overactor

Oh, I didn't see your post. It's wild how much difference timing and slightly different wording can make. I couldn't have predicted which one of these posts would take off honestly.

Making accurate predictions in a specific domain doesn't imply understanding, sure. Universal autocomplete can make accurate predictions in all domains. Apples to oranges.

No human can solve the Riemann Hypothesis. Why do you expect an AI to do it before you consider it able to understand things?

I think this misses the point made by Hinton. The point is that, if a model is to reach a certain level of autocomplete, it needs to understand what it is talking about. Imagine the result of what required deep thought from you is included in the training data, along with billions of other such exchanges. At some point the only way to improve training score is to develop the ability to "understand" things. (That is, if the model isn't big enough to memorize the training data.)

Universal autocomplete implies understanding and the only ending question is if our current architectures and model sizes are able to achieve any level of it.

I think it's a little harsh to claim that the LLM didn't understand words. Sure it's far from perfect, but it mostly gives coherent answers. The AI is instructed to interpret each scenario as deadly, so it will typically do that, even if it doesn't make much sense.

That was meant as tongue in cheek comment. Sorry if that didn't come across.

Touché.

Yeah, the AI is far from deterministic.

That's definitely very worrying.

Also, I can't take any credit for creating the game. I'm just a fan sharing something cool.

You actually can use custom scenarios by clicking the pencil icon on the prompt selection screen.

I think I read that some companies are working on using LLMs to improve/generate dialogue from side characters.

I absolutely love that idea. The AI could absolutely pick a negative trait when the player survives. Maybe they gain a positive trait when they die? That way it might balance out a little.

That's an interesting way of thinking about it. The AI definitely always rejects any attempt to make nothing happen.

I can't promise you this, because I haven't seen the actual code. But in the discord one of the devs has alluded to a bug occuring when the AI for some reason doesn't decide if the player survives or not. I'm pretty sure the LLM decides if the player lives or dies, but it's definitely pretty random sometimes.

It's a bit hit and miss to be honest. You can throw some pretty wild stuff at it and often it reacts surprisingly well. On the other hand it also often crucially misunderstands things and generates pretty nonsequitur responses.

The good outweighs the bad in my opinion, but YMMV.

Yeah, accepting death and prayer are both sort of like cheat codes.

A strong gust of wind could definitely knock you off balance enough to lift your foot if it's unexpected. But yeah, it doesn't really follow from the plan. If you're looking for a fair and balanced experience, this game won't offer it.

For the record, I thought your plan was very creative and it does seem like the AI didn't really get it, which can be frustrating.

AFAIK AI players is on the roadmap, but the devs are mostly focusing on ironing out bugs right now.

You can pay alone, no problem. Though it's obviously more fun with people. The AI is pretty arbitrary and there are many ways to cheese the scenarios. I think that adds to the fun though and it's best played with people just trying to have fun and not super pressed by wanting to win.

The game is still in beta and a bit buggy, but it's a great application of AI where its weaknesses can actually turn into strengths as it makes for funny output.