LLMs in Games - Infinite Stories, or Infinite Hype? - Chris Simon | grokludo 13

Chris Simon is a technologist who's given talks about AI – specifically LLMs, or Large Language Models such as ChatGPT. He's best known for calling it a hype-fueled dumpster fire.

Despite the spicy title of his talk above, his views around AI are quite nuanced. After researching the "LLM supply chain", including the process for how these models are trained and reinforced, he made an ethical decision to not engage with them at all – but he's still aware of the potential upsides, and charts a likely path forward after the hype bubble has died down.

"Sometimes people point out that every hype bubble we've had in the past has left behind an infrastructure layer that's served the future in a way that was not predictable," says Simon.

As such, he believes smaller, highly specialized models that run on your GPU are "a very probable outcome of this whole phase."

AI is going to be an increasingly big topic in games. From things like art and code generation, to chat moderation, to dynamic difficulty systems, and all the way to engineless games that use a neural network to generate images based on user input. And there are many more potential uses.

So it's a topic we'll probably return to, but one of the biggest ways people are using LLMs is in story generation and dialogue. The idea here is that one could chat to an NPC indefinitely, or let a robotic Dungeon Master do all the lore work.

Chris Simon says it's not so simple, and it's easier to see the problem when you analyze LLM outputs at scale.

Referencing a conference organizer who saw thousands of speaking submissions, Simon says "The creativity is not actually there. Because if you ask 2,000 people, you get 2,000 submissions. You ask the LLM 2,000 times, you get about five submissions with minor variations."

Simon goes into the many inherent biases that the larger models have, as a result of scooping up all the text on the internet (including its rough edges), as well as all of literature before the civil rights movement.

While it's possible that the EAs, Activisions, and Ubisofts of the world could train specific models without these biases, it's more than a coding problem. It's about server farms, GPUs, and the hyper-competitive market for AI engineers.

I believe it's more likely these companies will reach out to incumbent enterprise services – meaning those inherent biases will make their way into games. As if on cue, EA just signed a deal with Stability.

While I can see grokludo returning to this topic, given its vast scope, I'm glad Chris Simon is the first person on grokludo to talk about AI. Because just as many games and NPCs will use the enterprise LLMs as a foundation, all our future AI chats will use this interview as a foundation, about the hidden costs of AI in games.

Thank you all for supporting grokludo!