Regarding debunking the skooma merchant murder anecdote:
> The addicts live in a locked cabin, so it’s unlikely for the player to enter it unless they are specifically looking for it.
This is overlooking a crucial, obscure, and unintentionally hilarious detail: not all the skooma addicts are in the cabin! Out in the world are two NPCs who make a monthly inter-city trip to the den to get their fix. However, due to a bug where these NPCs are assigned to the wrong faction, they can't actually get through the locked door of the den, so they'll stand outside the door drinking skooma forever, unable to progress to the step of the AI package that would eventually return them home to their usual schedules, unless the player unlocks the door for them. https://en.uesp.net/wiki/Oblivion:Trenus_Duronius
Interesting detail, thanks for letting me now. I had a look at the AI packages of all three visitors (Gelephor, Gellius Terentius, Trenus Duronius) and at least in the base game (without UOP) none of them carry skooma nor are scripted to find it. So even though the game implies they are skooma addicts via dialogue & environmental storytelling, from a purely technical POV they are not addicts. Getting stuck outside the shack checks out, though I don't think faction membership is the reason for that — they simply don't have the key to the door.
The closest thing we got to the idea of Radiant AI is probably Dwarf Fortress.
But entirely goal-driven (and thus unpredictable) game AI systems like this are usually at odds with story-driven gameplay where the outcome needs to be deterministic (or at least "winnable") and the player is the hero which the story is built around (while games like Dwarf Fortress don't have a pre-defined story, and also no player character to take care of, and the whole fortress being wiped out because of comically unpredictable events is a large part of the "fun").
Another similar game to Dwarf Fortres is Song of Syx [0]. It's more accessible then DF and I think they can have up to 20,000 entities active in the world at a time. The world map is pretty huge, and the player gets to control a one group among many. Every entity in Song of Syx is individually modelled, though probably not in quite the details that DF is known for.
That was also my thought. How does the world behave 100 hours into the simulation? If half the town residents have managed to get themselves killed by guards and some of the shopkeepers are gone, it's a bad outcome. Complex sims have emergent behaviors that are hard to tune.
The other thing is a bit more subtle. It's a big open world and all NPCs need to be active continuously for that sim to work. So you have a big N to squeeze into a tight per frame CPU budget. Also, things like path planning or object interaction only work if some information like object positions and pathfinding maps are kept in memory the whole time for the entire world. This sounds very challenging on a 2005 era PC.
The thing about Oblivion that the simulation tends to run up against is that hitpoints and death are an abstraction: a real human would die much more easily from the injuries that get inflicted, but also a real human would avoid a lot of problems in the first place (and have families that would take over their shop, and live in cities with more than twenty people hanging around). You run into trouble when one part of the simulation is taking things as symbolic while another part is taking it as literal. If you want it all to be literal you've got to be willing to go super deep into the emergent simulation.
In time since Oblivion we got games like Divinity: Original sin 1/2 where you can kill pretty much every character in the game and it will still be finishable.
The essential NPCs could also be flagged essential, or maybe have a variation of that flag where only way given character dies is if say 1/4 of the damage dealt to character is from player (so NPC can't accidentally kill important NPC basically).
Also, radiant AI can also just... not run on the plot significant NPCs.
Finally, Bethesda games aren't known from main story being the main selling point.
I think it's more than essential NPCs though. Already in Oblivion those couldn't die anyway (Morrowind was the last TES game where you could get locked out of finishing the main story if you killed the wrong NPC).
But fully emergent behavior would likely destroy some player's experience in other ways - towns without shopkeeps, most quests ruined, little staged moments going away, etc.
Some games in the Ultima series did, but Morrowind didn't, which is why Radiant AI was developed in the first place. The first chapter of the article is about that.
There's probably some mathematical way to express that... it'd be interesting to look at Todd's mythical "Radiant Economy", create a dynamical system model/game-theoretic mode, and try to prove that in the long run everyone doesn't end up broke or a millionaire.
I think Veloren has a sort of dynamic economy where NPCs trade in and consume goods. Well, maybe not NPCs, but at least settlements as a collective, or something like that. I'm unsure of the details, but I remember prices being different between settlements, and prices changing based on local NPC inventories.
> create a dynamical system model/game-theoretic mode, and try to prove that in the long run everyone doesn't end up broke or a millionaire.
Simply ask yourself which factors in the real world lead or don't lead (depending on your political stance) to this outcome, and you likely have found the relevant factors that you have to include.
To be fair, the velocity of money can be be significantly higher in a video game, and you're much less likely to have innovations reshuffling the market. It seems inevitable that extremal states would be more prevalent than real life.
With the point "the velocity of money can be be significantly higher in a video game" you actually outlined a serious problem (and a potential solution):
The rest of the in-game economy (including its pricing) doesn't fit the money circulation velocity, thus we get problems.
After playing Starfield I don't really have any expectations for Bethesda to deliver on anything interesting anymore. The progression from Oblivion to Starfield has been one of becoming less like a small shop with character willing for its developers to take big risks with unique and intricate features, and more of trying to be a generic AAA studio that prefers predictable blandness. I don't think you can really hope that they'll magically return to making games the way they did 20 years ago.
They seem to now be under the mistaken impression that radiant AI is to get more content out of the game by implementing infinite permutations of simple quests, and that customers will think X things permuted Y ways is X*Y content and not just X+Y content. But the purpose of radiant AI was, I think, to make the world feel alive and even unique. Which means I really shouldn't even see every x in X or every y in Y.
If you want an interesting implementation of the same concepts as in Radiant AI I recommend checking out Dwarf Fortress. Every dwarf fortress world is essentially an entire history of thousands of radiant AI interactions up until you enter it, at which point your adventurer/fort becomes part of the world and continues the radiant interactions with its civilizations/wildlife/monsters/etc.
I think DF is probably the ideal existing game to considering adding LLM-characters and conversation to as a drop-in augmentation. DF already has the simulation and generation of realistic characters and stories working, but unfortunately it's very formulaic to interact with it as an adventurer. In that case the game actually is quite "alive" already, just without a voice.
To me, Starfield is a massive admission that either the developers don't understand what made their previous games work - or that no one will step in at a top level and prevent them breaking that core.
The modern TES games have been all about environmental storytelling, exploration, combat and crafting. All else is secondary.
Whether you like that focus or not is up to you, but that's the draw of games like Skyrim and FO4.
But Starfield completely broke it. They wanted hundreds of planets to explore, but the only practical way to do that is procedural generation. No one wants to explore procedural spaces. There's never anything interesting in them. You can't do environmental storytelling because that requires a human hand.
Due to engine limitations, making all the travel seamless was also completely impractical.
So instead of seeing a location and going "Oh man, I want to go there!", then just walking there encountering distractions on the way...it's Loading screen -> Loading Screen -> Loading screen -> Generic planet with nothing interesting to see.
How no one at a top level said "this can't work, the game's concept is bad, start over" is baffling. No one had a vision at the top level for how the game was supposed to work - or that vision was just wrong.
If Bethesda can't understand the fundamentals of their own best-selling game, I don't see how they can make a sequel.
> No one wants to explore procedural spaces. There's never anything interesting in them. You can't do environmental storytelling because that requires a human hand.
I would say, rather, that no one wants to invest the development effort to make them interesting enough to explore.
In my view, you can either use procgen to make development cheaper or to make it more interesting to explore, but not both at the same time. The roguelike genre was invented because the developers of rogue wanted to be surprised by their own game. And it worked to an astonishing degree.
But you've got to design in the systems that are interesting to explore, rather than relying on the amount of content.
Everyone hopes that you'll have multiplicative results so that content X times content Y goes exponential. But with procgen the multiplicative effects are more from different systems interacting; having a sword with different stats feels same-y, having a sword that combines two gameplay effects starts feeling more interesting, having a sword that integrates with a procedurally generated narrative and a system of tracking per-weapon kills that dictates your reputation among monsters starts feeling like there's a lot more to explore.
Nethack is famous for having a zillion different hidden reactions that let different parts of the content work together in surprising ways, as anyone who has tripped down the stairs while wielding the corpse of a cockatrice has discovered. Dwarf Fortress has a zillion different moving parts, so that the giant shambling golem built out of salt can be defeated by shoving it into a lake. Caves of Qud lets you bring a chair to life and then use your psychic powers to swap minds with it and then go on to play the rest of the game as the chair (with rocket launchers).
They've all got a lot of interesting environmental storytelling, but in absence of the human scripting have to work a lot harder for it. A lot of games, unfortunately, stop at the X+Y generation, without building in the synergies to make the different values of Y unique and expressive enough for the players to care.
> No one wants to explore procedural spaces. There's never anything interesting in them.
I think practice shows this isn't true, Minecraft is pure procedural generation and people love wandering through Minecraft worlds.
I haven't played Starfield, but based on what you said the difference is in the complexity and amount of stuff, in Minecraft you don't have to go far to find more new things. Even if you're very familiar with the game you can still come across very unique areas, it's rare that I feel like I wasted my time by just wandering around a map.
Additionally Minecraft solves the story problem by simply not having one, which works fine for the kind of game Minecraft is, probably not so much for Starfield.
Starfield (and I speak as someone who put in a couple of hundred hours to the game) has a wealth of problems, but one of them is that they messed up the distribution of even the small-ish number of points of interest that they have.
They put some sort of cooldown timer on them, set way too short, so players see the same half dozen over and over again.
A modder discovered the timer and set it longer, and suddenly found a load more content that very few people had come across before.
Bethesda had been doing procedural generation since forever though: Have you played Daggerfall? It's always been part of their studio's DNA.
Bethesda has always relied on top of the line technological innovation that makes us forgive all the jank that came with it. Whether it was a bad combat system, a level scaling mechanism that just doesn't work, uncanny graphics... this has always been there. It's the opposite of the old Nintendo Way, where the games always were less ambitious, but had so much polish that the games counted as mirrors.
We've reached a moment of much diminished returns though. 5, or even 10 year old games aren't so technologically inferior that they are uncomfortable. A very shiny things has more trouble covering for jank, and high budget games are just so expensive that neither coherent vision. nor significant innovation are likely. So the Bethesda way is just not workable anymore.
What I'd want Bethesda to do, Breath of the Wild/Tears of the Kingdom meets Morrowind/Oblivion, is just really hard to wrangle logistically. Getting anything done under those kinds of constraings just takes too long.
Procedural generation is fine. But you can definitely see that Starfield was intended to be a platform for user-generated content straight from the start, and I think they must have convinced themselves that they didn’t really need to care too much about the game itself, because those chumps - sorry, players - would add all the content for them on their own. It’s like Metaverse all over again. They forgot they actually needed to make something worth playing and users’ time investment before it would become a money printing machine. Also, probably like 4 people who worked on Daggerfall still work at Bethesda and most of their games between then and now didn’t use procedural generation much at all, so I don’t understand why so many people make this argument. Like oh it’s normal for me to put DSLs in my software projects, here check out this git repo I worked on 25 years ago when I was in college, our customers should have been prepared for the shit job I did with it this time.
Actually, I think I would be completely fine with Bethesda just churning out TES POI and storylines without trying to do anything significantly more complicated than what they did in Skyrim. Just focus on the world building and the story and do some simple gimmick that’s a little more creative than “shouts/dragonborn but in space”. I suspect most other players would be happy with something of similar scope.
> I think they must have convinced themselves that they didn’t really need to care too much about the game itself, because those chumps - sorry, players - would add all the content for them on their own
Were they wrong? Skyrim sits at 70k mods after who knows how many years. Starfield has 10k already. I’ll admit it might not go as far as Skyrim, but still.
I feel the fact they did procgen is not as bad as the fact that what was not was just slightly less compelling than usual.
That is a major issue with Starfield, but it also felt like Bethesda missed the improvements happening in other games in the last decade or so. Many games now are much more cinematic in their storytelling, often with full motion capture. A very recent comparison would be BG3, which is very cinematic despite being almost impossibly large.
In Starfield you have a mostly static view of your questgivers talking. Which was fine 10-20 years ago, doesn't feel as engaging today when many games do it much better.
It's also not only about this aspect, you can make engaging stories with old-school methods. But the writing could not save the aging presentation here, it appeared very bland and tired to me.
What absolutely didn't help was the persuasion minigame, where you essentially broke all pretense of having a story-based reason to bypass a certain check. Persuasion checks are very common in RPGs, I've never seen them done so terribly as in Starfield.
The environmental storytelling certainly was the highlight of previous Bethesda games. But the main and side stories often were engaging as well. In Starfield they felt aggressively bland and mediocre in a way I haven't really seen in other games.
They've always been terrible at animation. 10 or 20 years ago, their animations have always been the absolute worst by any contemporary standard (the art used to be too - see Battlespire for some terrible art - but they improved it). Maybe part of that was because of the engine, but I think they just never had the culture for it.
They clearly did try to improve their animations in Fallout 4 in 2013-2014, which is the timeframe the most development happened, so it's not like they're oblivious to their biggest shortcoming as a studio. So what they did in F76 and Starfield is just a regression.
Proc gen can be engaging if emergent content is complex, dynamic, and novel. But again that goes back to Radiant AI being a vessel for generic fetch quests in the newer games, while in a proc gen game you would think there would be a major, if not the major dev focus on fleshing out the system in other ways (from dynamic tribes and factions to more fully fleshed out STALKER-esque persistent fellow space travellers with agency). The final missing component would be inspiration in design of the pieces, so they interact together in interesting but emergent ways, which is of course another element that the game sorely lacks.
I recognize all of that as true to some extent, but still I have 230 hours in Starfield, and I haven’t even finished all the quests.
Does that truly constitute a failed game?
As far as I’m concerned their biggest mistake was not having something to travel around in the planets on the start. Walking around to the interesting locations was annoying.
Then there’s a bunch of pointless systems like the colony system, and the whole space magic thing, but the rest is still a bog standard Bethesda game with 10000 different handcrafted unique locations for me to explore following a bunch of sort of interesting questlines.
Just because you enjoy it doesn't mean the game doesn't have problems, or it didn't have a negative affect on BGS's image. The reviews vary location to location, but on Steam it's at 55%. That's unfathomably bad for a Bethesda open-world RPG title.
A lot of people bought Starfield because it was a Bethesda game. A lot of those people will re-consider the next time such a game comes out.
Even years later, people are willing to put up with all of Skyrim's jank, bugs, performance problems, terrible animations and visuals, bad story and the rest because the core gameplay loop of exploration is so strong. It carries the entire game.
Starfield is missing that core that holds it together.
> Just because you enjoy it doesn't mean the game doesn't have problems, or it didn't have a negative affect on BGS's image
All of their games have a negative effect on BGS’s image. There’s no company that has more people complaining about their games. It’s going to take a whole lot more than a single terrible game to get people to stop buying them.
I’ve played fewer hours but gone through the whole story and about eight trips through the unity.
I don’t think it’s a failed game so much as one that’s not lived up to what it could be. The story is occasionally great but not always so. Most of the procedurally generated planets are entirely pointless and dull. You can see where they abandoned and downscaled ideas because there are still rough edges - the ‘fuel’ system that never was, for instance.
Overall I enjoyed the game, but it definitely falls into the same “banality of the infinite” trap that No Man’s Sky does
Yes. IMO Starfield's biggest failure is in the creative department. It is not interesting at all (for me) in terms of things like writing and voice acting etc. It is not a technical problem that can solved by innovative game mechanics like a roided up version of radiant AI (whatever that is).
Games like RDR2 and Witcher 3 left such a mark on me becauss they had bold personalities. Starfield in comparison feels like corporate memphis despite a nice Nasapunk foundation.
The funny thing about using AI to create an infinite amount of bland quests is that there is literally no audience for it. The people who play the game through once or twice aren't going to care about it and the people who want more of the game will download one of the thousands of mods created by the community. Oh, wow, you used AI to come up with a quest where I have to go to a cave and kill a creature. Amazing.
I think Starfield gets a lot more flak than it deserves. Yeah, compared to Fallout 4, where there's something hand placed to observe or interact with seemingly every 100 ft in any direction, the world feels barren. But I think the departure is intentional; Starfield felt much more like a spiritual successor to Daggerfall than to anything since Morrowind. Overall, I spent less time in Starfield than in older Bethesda titles, but I liked what was there, despite it being less dense, and I spent more time than I have in many other games.
Why should Bethesda have to refine the same exact formula over and over? That would just turn into what Ubisoft does with Assassins Creed, pumping out soulless entry after entry into the franchise. In other words, Starfield was Bethesda taking a risk and trying to introduce unique features rather than releasing yet another another predictable "Bethesda RPG".
Perhaps Starfield was the most important Bethesda release. The animus toward Starfield will serve as an enormous signal/reminder to course correct away from this "unique feature." One can hope.
> Why should Bethesda have to refine the same exact formula over and over?
Because they are the only ones who can pull off that formula, and when they stray from it, they end up as just one mediocre title in a sea of similar mediocrity.
That's a fairly reductive take in my opinion. Oblivion made the game approachable to a much wider audience. That may have been to the detriment of the core gameplay but it wasn't a _clear_ step backward. Some things went backwards while others moved forward.
I'd be surprised if the majority is overwhelming since DF has sold a million copies on Steam so far. For comparison, Civilization V has sold about 10M.
RimWorld is only superficially comparable to Dwarf Fortress, if you want to talk about basic gameplay then sure, but what makes the latter special is the immensely complex world simulations and interactions going on in the background, RimWorld has absolutely nothing like that.
Games have a similar lifecycle to social scenes. Now and again, an amazing game comes along that captures the imagination of gamers. Usually it's made by really creative and innovative people with a clear vision and direction. Also these people usually have taste, which is a crucial element.
Their good taste attracts a bunch of early adopters, people with a finger on the pulse and who are eager to play and appreciate the game for what it is. But this interest attracts poseurs, people who play the game but just to say they are, to feel included and a part of something cool. There are far more poseurs than otherwise, so at this stage the scene can grow exponentially. This growth attracts vultures commoditize the scene in the form of penny pinching and "enshitification" through dark patterns. Monied interests strip out everything that made the game interesting and fun (because a good, fun experience isn't profitable), and then they milk it for everything it's worth until it's a dried corpse. These are the people who are driving the bad gameplay decisions and who aren't listening to the taste makers.
Usually in games this comes in the form of a pivot to MMOs. By that measure, TES died in 2014.
IMO this also applies to Final Fantasy (RIP 2010, plenty of new FF games but nothing that recaptured the magic of 6 and 7) and Warcraft (RIP 2004, no new warcraft games since).
Reminds me how more people play oldschool runescape than the newer version. Anytime jagex tries to implement some change to that game they poll the community forums. Seems to work alright for keeping people around.
As a Bethesda fan (spending 1000s of hours combined in Fallout and Skyrim), I enjoyed reading this post. Especially liked the use of creating your own NPC to test the various scenarios. I just now started playing the Oblivion remaster for the first time and I find that I am liking the NPC interactions / liveliness a lot more compared to their later titles.
The one item that stood out to me was:
"Todd’s mid-fight dagger acquisition
Verdict: Impossible in the final game unless scripted to do so"
I do not disagree with the verdict for the final build of the game but I recall observing something similar in Fallout 3. I had stashed a mini-nuke launcher and ammmo in the Megaton player home. Some sort of conflict transpired (do not remember what exactly, perhaps I provoked an NPC for fun), I witnessed one of the town-folk run into my player home (in its own cell) and come back out with my weapon. It is possible with 1000s of hours in Bethesda games I am just mishmashing memories together but I am pretty sure this is what prompted me to eventually download a player home mod (and eventually learn G.E.C.K. by "remastering" it).
It's an interesting anecdote, but from my understanding of the system that simply shouldn't be possible. Your house's interior cell isn't loaded into memory when you are outside in Megaton, so there's no way for the NPC to access your items. I think this fundamental limitation holds true for every version of the engine, from Morrowind to Starfield, but I'd be glad to be proven wrong with concrete evidence.
When I played gothic, I was in the wilderness and in the process of being killed by some beast. Completely unexpected, a core NPC (Lester?) joined the fight and slew it. It turns out, he makes a walk between 2 camps every day, and happened to be around just at the right time.
While already impressed by the AI, I was blown away by this l behaviour. He goes between 2 places that can't exist in RAM at the same time, and interacts with the world when it happens to pop into existence around him.
Oblivion has multiple NPCs with complex schedules that involve travel between cities. And yes, you can run into them on the road. Best example is the countess of Leyawin, once a month she visits her mother in Chorrol (opposite side of the map) along with her personal guards and advisor.
Radiant AI does work exactly like that. The game keeps the global cell-level pathfinding graph in memory at all times, and uses it to simulate NPC travel outside of the loaded area.
Technology wise they should be comparable, but the way its used feels different. The Gothic world felt much more alive than Oblivion, even if Gothic is from 2001 and Oblivion from 2006
Interesting read. Got me thinking, I’d love to see what happens when modern AI meets open world simulation. Not just prettier graphics, but actual reasoning NPCs. Imagine arguing with a World of Warcraft innkeeper about the price of ale. Priceless.
Wiring a chatbot to dialogue is less interesting to me than the possibility of AI directing scenes and orchestrating reactivity across multiple characters. A reasoning model can ensure that the world responds to the player in a reasonable and narratively interesting way, without having to script everything or make individual characters particularly intelligent.
We're used to thinking of game AI as a property of the entity it's attached to (the NPC, the enemy, the opposing player) but an LLM can sit above that, more like a dungeon master.
Wasn't this the goal of the Director AI in Left 4 Dead?[1] Monitoring player progress (or lack of it) and tailoring how zombies and items spawned outside of script events, and in L4D2 how the map, pathing, and weather worked in order to maximize tension or encourage progress?
Not possible, because can't be guardrailed with 100% accuracy. You'll ask it something outside of the Warcraft world (e.g. US politics), and it'll happily oblige. I imagine NPCs will generate really weird immersion breaking stuff even if you cannot freeform interact with them anyway.
You can do that also while playing a traditional tabletop RPG. Players typically don't do it because why would they ruin immersion?
I understand that in multiplayer with strangers it would be a problem because you could affect other players' experiences, but in a single-player game I don't see this as a big issue, as long as the NPC doesn't spontaneously bring immersion-breaking topics into the conversation without the player starting it (which I suppose could be achieved with a suitable system prompt and some fine-tuning on in-lore text).
If it's the player that wants to troll the game and break immersion by "jailbreaking" the NPCs, it's on them, just like if they use a cheat code and make the game trivial.
It's still gonna be hallucinatory AI slop. For the same reasons it makes uninteresting quests and boring planets. It's lazy and it can't replace actual writing and art.
AI is great for getting tasks done where you can pull the information you need out of the slop. For quality immersive entertainment it's not there.
I’m not at all sure of this. You can use classifiers, fine tuning, and prompting to mitigate the issue both on user input and model output. And you’d probably want a bunch of fine tuning anyway to get their voice right.
> Not possible, because can't be guardrailed with 100% accuracy. You'll ask it something outside of the Warcraft world (e.g. US politics), and it'll happily oblige. I imagine NPCs will generate really weird immersion breaking stuff even if you cannot freeform interact with them anyway.
> Not to mention the current token cost.
You of course have to train the AI from ground up and on material that is as much as possible only related to the topics that are in the game world (i.e. don't include real-world events in the training data that has no implications in-universe).
You don't, for example, expect some ordinary farmer or tramp in the game world to know a lot about the (in-game) world or be capable of doing deep conversations about complicated topics.
So I don't think the necessary amount of text that you need to train the AI on is as insanely large as you imagine (but of course nevertheless a lot of texts have to be written - this is the price of having "much more dynamic" AI characters in the game).
Write a couple of lore books, in-universe cyclopedia, some character sheets and exclusively train on them. Maybe some out-of-game lore for cross-over universes!
The question that poses to me is the quantity of writing you need for training before you can reasonably expect a generation system to produce something new and interesting, however much work on the right knowledge is in the right place, and is worth the costs for how you expect the player to interact with the game beyond the manual work.
I doubt there's telemetry in the elder scrolls games, but I'd love to know how many go around the world exploring everything the characters have to say, or reading all the books. How many get the lore in secondary media, wikis or watching a retelling or summary on youtube. On a certain level it's important they're there as an opt-in method to convey the 'secondary' world lore to the player without a "sit down and listen" info dump, plus give the impression it was written by someone so these objects would would exist organically in the world or certain characters would talk about those topics, but I wonder how much of the illusion would still be there if it was just each book having a title.
For this to work you pretty much have to start from scratch, putting in "obvious" things like "the sun exists and when its out it casts light and shadow" and "water is a liquid (what's a liquid?) and flows downhill". Is there a corpus of information like this, but also free of facts that might be anachronistic in-universe?
Is that feasible? I was under the impression that fully training an LLM requires untold mountains of data, way more than a game dev company could reasonably create.
You are correct. The fact that so many people are saying “lol just train it on text about the game bro” reveals how little people understand how these models work, how they are trained, etc.
Microsoft's phi models are trained on a much smaller dataset. They generally aren't as amazing as the models that get talked about more, but they are more than enough to get the job done for npc lines in a game.
With the advent of unoptimized UE5 releases becoming the norm and the mentality of shipping badly broken games by default and them only being in a good state years later if at all, I’m not sure running an LLM on device would be a good idea.
I enjoy getting my ale at the click of a button, and keep my arguing capabilities for stranger online.
There may be a place for AI driven games but there is literally no reason to shove it everywhere. Pre-written dialogue is much more enjoyable to engage with on the long term, contrasted with having to think about phrasing for an NPC that spouts generic fantasy speak.
After hearing the "everyone pickpockets everyone and goes to jail and/or dies" anecdote for the "original" Radiant AI, I'm beginning to suspect that the following are incompatible:'
– there's always enough interesting characters to interact with to give quests etc.;
- live simulated world with emergent behavior that involves characters disappearing;
The "always enough interesting characters" problem needs to be solved by something along the lines of "if an important NPC dies, the role is passed to an heir". But ... also the world needs to be less murder-y, and (related) actually have a closed economy.
The article mentioned the problem of fitting audio on a single DVD (which would only be exacerbated by fallbacks, and no please don't consume my entire SSD) ... there certainly was a regression in video game creative dialogue when everything started to be voiced. And voice synthesis is an example of one of the rare problems that AI might actually be able to solve fairly reliably, though it's not clear if the jarring exceptions would be more of a problem outside the current utility problems. Though given that the individual input words should be known, probably just converting text to phonemes would suffice.
Honestly I think there is a fundamental incompatibility between: some sense of simulation or realism, and a high enough density of interesting events per character per hour to meet the player's entertainment expectations. A functioning society just can't supply enough arrests, trysts, bandit kidnappings, secret identities, feuds, marriages, etc etc, without rapidly tearing itself completely apart. There's a reason basically every TV show feels like it going off the rails after a few seasons: you can't lay rails in front of you as fast as episodes consume them. It only works at the beginning because you're borrowing against the stock of events that occurred in-universe before the show began.
Dwarf Fortress kindof solves this by zooming out to increase the character count, as well as the standard fantasy trick of super-charging the economic productivity of everything. Letting 1 dwarf feed 15 by working part-time on a 25 square meter plot of mushrooms helps a lot.
I have remembered the phrase "Radiant AI" from the Oblivion Marketing when it came out, 2005ish I guess, when I was in high school. I'm glad it stuck with someone else as much as it did for me: the hype, the disappointment, but also the wondering what it could have been, because it sounded like a legitimately very-cool game feature except for the part where it didn't exist.
What an amazingly well researched and interesting post. I’m very grateful to the author for having done the legwork to research all of this.
I loved how they were able to peel back the Todd Howard reality distortion field to really understand how Bethesda went from that famous E3 2005 demo to what we got in the end.
>> The technology powering this next generation title is doing so much more than simply making everything look great, it’s also changing the rules of how virtual game worlds function. As mentioned before, the area of Tamriel that is the setting for Oblivion is populated with 1,000 NPCs. Unlike current games, these characters don’t simply disappear once the player leaves the area, they exist 24 hours a day, seven days a week. Every character has its own virtual life and its own schedule to follow.
You know who really did this? It's a game called Rain World [1]. In Rain World, the world keeps turning when you're gone. Literally: predators and prey go about their way, chasing and fighting and eating each other, after you die. When you come back they don't respawn. The game simulates their actions while you're away and you meet them again in medias res, doing whatever they were doing while you weren't there.
And what was the reaction of the gaming public to that? A typical reaction on release was this article by Brendan Caldwell on Rock Paper Shotgun, whence I quote:
Modern platformers that want to be difficult have learned the value of a quick and nearby spawn. Fell into some spikes? Never mind, says the game, and one second later you are at the last brick wall you leapt from. The slugcat doesn’t get this treatment, instead it is transported back to the nearest save point, the last hibernation chamber. The things you have done to the environment have been undone, the parts of the map you revealed have been recovered in shroud. You are ten screens back from where you were, only now the predators and prey will be in different places.
(...)
The oddest thing about it is that, like the controls, this difficulty feels entirely deliberate. It is like Rain World wants to have the strength of difficulty we find in Dark Souls. But that classic of dying and re-dying had the impetus of soul currency, a sense of gambling, a sense of pace, and the relief of clever shortcuts with near-perfect geography. Not to mention the HUGELY SIGNIFICANT gesture of always putting the enemies reliably in the same place, like a solid, immovable set of spiky hurdles. You always had the means to overcome and defeat them. You just needed to learn.
In other words: "What? I can't just memorise enemies' positions so I can defeat them by muscle memory alone??? I have to think?? Each time?? During a game????"
:Throws controller:
Yeah, so much about AI simulating enemies that have an independent existence.
I'm not sure about the limits of the engine's complexity, but the game "shadows of doubt", a procedurally generated murder mystery game, has a giant sandbox with characters that have jobs, partners, visit restaurants, and more.
I don't think the NPCs can do more than a handful of actual actions, but the way you can find who a character met by watching the security tapes of a particular restaurant from a particular time because on of the bat's neighbours says "I saw this person here last night" when you ask about your murder victim is extremely impressive.
There is definitely a sense that you've seen everything after a while because of the limitations of procedural generation, of course, but a sandbox like that combined with scripted quests would make for some really fun gameplay outside of the main quest.
Honestly the big problem isn’t the tech. The AI techniques you use most often are decades-old and well-known, LLMs don’t really enter in to it except for generative dialogue.
The problem is that except for in a handful of cases the idea is often more appealing than the reality.
Nah, the problem is that the enjoyment vs effort graph is a valley. Just adding some simple procedural behaviours isn't all that interesting and possibly creates bugs.
You need to spend a lot of work and add a bunch of behaviours and interacting systems for line to start going back up and arrive in place where games like Rimworld, M&B: Bannerlord or Dwarf Fortress are.
Like if your Radiant AI makes NPCs bandits go and attack NPC caravan, that's not all that interesting, and hell, player might not even notice and think it is scripted. Because aside from some quick loot there is no impact whatsoever on world, you can get rid the world of every banding within 10 mile radius and nothing will change aside from amount of loot in your inventory.
But if you do similar thing in M&B:Bannerlord... there is actual (if simplified) economy there. You CAN starve a city if you just kill all merchant caravans going in, and raid the villages, and the city economy will go down, they will man less guards on siege and have less resources... and on other hand you can make sure local economy flourishes and that will cause prices to go down on stuff, which will cause city that now has access to cheaper weaponry to have more guards.
If killing a bunch of bandits made city prosper a bit more (or vice versa, attacking traders and caravans made it poorer), if clearing local mine made some miners to move in to provide to city, if sabotaging army camp made a dent into political situation (imagine winning Skyrim rebel/empire conflict by sabotage like that), now we're starting into it being interesting rather than a gimmick.
Yes, there is a development cost, but it’s a feasible system. If the payoff is deemed worth it, it can be done at the expense of something else. It’s a design decision, not a limit of technology or even budget.
The problem is that it’s hard to make a game a better experience this way, and for many games it would distract from or confuse the core experience, making it worse.
It’s a well-explored problem too. Anecdotally, in my career I’ve worked on three games where this kind of system (at various levels of complexity) was proposed. Game designers and programmers love this stuff (I do). In the end these ideas were abandoned simply because they didn’t make the player’s experience better.
If a tree falls in the woods, and no-one is around, does it make a sound? If the player encounters that tree lying on the ground, do they care that it’s fall was simulated after some event, or is the impact the same as if a level designer or procedural generation system placed it there? Will they even notice it? Can we make sure the simulated tree falls in a way that doesn’t break navigation systems, or cause a collision issue where the player can get stuck, because then they’ll definitely notice it in the worst possible way, etc.
These are not impossible problems but it really takes a special type of game to make it not only worthwhile, but better for the player, and probably a special type of player too.
I don't actually want to have a deep, philosophical conversation with a blacksmith.
I just want to see that blacksmith close up shop early because he's feuding with the town guard, or give me a discount because his daughter just won the local archery competition. I want a world that reacts to itself, not just to me.
The goal shouldn't be to make NPCs that can pass the Turing test, but to make a world that feels like it has a pulse.
Agreed, that's the real dream of open world RPGs: dynamic worlds. Perhaps modern AI techniques can help in that a bit, but what you really need is an incredibly intricate simulation.
Do you really want that in a scrolls game though? I want the blacksmith to be first npc in the town, more or less always there, with 1 button on the dialog tree to get to the shop menu for me to unload an entire dungeon of loot onto this blacksmith. And he better have ore and leather strips.
And I want to be able to game player statistics using a combination of spells and potions so I can pickpocket the blacksmith and then sell their stuff back at marked-up prices. The traditional RPG numbers-and-skills-and-formulas part of TES was a great joy to exploit.
It doesn’t need to be a deep philosophical conversation. You could be striking up a “buy now pay later” business deal or asking him to produce a specific type of equipment according to your specifications, etc.
It’s entirely possible that Radiant AI in its entirety is actually in the original oblivion and the remaster.
It’s just that they either forgot to enable the build flag, or part of their production release is to pick a random commit as gold master.
As people had already parted with their money it’s been given the same priority as the game breaking bugs - which is to say it was left for the community to fix.
Maybe they will put out a “hotfix” in another 15 years to enable it.
I strongly believe that no bethesda employee has ever played a release version of their games.
Radiant AI is in Oblivion and every game they've made since then. There's nothing to enable. The issue is primarily with the game content; it's used all over the place, but in the final game it's not very impactful.
By “in its entirety” I mean as it was promoted and originally demoed. The release version is vastly different to the pre-release demos (some are linked in the article)
Yes, I wrote the article. I wouldn't say the system in the release version is vastly different to the pre-release demo (there was only one to my knowledge, the E3 2005 one), as that just demonstrates a tightly scripted sequence of events, which one of the developers was open about even before Oblivion was released. Some things about the system definitely changed (such as disallowing NPCs to pickpocket from the player), but I don't think we have any evidence of whole systems or major behaviors that were actually implemented at some point and cut before release.
That said, I’m pretty sure that they said the e3 demo wasn’t scripted (edit: the quote in your article confirms it, too).
We were expecting, at the time, a game like in the demos. But as you stated, it’s probably more content related, in that they didn’t actually schedule much (or any) complex combinations of those packaged behaviours or npc2npc interactions as shown in the demos - leaving only simple instances of the packages you described. Maybe the dependency chain of goals has some concrete limit, for example.
It’s mostly just “go here”, “find food”, “eat food”, “sleep” (which I suppose emulates life, but isn’t what we were expecting).
Although I guess that the amount/complexity of wrangling the behaviours of 1000 (???) npcs to stop the game being unplayable due to goals being destroyed is why it’s just so passive in its release form.
There's another quote which explains what they meant by "it's not scripted": it's not using their (text-based) scripting language, but the entire sequence is more or less 100% deterministic, using AI packages to control the behavior:
> The reason it’s AI and not scripting is because it uses goals and rules to determine how something is going to be accomplished.
> In the sense that it’s a sequence of events that happen in a particular order, you might consider it scripted, but the way you set up those events, and how the actors accomplish them, is not scripted.
I was referring to you saying above that the e3 demo was “tightly scripted”. I never suggested it was, just that it was much more complex than what was released.
I think more likely scenario is that in QA testing there was so many edge cases that between demo and release they disabled a lot of it; limited amount of power of consoles might've also been a factor
> The addicts live in a locked cabin, so it’s unlikely for the player to enter it unless they are specifically looking for it.
This is overlooking a crucial, obscure, and unintentionally hilarious detail: not all the skooma addicts are in the cabin! Out in the world are two NPCs who make a monthly inter-city trip to the den to get their fix. However, due to a bug where these NPCs are assigned to the wrong faction, they can't actually get through the locked door of the den, so they'll stand outside the door drinking skooma forever, unable to progress to the step of the AI package that would eventually return them home to their usual schedules, unless the player unlocks the door for them. https://en.uesp.net/wiki/Oblivion:Trenus_Duronius
reply