Hacker News Reader

For what it's worth telling GPT-4 you want it to simulate a text adventure game that lets the user enter their actions and dialogue, and pasting the setting and some characters works much better than AI Dungeon did when I tried it last year.

GPT-4's context size is much larger, and when it starts to misbehave, you can just remind it how the game works and it's good to go again.

I'm sorry or you're welcome.

While I agree that violence in videogames is the "easy way out", the default mode of interaction that unimaginative designers reach for, I disagree this is because of limitations in AI.

Game design has a lot of tools in its toolbox. "Emerging play", "intelligent actors", etc, are only some, but not even the most important factors for interesting gameplay that is nonviolent.

AI is not the be all, end all of game design. Not even close to it.

> While I agree that violence in videogames is the "easy way out", the default mode of interaction that unimaginative designers reach for, I disagree this is because of limitations in AI.

You have an obstacle in your way. You use your skills to aim the crosshairs correctly and pull the trigger at the correct time. If successful, you are rewarded with the obstacle being removed. Then the loop repeats for just long enough so that you don't feel fatigued. It's a power fantasy and satisfying when balanced so that it's just within your reach.

Mapping it to guns, combat and violence is definitely the easy way to do things. Also one that sells a lot.

If you think about it, though, a similar loop exists pretty much anywhere there is a challenge in a game, violent or otherwise. You can turn that loop into a social skills challenge, but it's generally easy to see through those to the rules behind them, making them feel mechanical and hollow. That's one area where AI may be able to help, it's hard to tell.

This is an interesting conversation!

I agree with your characterization of the mechanisms for violence in videogames, and I also agree that if you abstract them a bit, they generalize beyond just violence.

Where I disagree is in your assumption that there is a similar loop "anywhere there is a challenge in a game". Deduction-based games aren't like this (e.g. Return of Obra Dinn). Puzzles aren't like this. Boardgame-like videogames aren't like this. Traditional point-and-click adventures aren't like this (they suffer from other problems though, beyond the scope of this conversation). Portal 1/2 are not like this in its core gameplay elements. Story-heavy, on-rails games aren't like this (though only because there is barely any gameplay to them -- and let me be clear that I think "movie-like" games are a game design dead end).

I do enjoy mindless fun and I'm not against violent games; I just think they are the easy way out for game designers. But I think ultimately making a "good" game goes beyond replacing shooting with some magic silver bullet (AI or whatever): ultimately you need good game design, there's no escaping this.

I don't think it's a coincidence the games I find more engaging are those that break the mold somehow, even when there's no "novel" gameplay element to them. Molleindustria's "Every Day the Same Dream" moved me profoundly, but there's no real tech breakthrough in it. It's just a bog standard Flash game.

I do think AI could be successfully employed in text adventures, or games where conversation with an "agent" was a core gameplay element. Imagine being able to Voigt-Kampf a replicant in an hypothetical Blade Runner game! (no, not Westwood's, which had a sucky VK implementation).

So the way this framework works with puzzle games is similar but slightly different. The important thing is to think about it in terms of what player skills are being tested to overcome a problem.

Instead of being a challenge to players’ reflexes and coordination - physical input on the keyboard/pad - it’s a mental challenge instead - exploration, observation, reasoning, etc. In this case the physical input is trivial but the challenge loop has the same shape.

In the case of Obra Dinn, each time you add something correctly to your journal you’re completing that same challenge loop. In an abstract, theoretical way, of course, but I find it useful (I am a game designer). Other games blend physical and mental skills together and of course there are many different skills for different games to test.

You’re probably right in that this doesn’t apply to every kind of gameplay, but most games drive you forward by putting some kind of challenge in your path, even if they’re not violent. I don’t know if that’s because it’s easiest, if it’s a convention from the early days of video games that has stuck, or if it’s a fundamental part of human nature.

Well, you need a challenge for it to be a game. Something to "play with". I suppose only aimless sandbox games escape this.

I don't think simply saying "this has an obstacle to overcome before you move to the next obstacle" is enough of a similarity. The gameplay loop of Counter Strike is nothing at all like the one in Obra Dinn or Return to Monkey Island. I don't see any similarity beyond "you need to solve a problem in order to continue". CS is like an e-sport; Obra Dinn and Return to Monkey Island are like theater plays with puzzles. One is made for replaying because you get better with experience; the others are meant to be played once, because once you solve the puzzle that's all there is to it. And so on.

Which skills are getting tested also makes a world of difference. If something tests my reflexes, it's radically different from something that tests my logical thinking. It's why gamers are drawn to some games but reject others!

In my opinion, all they have in common is that they are games.

> I suppose only aimless sandbox games escape this.

Flow theory suggests that players will self-direct to find challenges of their own making. A sandbox game needs to provide enough opportunities for that and make it suitably rewarding.

> The gameplay loop of Counter Strike is nothing at all like the one in Obra Dinn or Return to Monkey Island.

No game has one single gameplay loop, they're all meshes of smaller loops, but as you go deeper you always(?) end up seeing the same structure, almost fractally. The reason that I believe this model applies to both Counterstrike and for Monkey Island is that you can understand the difficulty progression for both games in the same way.

If the test of skill for aiming and shooting is primarily being able to move the left stick precisely and pull the right trigger at the correct moment, then I know that the challenge will be harder if there is more precision required (make the range of valid x/y position values returned by the stick smaller, e.g. by reducing the size of the enemy on screen - distance or physical size); or by reducing the timing window in a similar way.

In a linear shooter game you will use this (applied to every mechanic) to steadily increase difficulty throughout the game. In Counterstrike you can use it to balance and measure relative weapon effectiveness and also analyse why some positions are easier/harder than others.

In a puzzle game like Monkey Island (which predates all this theory) you provide the player an objective and put a number of complications in the way[1]. You can make a puzzle harder by adjusting the number of complications in the way. You can also adjust the difficulty of each complication. If one puzzle type requires observation skills you know you can make such a puzzle harder by making the object harder to see on screen: size, shape colour, animation, etc. If there is a memory test component, you make the puzzle harder by increasing the number of things you need to keep in working memory at one time.

You can then reuse those puzzle types again and again, controlling the difficulty of the challenges over time by increasing the difficulty of individual skill parameters and combining them with new ones. The same thing you would do in a linear shooter. The Witness is a perfect example of this.

This is not my own personal theory by the way. It's also not the only way of looking at game design, but it's a practcal method I've seen all over the AAA space, at least. It's not strictly formalised across the industry but Ubisoft internally called it Rational Game Design[2].

[1] https://www.youtube.com/watch?v=zsjC6fa_YBg [2] https://www.youtube.com/watch?v=xqHAjwrNp70

Exactly! I'm excited for NPC to act like AI, and have full fledged conversations. Now, if they're allowed to have previous memories of their interactions of not could be interesting, since that is currently banned by ClosedAI

I fear how buggy that could make future games.

Some essential NPC interaction gets stuck in a weird loop / starts speaking Chinese / gets triggered into a Tay like paranoid Nazi rant while adamently denying the existence of the key it is holding that it's meant to give to the player to progress.

At worst, it's going to lead to a "is this a good NPC response" pipeline, where the model spits out lines based on its given prompt, and each one is approved by a proofreader.

The minutes of time spending on thinking up what an NPC would say is now (about 10 seconds of speedreading 5 lines of generated results + 5 seconds in clicking either the 'approve' or 'reject' button): That in of itself is at least a 10x increase in available topic queries for NPCs.

> AI Dungeon, reveal that there will soon be new methods of maintaining a coherent story in a simulated game.

Somewhat coherent, for a few paragraphs. Still impressive.