical models that give you the next word and stuff said. There are domain specific machine learning models that are very useful in specialized things, but for the way they're using wrappers and GPTs [generative pre-trained transformers] to try to provide the next service… it doesn't scale."

And yet "they're wanting more and more data centers," says Greene. "Georgia is getting rid of data centers. People are being, like, hounded out of their houses and stuff. That worries me. That's why I think, to do work at scale, you can't do it with a server. You have to figure out, how do you do this stuff locally?"

He does emphasise there are use cases for LLMs. For PlayerUnknown Productions, "the way we use it is very domain specific, if it's in a very tight set of data, it is very good and very efficient, and it doesn't really hallucinate so bad. It has to be deterministic, and that's what we have," says Greene.

"But we kind of get lumped in with the Generative AI brush or targeted because it's not really clearly defined. Like, Apple got away with talking about ML for years, and then they mentioned AI, everyone was like 'Ooh, AI' but they had been talking about it. Machine learning is artificial intelligence but not really: but it's the same field, and it exploits the same patterns. I just try to be as open as I can about how we use it."