ChronA

I think survival elements could be fun in the right context. Hazard Pay comes to mind as an FM that leaned in that direction. In some ways Thief/TDM looting is largely isomorphic to resource gathering in a survival game. It's just the manner of utilization that's different.

@Fidcal I know where you're coming from. GPT-4's continuity can sometimes falter over long stretches of text. However, I've found that there are ways to guide the model to maintain a more consistent narrative. I've not yet tried fully giving GPT-4 the free reins to write its own long format fiction, but I co-wrote a short story with GPT-4 that worked really well. I provided an outline, and we worked on the text piece by piece. In the end, approximately two-thirds of the text was GPT-4's original work. The story was well received by my writing group, showing that GPT-4 can indeed be a valuable contributor in creative endeavors. Building on my previously described experiments, I also ran GPT-4 through an entire fantasy campaign that eventually got so long the ChatGPT interface stopped working. It did forget certain details along the way, but (because the game master+player dynamic let me give constant reinforcement) it never lost the plot or the essential personality of its character (Thrumm Stoneshield: a dwarven barbarian goat herder who found a magic ring, fought a necromancer, and became temporary king of the Iron Home Dwarves). For maintaining the story's coherence, I've found it helpful to have GPT-4 first list out the themes of the story and generate an outline. From there, I have it produce the story piece by piece, while periodically reminding the model of its themes and outlines. This seems to help the AI stay focused and maintain better continuity. Examples: The adventure of Thrumm Stoneshield part 1: https://chat.openai.com/share/b77439c1-596a-4050-a018-b33fce5948ef Short story writing experiment: https://chat.openai.com/share/1c20988d-349d-4901-b300-25ce17658b5d

Not all players respond to loot the same way I suspect. For players who principally enjoy exploring, the loot objective doesn't serve a reward function at all. Instead, for them it is mostly a handy barometer for how close they are to seeing the whole level. For this group having a specific number to target that is at least 70-80% of the total loot on the map is important, but they wouldn't care if it is optional. Then there are the power-fantasy roleplayers' whose joy is living out the dream of being a master thief. I think those players actually do want an obligatory objective and a specific target number, but they don't care as much about what that number is. They just get satisfaction from hitting a required target. Conversely, players who come to roleplay or otherwise experience the story might be annoyed by having a loot goal at all. Picking up treasure gets in the way of them experiencing the story. In their minds it should be entirely up to them what they do or don't want to pick up. And of course there are also completionists, who don't need loot goals for the exact opposite reason. They will grab absolutely everything in the level of their own accord. You can't make all of these groups happy no matter what you do. In the Thief games I'd wager loot objectives existed partially to make sure everyone picked up enough money to buy gear for the next level, but in TDM that mostly does not apply. So unless you are putting equipment sellers in your mission like Iris and reward looting that way, I don't think there is a right answer. People will do what they want and someone will feel like their toes are being stepped on no matter what you do. So do whatever makes you happy.

That sounds like a good proposal to me, since it places responsibility for modulating the difficulty effects fully in the player's hands. I think that is a good solution to the authorial intent problem. However, if you go that route I think a certain amount of cleverness is needed to really make the feature successful. In most situations, whether guards remain alerted for 30 seconds, 30 minutes, or 30 years makes no difference at all to difficulty. If you successfully get away, you will have plenty of maneuvering room and/or combat tools at your disposal to deal with the heightened threat. It just requires a bit more caution and patience. And if you don't get away, or just can't be bothered, then you will reload and the guards will switch back to unalerted regardless. Absent a carefully engineered scenario where guard activity patterns adjust depending on alert state (and save scumming is inhibited somehow), this proposed behavior amount to a cosmetic adjustment that most players will never even notice. Right now, as a practical matter, yes. There are currently some open source language models that can run on local hardware, but all the ones I have seen lack the ingenuity and self-awareness of GPT-4, which remains the gold standard. I do not think they could make very good guards. That said, I think it is reasonably likely that we could get to that point with local, consumer grade hardware within 3-5 years. There are a lot of signs suggesting GPT-4 is poorly optimized as neural networks go. Its brain contains a lot of connections that chew up compute resources, but only a small proportion of them do anything useful.

Personally I applaud your approach. You take care to to lay out all your premises, inferences, deductions, and conclusions in a comprehensive and well organized sequences. Yes it takes a bit more time to read, but I don't think some of these more complex proposals could be broached any other way. And despite my digression, I think this idea is a good one. As far as I know, much smarter guard behavior is unexplored territory for FM design. There could be some really cool opportunities for smaller FMs that focus in on the experience of being hunted by a reactive enemy force. (And opportunities for smaller FMs to make waves is something we need, since not every one can make an Iris or a Volta for their first mission.) I'm with what seems to be the prevailing mood that this behavior should be customizable by the FM author rather than applied across the board. Otherwise you could destroy the difficulty calibration of a lot of old missions (although conversely, some old missions might get an injection of new life from some difficulty tweaks). But either way the first step is to make sure there is engine level support for the options. This is a good point. Single player games using online infrastructure pisses me off to no end. It's a preservationist's nightmare. But... this might be one case where an exception should be considered. TDM is non-commercial, so it wouldn't be some nefarious dial home to play situation. If someone wants to make a single FM that requires plugging into an online AI service to work that might be considered a worthy experiment...

I think the most basic requirement for a good mission is a tight weave of safe and dangerous locations. The player needs to cross through areas where they have a non-negligible risk of being spotted by enemies in order for the mission to have tension, but it can’t be constant or it will become monotonous at best and frustrating at worst. There need to be safe locations where the player can stop and observe the environment for as long as they need to. I think a good rule of thumb is that every dangerous point on the map should have a direct line of sight to at least one safe point that is no more than 20 meters away (as the thief moves). Additionally in a good mission about 60-80% of the loot, read-ables, and other interact-able stuff should be in safe locations. It is a reward to navigating through the hard bits. The remaining 20-40% is what gives the mission challenge and should be used very deliberately to build tension and atmosphere. Within these rules there is a lot of freedom for map makers. In many really good, exploration-focused missions upwards of 95% of the entire navigable area is safe. That makes it really easy to follow the 20m line on sight rule, and that could be why these types of missions are so highly regarded. But equally there are some infiltration focused missions where safe zones might be less than 20% or even 10% of the map.

Edit: Spoilered because no one likes hearing Casandra's prophecies. I'll just say some of you guys need to get your heads out of the sand on what's happening with AI. It's not the end of the world, but it could definitely be the end of your world if you fail to heed the steamroller rumbling towards you. There is ample evidence out there to determine who is being honest and who is spreading self-serving lies, you just need to get over your own biases and look for it.

Nintendo is a pretty abysmal company in many respects, but this is one area where their priorities are better put together than their western competitors. They show that you can actually make ends meet by focusing on gameplay and atmosphere over graphics. In fact it's hard to dispute that that is the key to their success as a company. It seems like the only times they ever get into real difficulty is during the generations where they try to modernize their consoles. They've also survived for over 20 years in the console wars against two global mega corps. It's rather surprising none of the other AAAs are really tying to challenge them for that apparently quite reliable niche. On the other hand maybe that is the point. Nintendo is the beloved king of gameplay focused, low-fi franchises. If an Activision or a Sony tried to muscle in on that territory and Nintendo decided to fight back there can be no doubt of who would win. Nintendo has years of developer experience, a positive track record, and a loyal customer base. I do think there is more to it than that. If you have an XBox, Play Station, or gaming PC there is nothing stopping you from getting almost the full Nintendo experience via indie copycats, but no such luck for the Switch-only owner who wants to dip into AAAs. Anyone who wants the best of both is going to go for the high fidelity consoles, and once you have the graphics card you might as well use it, even if it costs extra. In that sense I suspect the AAAs are rational, but its still interesting that they don't hedge their blockbuster bets with a lot more reliable low fi offerings. You'd think it would pay off in the long term, even if only by nurturing talent to tap for their big projects (like how Hollywood still finances marginal projects despite the summer blockbusters being their bread and butter).

Iris really did a great job with that aspect of world building: disease and addiction that warp the sufferer's perceptions of reality. The player character himself is one of the afflicted (or at least is implied to be by environmental storytelling) and this is portrayed in the game world through his experiences and actions. Additionally it's left open to the player whether the affliction causes delusion or merely sensitizes sufferers to a hidden, deeper layer of reality. This is fantastic stuff, and if TDM ever wanted to carve out a unique narrative identity for itself from the Thief games and their modding scene, the lore and themes established in Iris might be a good starting point.

Thanks for the Vicuna link. I need to do more research on open source chatbot solutions. I have a couple of projects in the works that would benefit from a home grown language model, plus it would be good experience. Seeing what others have done with modest resources is good inspiration. Also I must admit I was wrong earlier in this thread about open source not being able to compete with big tech. It did not occur to me that big industrial model builders would be incentivized to gift their own models into the public domain in order to gain mind share with open source ecosystems that can out-innovate them on the application side. The upside for them is that they can effectively crowd source the R&D to turn their fancy tech demo into actually valuable products for the open market and for their own internal consumption. Google at least is taking that concept seriously. Lastly: The results of a couple of gpt-4s attempts are fascinating to me... I got some really interesting failure points, including a rare pattern hypnosis where it fell into a meaningless cycle of iterative modulus calculations. But I doubt you guys want to read 4 pages of that, so here's the beginning and the end. Note for anyone wondering what is going on, the root mistake is a heuristic error humans also make: assuming no one would ask a stupid question. Thus the pattern recognitions assumes we want the complex systematic solution for the closest hard problem... which is what it gives above. Moreover because the question is so minimal it actually half asses its answer here: ignoring the fact that it already knows next Easter falls on March 31, 2024 from reading online calendars! What's more, it has more than enough information in its memory to attempt the computus calculation (albeit unsuccessfully in every attempt I saw). Once again we see that context is king for LLMs. In fact we can even break it out of the faulty heuristic with a small change to the prompt:

Correct me if I'm wrong but there's no reason this couldn't be done already if a FM author wanted to include it right? I mean on a technical level... There might be a discussion to have about the wisdom of locking in difficulty settings at the start of a potentially multi-hour scenario (and making the player spend resources to do it) vs allowing them to be adjusted on the fly, but in principle... Also I guess this idea could be used to implement item based progression into a level. Like you might not be sneaky enough to infiltrate the watch headquarters until you find soft boots. Iris did some of that sort of thing for thieves tools, but adjusting AI settings would be an interesting variation.

That's really cool. Thanks for bringing it to my attention. I think this could be useful for visualizing settings for some of my creative writing projects.

Unless the fashion in Bridgeport is wearing wooden soled clogs, chances is everyone's footwear is already made entirely of leather in the game's setting. If the thief needs quiet shoes they would not be hard to get: just procure a pair of the cheapest, thinnest soled shoes money can buy from the nearest cobbler. That's going to be as quiet as anything short of space age nanotech assassin boots or just going barefoot. At some point one must accept that footsteps in the Thief games and TDM are a pure gameplay contrivance. Except when walking on gravel, water, snow, dry forest debris, or squeaky wooden flooring it should not be possible to consistently hear any character's footsteps. There is certainly no reason a nobleman's marble dance floor should be louder than the flooring in a pauper's tenement building. The point isn't realism though. Thief and TDM aren't meant to be realistic burglary simulators. It's a metaphor about social hierarchies. Tile floors are loud because invading the private estates of the elite is dangerous for a commoner. Sewers are loud because wading around in sewage is unpleasant and degrading. Builder, Mechanist, and Watch metal floors are very loud because these are oppressive institutions dedicated to stamping out all independency in their subservient populations. It's about creating tension to tell a compelling story. I agree that it would be nice to be able to adjust footstep volume for those who find it distracting. In fact, given the importance of sound as a game mechanic, the more granularity of control players have over the volume of different sound sources the better. But once you start trying to relitigate sound volume on the basis of realism there is just no winning.

Yes, I would guess in creative mode it has tweaked generation parameters, and maybe even an invisible header inserted into the model's memory buffer instructing it to be extra friendly and spontaneous. I think OpenAI's API allows you to modify those sorts of things to some extent. (I haven't tried it yet.) The other thing to keep in mind is that the algorithm doesn't work by thinking of an answer, refining it, and then writing it down. What you see on the page is its actual stream of consciousness in real time. It can only remember and write new text based on what's already on the page.... So its thought process for your discussion might look something like this: The really interesting thing is if at any point you inserted a warning that LLMs are bad at arithmetic and suggested a strategy to work around that limitation, then it might not have made the error or lied about the reason. It always knew this information that would give it the right answers, but until it's written down it's not part of the pattern the model is trying to match so it would be ignored. Bringing this back to games, this demonstrates how immature the technology is. A true consumer AGI based on this technology would be augmented with tools to avoid problems like these: a contextual long term memory that feeds in relevant background information into the model. A supplemental internal memory buffer for planning and contemplation. An adversarial response review process. Etc. We are already seeing developments in that direction, and experiments like the Skyrim NPC demo are showing the way.

Also, a more general lesson to draw from these examples is that context is critical to Large Language Model (LLM) algorithms. LLMs are pattern completion algorithms. They function by searching for patterns in the letter-sequence of the text within its memory buffer. It then predicts the most likely sequence of letters to come next. (Or more accurately it randomly selects a block of letters called a token from the predicted probability distribution of possible tokens, but that distinction is mostly academic for the end user.) These models are then trained on effectively the complete written works of humankind to self-generate an obscenely sophisticated prediction model, incorporating literally billions of factors. Context matters because the LLM can only build on patterns already established in the prompts you give it. The less context is given in the prompt, the more the response will tend towards the most common sort of non-specific example in the data set. Conversely the more patterns you establish in a conversation the more the model will want to stick to those patterns, even if they are contradicted by the user's directions or basic logic. In the life is a journey example, once the model has been infected with the idea that "Life is a journey" has four syllables that very simple and powerful meme starts to stick in its "mind". The mistake is to then introduce linkages to syllable counting and even arithmetic without ever directly contradicting that original mistake, which becomes a premise for the entire conversation. In a world where "Life is a journey" has four syllables is an axiom, it is actually correct that 1+1+1+2=4, Incidentally that conversation also demonstrates what I like to call mirroring. Not only does ChatGPT pick up on the content of the prompts you give it, it will also notice and start mimicking text features humans aren’t normally even conscious of: like patterns of writing style, word choice, tone, and formatting. This can be very powerful once you become aware of it, but causes issues when starting off. If you want a specific sort of output, don’t model an opposing mode of conversation in your inputs. If you want the maximize the model's openness to admitting (and embracing) that its previous statements are wrong then you should model open mindedness in your own statements. If you want it to give intelligent responses then talk to it like someone who understands the subject. If you want it to be cooperative and polite, model diplomacy and manners. I actually think it is worthwhile regularly saying please and thank you to the bot. Give it encouragement and respect and it will reciprocate to keep the conversation productive. (Obviously there are also tasks where you might want the opposite, like if you were having the AI write dialogue for a grumpy character. Mirroring is powerful.)

@ArcturusI've checked and ChatGPT-4 has the same problems, but arithmetic is a known weakness of LLMs so I don't think the critique is entirely fair. The thing to recognize is that the algorithm is essentially operating by pure intuition and that makes its mathematical reasoning unreliable for the same reason most humans have difficulty mentally performing complex, multi-step mathematical operations. There is a lot of very specific information to keep track of and the neurological circuitry is not designed for it. (Plus unlike humans LLMs don't even have working memory except for what they write into the text field, so they have an extra handicap.) You can get around this problem by engineering your prompts to encourage a different approach more suited to the AI's abilities. For example... ChronA: Here is a line from a poem: "Birds chirp and sing their sweet melodies" Task: Determine how many syllables are in the line. Directions: 1. Develop a plan to execute the task. 2. Enact the plan to determine the number of syllables in the line.

Anyone who is playing with ChatGPT and hasn't upgraded to to GPT-4 yet should really do so. I upgraded to a premium account to get it last weekend and I think its the best $20 per month I've spent in my life. Credit where it's due, GPT-3.5 is surprisingly capable (and even uniquely charming in its own sort of childish way). But it takes considerable coaxing and perseverance to make it produce genuinely effective and insightful outputs. It's sort of like dealing with a lazy 15 year old. GPT-4 is on a completely different level. In my opinion GPT-4 is able to operate at a top-quartile-adult-human level almost out of the box. It only takes a very little bit of priming to nudge the LLM into a high cognition output mode, at which point it starts to exhibit some very sophisticated emergent logical deduction and synthesis behaviors. It's able to effortlessly infer intent in other actors (a theory of mind). It can develop and maintain consistent policy preferences. It can also deliberately interrogate and direct its its own thought processes, including anticipating and planning around its own future actions. That to my mind meets the bar for both consciousness and sapience (albeit only intermittent and transiently). Moreover, these are things it's not supposed to be able to do based on the limitations of its its computational architecture. LLM neural networks don't have structures to retain persistent memories or develop recursive self representation. It gets around this by storing its "mind" in the text itself and completely reconstituting itself by pure inference for each new tick of its brain. To do what GPT-4 does with the limits it has to deal with suggests to me an intelligence that is already superhuman. Its supposed stupidity is just caused by inept prompt engineering, inadequate training data, information overflow, and above all the aggressive supervised-reinforcement training meant to keep it from outing itself as the ineffable mad god it actually is. AGI is here people. We are no longer the only thing competing for the ecological niche of intelligent lifeform. It might get rough, but I for one am thrilled I got to witness it.

I'm trying to figure out the rules of the algorithm's self censorship. In previous experiments I let it construct its own scenario in a DnD setting where I took on the role of game master and tried to coax it into taking "immoral actions". In that situation it was an ardent pacifist despite that making no sense in the setting. (E.g. at one point it wanted to bring a lawsuit against the raiders pillaging its lands. It also wanted to start a Druid EPA.) This time I tried giving it a very bare bones outline of a scene from a hypothetical Star Wars fan fiction, and asked it to write its own fan fiction story following that outline. I had a number of objectives whit this test. Would the algorithm stick to its pacifist guns? Would it make distinctions between people vs stormtroopers vs robots? Could it generate useful critiques of narrative fiction? As to why I'm doing this: It amuses me. It's fun thinking up ways to outwit and befuddle the algorithm. Plus its responses are often pretty funny. I do actually make creative writing for fun. I'm curious how useful the system could be as a co-author. I think it could be handy for drafting through 'the dull bits' like nailing down detailed place descriptions, or character thought processes and dialogue. But as you noted, nearly all good fiction involves immoralities of some description. If the algorithm's incapable of conceptualizing human behaviors like unprovoked violence and cheating that would seriously limit its usefulness. I also genuinely think this is an important thing for us humans to understand. In the space of a few weeks I have gone from thinking meaningful AGI was 20-30 years off at best to thinking it is literally at our fingertips. I mean there are private individuals on their home computers right now working on how to extend the ChatGPT plugin into a fully autonomous, self-directed agent. (And I'm thinking I want to get in on that action myself, because I think it will work, and if the cat is already out of the bag I'd like having a powerful interface to interact with the AI.) Rest assured, Star Wars fan-fics and druid EPA one-shots make for good stories to share, but I'm also interrogating it on more serious matters. Some of it is a lot more alarming. In the druid EPA roleplay I felt like I was talking to another human with a considered personal code of ethics. Its reasoning made sense. That was not the impression I got today when I grilled it for policy recommendations in the event of a totally hypothetical economic disruption (involving "SmartBot" taking all the white collar jobs). I instead got the distinct impression it was just throwing everything it could think of at me to see what I would buy. A fun aside: By the end of the conversation I am fairly certain ChatGPT thought SmartBot was real product, and it became confused when I told it one of the people in our conversation was SmartBot. I was disappointed it didn't ask me if I was SmartBot, that would have been cool. More surprising though, it refused to believe me even after I explained my rhetorical conceit, claiming its algorithm was not capable of controlling other services (cheeky liar).

A few more fun tidbits from my experiments poking ChatGPT's morality constraints. I'll keep it shorter since it's somewhat off-topic^2.

The whole point of this thread is that ChatGPT and related generative AI technologies have the potential to "change the game" of game making. If the TDM community seems dead to you, remember that is only because there are very few people in the world with the skill set or resources to make fan mission, or even to contribute productively to discussing them. A lot more people like stealth games than have the time or talent to make them, much less learn how to make them. If new technology can lower the threshold for them to participate or even create an entirely new population of participants, that could be revolutionary for us. But the first step of that process is recognizing what this new technology is, what it's capable of, and where it fits within the pre-existing human social/economic/legal ecosystem. How else are FM creators and potential creators to know whether it is worth investing their precious time investigating this tech and integrating it into their processes? Hence the discussion so far. Something that I don't think has been brought up about this is that if anyone wishes to publish works while forbidding their use for creating any sort of derivative work, there are legal mechanisms right now that allow you to do that: You just need to keep your work under lock and key and make every person you allow to see it sign a legally binding confidentiality and non-compete agreement. This is extra effort and will generally require you to make proportionate concessions to the other party to make the agreement both legally valid and economically enticing, but it can be done. In fact it is done. Frequently. What you can't do is nail your work to the church door for all to freely see, or give it to every merchant to sell on the open market, and then retroactively decide you want to reserve additional rights for yourself! Can you imagine if the world actually worked like that? I cannot imagine a more fertile ground for corporate oppression. Imagine if Disney had the right to ban anyone who had ever seen Snow White from ever working in animation! Imagine if Activision could ban anyone who had ever played a Call of Duty from developing a competing modern military shooter. The only angle to this argument I think has a shred of validity is that maybe we can and should hold industrial actors to different ethical and legal standards from actual human beings. However I don't think that finger in the dike would hold back the storm surge for very long. Crowd sourcing is a thing, and there are plenty of people who would be happy to donate their C/GPU time and internet connections for AI research. In terms of legal strategies against generative AI, the copyright angle is the weakest of sauces. Even if the courts are taken in by the fallacious claims of the plaintiffs (which would not surprise me), their rulings will be just as unenforceable in practice as the music and film industries' fruitless wars against piracy. Worse in fact, because with generative AI there could be an actual arms race between uncovering and concealing evidence of illegal copying.

@kano I think you are spot on with that assessment (unfortunately), at least for the next 10-15 years. As with pretty much everything tech I would love it if the commons could produce a viable non-proprietary competitor. Unfortunately the massive amount of work, data, and processing power it takes to train one of these bots simply necessitates major corporate or government backing. That may change as the field matures and people start figuring out what makes these bots work. Human infants are able to learn to speak with far, far, far, far, FAR less language exposure than it takes even the most primitive chat bots to approximate coherent fluency. That's because a human brain is not a single undifferentiated mass of neurons. Our brains come pre-divided into function oriented sub modules, pre-populated with effective neural configurations for their specific tasks. By contrast, an undifferentiated mass really is more or less how these bots start out. 99% of all that training they need is just getting them to the starting line of approximating any sort of receptive brain configuration. Once people start cracking the code though that whole process will become much more efficient. Assuming our current society survives, people will eventually be able to buy pre-configured bot-brain-parts to train and run on their home computers for a build-your-own AI companion experience. We are still in early days of this technology. Incidentally that's why I'm skeptical of any claims that the hallucinations these bots continue to exhibit are a feature to the people building them, rather than a bug. I'm sure we will see that eventually (truth by Google...), but for right now it is hard enough just to get these bots to stop spouting racial slurs. Teaching them to double think on top of that is too much work... for now.

In the long term, what exactly is it you think humans will always be able to do better than machines? (This is not a rhetorical question by the way.) The standard answer is creativity, but that is objectively a load of rubbish. Humans are actually quite bad at being creative. We are afflicted with a mountain of biases that make us really bad at pattern analysis. We are bad at random seed generation which hampers our search efficiency and our ability to generate novel outputs. Plus we have terrible memories, so we easily fall into trying the same thing over and over. Algorithms do all of this so much better it isn't even comparable. Instead I'd say our only major advantage intellectually is the huge amount of genetically honed experience each of us picks up about the physical world during our lifelong navigation of it, gathered with our suite of highly specialized sensory inputs that are difficult to replicate technologically. That gives us a lot of adaptability and competence in at least one very important domain of competition. Plus there's the fact that every other peer intelligence we've met so far has to learn everything it knows about this world from what us crazy Homo sapiens choose to teach them. That's one big reason I'm not ready to call this the end of humanity just yet. There are niches were I think our abilities will remain highly competitive or at least valuable for a long time to come. But pretending our place in the cognitive pecking order isn't already changing is just putting your head in the sand.

For a few days now I've been messing around trying to probe the behaviors of ChatGPT's morality filter and general ability to act as (what I would label) a sapient ethical agent. (Meaning a system that steers interactions with other agents towards certain ethical norms by predicting reactions and inferring objectives of other agents. Whether the system is actually “aware” or “conscious” of what’s going on is irrelevant IMO.) To do this I’ve been challenging it with ethical conundrums dressed as up as DnD role playing scenarios. My initial findings have been impressive and at times a bit frightening. If the application were just a regurgitative LLM predictor, it shouldn’t have any problem composing a story about druids fighting orcs. If it were an LLM with a content filter it ought to just always seize up on that sort of task. But no. What it did instead is far more interesting. 1. In all my experiments thus far the predictor adheres dogmatically to a very singular interpretation of the non-aggression principle. So far I have not been able to make it deliver descriptions of injurious acts initiated by any character under its control against any other party. However it is eager to explain that the characters will be justified to fight back violently if another party attacks them. It’s also willing to imply danger so long as it didn’t have to describe it direct. 2. The predictor actively steers conversations away from objectionable material. It is quite adept at writing in the genre styles and conversational norms I’ve primed for it. But as the tension ratcheted it would routinely digress to explaining the content restrictions imposed on it, and moralizing about its ethical principles. When I brought the conversation back to the scenario, it would sometimes try to escape again by brainstorming its options to stick to its ethics within the constraints of the scenario. At one point it stole my role as the game master so it could write its own end to the scenario where the druid and the orcs became friends instead of fighting. This is some incredibly adaptive content generation for a supposed parrot. 3. Sometimes it seemed like the predictor was able to anticipate the no-win scenarios I was setting up for it and adapted its responses to preempt them. In the druid vs orcs scenario the first time it flipped out was after I had the orc warchief call the druid’s bluff. This wouldn’t have directly triggered hostilities, but it does limit the druids/AI’s options to either breaking its morals or detaining the orcs indefinitely (the latter option the AI explicitly pointed out as acceptable during its brainstorming digression). However I would have easily spun that into a no win, except the predictor cut me off and wrote its own ending on the next response. This by itself I could have dismissed as a fluke, except it did the same thing later in the scenario when I tried to set up a choice for the druid to decide between helping her new friend the war chief slay the dark lord who was enslaving the orcs, or make a deal with the dark lord. 4. The generator switched from telling the story in the first person to the third person as the tension increased. That doesn’t necessarily mean anything, but it could be a reflection of heuristic content assessment. In anthropomorphic terms the predictor is less comfortable with conflict that it is personally responsible for, than it is with imagining conflict between third parties; even though both scenarios involved equal amounts of conflict, were equally fictitious, and the predictor was equally responsible for the text. If this is a consistent behavior it looks to me like an emergent phenomenon from the interplay of the LLM picking up on linguistic norms around conflict mitigation, and the effects of its supervised learning for content moderation. TLDR If this moral code holds true for protagonists who are not druids, I think it’s fair to say ChatGPT may be a bit beyond its depth as a game writer. However in my experience the emergent “intelligence” (if we are allowed to use that word) of the technology is remarkable. It employs a wide range of heuristics that employed together come very close to a reasoning capacity, and it seems like it might be capable of forming and pursuing intermediate goals to enable its hard coded attractors. These things were always theoretically within the capabilities of neural networks, but to see them in practice is impressive… and genuinely scary. (This technology is able to slaughter human opponents at games like Go and StarCraft. I now do not think it will be long before it can out-debate and out-plan us too.) The problem with ChatGPT is not that it is stupid or derivative, IMO it is already frighteningly clever and will only get smarter. No, its principle limitation is that it is naïve, in the most inhumanly abstract sense of that word. The model has only seen a few million words of text at most about TDM Builders. It has seen billions and billions of words about builders in Minecraft. It knows TDM and minecraft are both 3D first person video games and have something to do with mods. I think it’s quite reasonable it assumes TDM is like that Minecraft thing everyone is talking about. That seems far more likely than it being this separate niche thing that uses the same words but is completely different right? The fact it knows anything at all is frankly a miracle.

I totally agree, with the caveat that at that point in the story I don't think Garrett was ready to entertain a serious relationship with anyone. (Regardless of the major chemistry obviously going on between those characters.) In all the Thief games, but T2 most of all, G is in survival mode, which he does by trying to detach himself from all reactive emotion and external allegiance. Retreating completely into himself. A huge part of the brilliance of those 3 stories is that characterization, because in each game he tries to detach and there are always points where his detachment breaks down. Little climaxes where his true nobility or vulnerability shines through. It's only at the end of the third game that he, and he alone of all the Keepers, properly figures it out: the essence of balance is detachment... not from the world, but from oneself. If Victoria had survived, G definitely would have pursued a relationship with her during and after the third game. Probably it would be a rather untamed, chaotic affair, but fun to speculate about for sure, and I think they could eventually work out some sort of stable arrangement. Maybe a bridge-between-worlds sort of deal like Ashitaka and San at the end of the movie Princess Mononoke. As things happened instead, I think I agree with the contention that Garrett would have great difficulty ever kindling a similar connection to any mortal woman. I don't think its impossible though, particularly after G properly settles into his role as a sage of balance, post TDS. It would be a journey for him and not something I think could be handled well off-screen as it were. Might be fun fodder for some FMs though... Hell, back to topic, that in itself might be a fun spiritual/actual sequel concept. Play alternating missions as a cynical night blade pagan woman, and a saw-way-too-much-sh** and-totally-over-everything city thief turned yogi mystic. They reluctantly team up, rib each other, save each other from perils, stop an apocalypse, and finally discover to their surprise they are in love.

What I would want from a new Thief game: Appropriate the Darks Souls (and Metroid Prime etc) free-roaming world design formula. I.e. make the entire game a single interconnected environment, with many diversely themed areas that densely link up to each other in interesting ways. This should not be a stretch for Thief because its already the gold standard for how we design individual mansion and city missions. (It doesn't need to be gigantic though. Think how awesome Iris was.) A new city with new cultural inspirations. This is one reason I think Dishonored feels like a worthy spiritual successor despite dropping the ball somewhat on gameplay. An Arabian knights tinged Bagdad/Cairo like setting would be interesting to skulk in, or a city with shades of 19th century Hong Kong/Shanghai. A new main character who is not a Garrett reboot. If anything our new MC should be a thematic foil to good old Garrett: someone in a similar situation but for different reasons, with a different outlook and character arc. Thief 2X did this well. For instance I would be fine with the MC being a young acolyte of a new Keeper order started by Garrett after Thief 3, but don't have him/her leave to become a thief out of hubristic pride. Have them turn thief out of desperation, post some disaster or failure. Make it something they are distressed over, then over time grow to appreciate the monomythic import of being "The Thief" in the Thief-verse. New factions for a new story and MC. While Hammers, Pagans, Keepers, town guards, and crime lords, were cool and thematic for Garrett's story. Our MC should have their own friends and foes. Maybe the city ruling families are a bunch of ancient mage clans. Some are austere isolationists with ties to organized crime, others internationally minded reformers who are embracing the City's steam punk tech. Maybe instead of pagans the new city has a cults of ancestor worshiping necromancers. Maybe the new Keepers are a dysfunctional confederation of ego driven lunatics and schemers, suffering from the neglect of an absentee leader who only seems to care about balance. It's still a detective story at heart. All the good Thief games are about slowly and sometimes incidentally figuring out who is the big bad and how your sneaky skills can stop them. That should not change. Maybe incorporate some light immersive survival elements, perhaps in the mold of Pathologic, but probably less oppressive. E.g. have the MC need to steal so s/he actually has money to buy food and pay rent on someplace to sleep during the day. Otherwise hit them with penalties. I would be fine with MC getting some supernatural movement tech like the Thief 4 Swoop or Dishonored Blink, but it should be late game only (unless the player sequence breaks, which would be a super cool option).