Live data from Hacker News

I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

joshfonseca.com

121–130 of 198 posts

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#121
post #109

Earlier quoted context omitted.

[flagged]

Do you try trick your calculator by entering large sums that it might not have been programmed to answer?

I spent a long time doing this as a child :)

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#122
post #90

Here's the code: https://github.com/vuciv/animal-crossing-llm-mod I was intrigued as to how it would intercept a conversation and then pause the game for long enough for the LLM to return a response, so I used https://gitingest.com/vuciv/animal-crossing-llm-mod to dump the 40,000 tokens into Claude Opus 4.1 and asked it: https://claude.ai/share/66c52dc8-9ebd-4db7-8159-8f694e06b381 The trick is the watch_dialogue() fu…

Wrote up a few more notes on my blog https://simonwillison.net/2025/Sep/10/animal-crossing-llm/

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#123

Earlier quoted context omitted.

Technical question: what guarantees that the memory address is stable? It means that allocations always happen in the same order in the game ?

It needs to be a global variable in C. A variable in a function scope or runtime malloc'd spot? No chance of finding a stable spot. Thankfully, a lot of old games love to use global variables because you can't run out of stack space or allocatable memory. Modern games shy away from that because even the tiniest consoles these days come with gigabytes worth of RAM, even a memory leak has to be a gigantic firehose to b…

> A variable in a function scope or runtime malloc'd spot? No chance of finding a stable spot.

It depends, on a system with no ASLR it's potentially still deterministic, but there are a lot of potential sources of non-determinism.

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#124

This is awesome. LLM-powered NPCs is one thing I'm most excited about in the future of gaming. Characters repeating the same scripted dialog over and over again is one of the biggest immersion breakers.

I'm unsure how useful they'll really be honestly! In many situations its very useful to know when your "done" talking with an NPC when they start repeating lines etc... There could probably be cool uses but I don't think it will be a pure "upgrade" as the repeating dialog is kind of a feature honestly. We'll have to see how it pans out xD

Feeling it would work best in more of a Dwarf Fortress approach game where it's more you have a sandbox with rules that cause a simulation to have emergent gameplay.

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#125

This is awesome. LLM-powered NPCs is one thing I'm most excited about in the future of gaming. Characters repeating the same scripted dialog over and over again is one of the biggest immersion breakers.

I spent a lot of time on ML driven dialogue and natural language understanding for games in the pre-llm days working with some folks building Xbox, PC, and VR/AR games, and I'm excited about the potential here. The major barriers to shipping in a production game though are still the same imho:

1. (predictability) Games like to have a clear arc and tend to use at least some of their NPCs to move that forward, it is harder to do this with a model that could make a choice you don't predict. They tend to have a set of items, quests, what-have-you that you need boundaries around.

2. (testing) Games like to test like crazy before launch, at least the AAA ones, so their QA folks just don't like a model that can have infinite responses/variants. Many then drop to a skeletal crew for maintenance and improvements after launch, where with ML models you actually need to keep improving the model, finding long tail bugs as more players interact with the system, etc.

3. (cost) Games are usually very cost aware, it's far cheaper to just have a set human-written dialogue path, then to run a model, even an offline one. Cheaper in both actual dollar costs if you're talking about a high end LLM service call, and CPU/GPU/memory costs if you're talking an on-box system.

4. (internationalization/localization) AAA games need to launch fast to many languages and locales, using a model for NLP and dialogue management/natural language generation adds added testing costs for each new language, that is just a very cheap translation normally that can be outsourced.

There have been some fun experiments in this space, and I expect to see this improve and become common use in the future, but it will take time and more work on how best to integrate a model into the flow of a game. I do love it for "presence" so talking to NPCs feels more human-like.

Nick and the crew at AI Dungeon (and related) have always done some interesting work in this space, trying out games where AI can be used in interesting ways.

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#126
post #104

Earlier quoted context omitted.

I guess you could just introduce a Symbol that marks the NPC as "I said everything I have" or gray out their text, or some other visual marker

Repeating the last line of dialogue is not just a way to indicate that there's no more dialog, it often also works as a remainder, giving you the most important kernel of information ("You should go to [place] and talk with [npc]"), in case you come some time later and forgot what you were supposed to do. You can indicate there's no more dialog in many ways, but you'd lose that secondary feature. Same thing if the NP…

So true. In such an LLM-driven game though, I would imagine the player would just ask the NPC: "I forgot what to do" or even "Can you explain it in other terms?" (if the quest description isn't clear enough).

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#127
post #81

Earlier quoted context omitted.

How can you tell? Either you're saying that LLMs have no "thoughts", and just regurgitate everything, which doesn't explain how they can reason in novel problems, or you're saying you can tell what's original thought and what's regurgitation, in which case I'll ask "how do we know your comment isn't you regurgitating an HN opinion"?

Generative AI can reason in novel problems in an analogous way that bands continue to make new melodies even after all these years using the same instruments as all the other bands in their genre.

I do agree, in about the same way as people.

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#128

Earlier quoted context omitted.

I'm unsure how useful they'll really be honestly! In many situations its very useful to know when your "done" talking with an NPC when they start repeating lines etc... There could probably be cool uses but I don't think it will be a pure "upgrade" as the repeating dialog is kind of a feature honestly. We'll have to see how it pans out xD

> In many situations its very useful to know when your "done" talking with an NPC when they start repeating lines etc... But that's not how real life works at all, right? You talk to someone for as long as you want to talk to them, or until they start sending signals that they are done talking with you . The way video game dialog works has always bothered me, it makes characters feel stilted and makes me care less ab…

"But that's not how real life works at all, right?"

How real life works is always a plausible interesting goal, but it's very often at odds with a bunch of other valuable goals for players.

A particular sharp example of this is sports video games. It might well be interesting (and certainly realistic) to simulate bad referees in a sports game. Horrible blown calls by tennis line judges, or missed calls by basketball refs, or bad umpire calls on pitches. Real-life soccer makes working the refs and their inability to see everything an art form, as far as I can tell.

Perhaps that's interesting, but the irony here is that real life refs are actually bad simulations of the original perfect game code in the first place, from a certain point of view. I think debates about the use of instant replay in sports gets at the heart of this, and one could imagine using real-time AI to help refs taking this conversation much further.

I think the sports case is a particularly sharp example, but it definitely holds with all sorts of choices in games.

For Animal Crossing in particular, I remember when I finally played it, it struck me after a while how much it had in common with recent MMOs (Everquest and World of Warcraft) that I had had fellow game developer friends have their lives severely disrupted by. And when I played the original Animal Crossing, I remember noticing specifically how careful the designers were in having players use up every bit of interesting content in a day after 45 minutes or an hour, so that eventually you'd run out of things to do, and that was the game's signal to put it down and pick it up again the next day. And I remember being struck by how intentional it was, and how humane it was... particularly given their goal of wanting to make a game that was asynchronously coop (where different family members could play in the same shared space at different times of day and interact asynchronously). As a game designer myself, I really respected the care they put into that.

Anyway, that's my immediate thought on seeing this (fascinating, valuable) experiment with LLM dialogue in Animal Crossing. The actual way NPCs work in these games as they are has been honed over time to serve a very specific function. It's very similar to personal testimonials by paid actors in commercials; a human expressing an idea in personal dialogue form triggers all sorts natural human attention and reception in us as audience members, and so it's a lot more sticky... but getting across the information quickly and concisely is still the primary point. Even dialogue trees in games are often not used because of their inefficiency.

I totally think that there will be fascinating innovations from the current crop of AI in games, and I'm really looking forward to seeing and trying them. I just think it's unlikely they will be drop-in replacements for a lot of the techniques that game developers have already honed for cases like informational NPC dialogue.

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#129

This is awesome. LLM-powered NPCs is one thing I'm most excited about in the future of gaming. Characters repeating the same scripted dialog over and over again is one of the biggest immersion breakers.

IMO this isn't going to be fun for most once the novelty wears off. Games are goal oriented at the end of the day and the great games are masterfully curated multi-disciplinary experiences.

I'd argue throwing a game wrapper around an LLM is a new LLM experience, not a new game experience.

Re: I replaced Animal Crossing's dialogue with a live LLM by hacking GameCube memory

#130
post #109

Earlier quoted context omitted.

[flagged]

Do you try trick your calculator by entering large sums that it might not have been programmed to answer?

Sneaky sneaky natural logs and that precocious savant, Euler!
Post reply on HN