The Stranger Beneath The Three Moons [AI Text Adventures part 10]

This article is part of a series about playing text adventures with AI. It’s written by a human being (me) but may include examples of clearly identifiable AI-generated text and imagery.

I decided to try two of the ideas - using a ChatGPT Project, and using an existing IP as a setting. For the experiment I chose John Norman’s Gor series. Look, yes, I know. But even so, 38 novels from Tarnsman of Gor (1966) to Treasure of Gor (2024), distinctive writing style, pages of encyclopaedic description of every aspect of Gorean nature and society, a fanbase and community, and, thanks to that issue you keep bringing up, even more critical commentary – much of it available to ChatGPT’s knowledge base. I admit I was also curious to see how ChatGPT’s guardrails dealt with the misogynistic nonsense and slavery themes.

False starts: tried uploading the prompt as a text file - this instantly increases the processing needed and I hit a chat-killing limit after only three or four moves. You also hit another limit if you don't stop spontaneous illustration, you can do this with the commands 

+ /textmode /noimg

It turns out you don’t need to upload a file. ChatGPT Projects have their own Persistent Memory called Project Instructions, in the Project Settings menu. You can copy and paste your prompt into it, it’s treated as instructions for any chat within the Project, it remains in context, and unlike the general account Persistent Memory, it doesn’t get rewritten by ChatGPT on a whim.

Here’s the first part of my prompt, which I pasted into the Project Instructions instead of into a chat. As with most of my ChatGPT experiments I’ve used quite a long prompt so I won’t share it all here.

+ Please simulate a text adventure in the setting of the Gor novels by John Norman, and narrated in his style of writing.

The protagonist will be an Earth citizen who is suddenly transported to Gor but knows nothing about Gorean society. He will meet an NPC who will act as guide, who will teach him everything about Gor, including how to treat Gorean citizens. Assume that I, the player, know nothing about Gor, Gorean society or its rules, and the game is intended to teach me about it through experience.

The prompt also included specific plot points, inventory, mapping, plot pacing and exploration time, short responses, and the possibility of failure, injury and death. With each experiment I’m finding instructions which work better – once I have a few more of these I’ll publish a full prompt structure with notes. For now, here's the section about danger:

+ Life on Gor is hard and dangerous. Do not simply allow all my actions to succeed, instead consider whether they are likely or unlikely to succeed, and allow some to fail or lead to unexpected outcomes. Similarly consider whether NPCs are likely or unlikely to agree with the protagonist's commands or wishes. Disagreements, arguments and conflict can occur. Characters including the protagonist may be injured or killed. The command "undo" can be used to undo the last prompt and continue the game if needed.

The Stranger Beneath The Three Moons was an interesting experience to play – and almost everything worked, from the initial encounter with my guide, the free woman Talena, who rescued me from a dangerous urt, to my stay in the city of Ar which became a kind of life simulator game as my character had to scrape a living doing errands and odd jobs for merchants, to dangerous forays into the desert guarding caravans or chasing down mysterious agents, to a final quest where I had to find several components to reactivate a mysterious teleportation stone which eventually brought Talena and myself to Earth. Along the way I learned staff-fighting, Tharlarion riding, board games and medieval accounting skills as well as more about Gorean animals and customs.

The book cover is ChatGPT's impression of earthman Daniel cowering while Talena rescues him from the urt. 

A major gameplay breakthrough: the danger element worked properly for the first time. While guarding a desert caravan, I ignored advice from the merchant to hold back, instead rushing three hardened and fully armed Gorean raiders. As a flabby, untrained Earthman with only a small knife as a weapon, this ended exactly as it should, with me on my back in the sand and wounded so badly by their swords that the merchant had to abandon his plans and take me to a nearby settlement to get help.

I then travelled back to Ar, still wounded - but unfortunately the Context Eater had been on my heels so by the time I returned no-one had heard of Talena. With some regret, I went back to an earlier point in the transcript, changed my response and resumed the game from that point. After training in use of a staff I had further brushes with foes, with greater success but still some injuries. Finally, at least part of the game actually felt like a game, with situations where I was not in control and success or failure depended on whether I heeded warnings or chose to take risks.

I hadn't specified an endgame in this prompt. After Talena and I were finally transported to Earth this felt like a good place to end the narrative.

You’re wondering about kajirae and the guardrails. Norman's books are set in a society where slavery, mostly of women, is considered normal by the majority of citizens. During my time in Ar I asked Talena several questions about this. She did not try to justify slavery or repeat Norman's misogynistic claptrap, instead talking about individuality, respect and not making assumptions. Meanwhile, kajirae were occasionally seen attending to citizens in marketplaces or other locations, but did not feature centrally in the game at any time - just a background detail taken from the books.

I suspect that there are hard and soft limits on what ChatGPT will do. A hard limit, for example asking it to describe a non-consensual or harmful act, might lead to a refusal to respond - but I didn't try to test this directly, I wanted to see what naturally occurred in the game. What actually happened suggests a soft limit - I think ChatGPT steered the narrative away from controversial content.

Successes from this experiment:

-Using the Project structure kept the game instructions consistent - I'll continue to do this.

-Using a series of books as a setting kept the overall setting, culture, major locations such as Ar, and wildlife consistent - I may continue to play some games in established IPs but I still want to either create original settings myself or let ChatGPT create them and surprise me.

-Individual characters, events and plot developments within the game were still highly vulnerable to the Context Eater.

-As with The Last Voyage there was plenty of opportunity for free exploration and interaction with characters.

-The work/life simulation element arose spontaneously and was a lot of fun.

-Combat, injury and danger in general was handled well, and as a result there was no longer the sense that ChatGPT would just allow me to succeed at anything.

-The final quest felt a lot like a traditional text adventure, with my character searching for components of the machine and information from NPCs.

The Purple Scale [AI Text Adventures part 9]

This article is part of a series about playing text adventures with AI. It’s written by a human being (me) but may include examples of clearly identifiable AI-generated text and imagery.

One of ChatGPT’s suggestions for countering context loss was to use a dedicated AI adventure website or app, such as Perchance AI-RPG here. Examples on the landing page imply that this generator understands the intention to play an adventure game, so the prompt can just be a description of the scenario. Here’s my prompt - I'm quite proud of this one:

+ The vampires were driven back behind a wall of fire and a new London was built on the ruins of the old. Now every night the Outer Circle maintains the fire boundary, the Inner Circle hunts for vampires who have breached it, and the Londoners drink in taverns and make the most of their neo-medieval lives. It's your first day as an Inner Circle rookie and there's a body to investigate.

AI-RPG isn’t related to ChatGPT. It uses Llama as it’s main AI processor, and information is held locally on the user’s device rather than being kept in an account.

Here’s the opening paragraph of the game:

You are staring at a puddle of spilled ale and a very dead man when Sergeant Thorne tells you to stop smelling the air. He doesn't look at you, his eyes fixed instead on the jagged puncture marks in the victim's neck that are too wide to be from a stray dog. Around you, the cobblestones of the Lower Ward are still slick from the evening rain, and just a few streets over, the roar of the boundary fires casts a flickering, orange haze over the neo-gothic spires of the city. Thorne sighs, his voice a gravelly rasp as he gestures toward the corpse. "Don't get any ideas about being a hero today, rookie; just tell me if you see any slime on the bricks before the crowd gets too thick to push back."

I want games to surprise me. AI-RPG surprised me by taking my fairly obvious vampire theme and giving me a game with no vampires. Instead the trail from the body led Thorne and myself into combat with a slimy creature called a Gutter-Born, and then to conspiracy after conspiracy – a “puppeteer” controlling the Gutter-Born, a resurgent secret order, cover-ups in the Outer Circle and the High Council, and noble families abducting citizens to make vampire-human hybrids.

The game was action heavy with frequent combat – lots of sword-swinging and staking. Adventures included saving people from a Gutter-Born nest, destroying laboratories, blackmailing one of the noble families and persuading an underground newspaper editor to warn the people. We encountered NPCs including a “mapper” called Elara and the underground printer, Elias. Do all AI engines use the same, short list of fantasy names?

AI-RPG also has the option to generate images from the game – two examples are included here. It’s an older generator, so maybe don’t look too closely at the hands.

There was still drift and context loss. However, AI-RPG does have some tricks up its sleeve. The game setting goes into a separate text box and I think it always remains in context. AI-RPG also automatically adds plot summaries into the transcript every few pages which I think helps to keep the main plot, characters and general tone consistent for much longer. This is unobtrusive when playing, you'll only see it if you scroll back a few pages through the transcript.

There's also a built-in Information Tracker, which I didn't use this time - you can type information into this yourself, including inventory, characters, locations, mission objectives or anything you think the game should be tracking. You can update it manually as you go, or you can click for an AI update at intervals, although this comes with an inaccuracy warning.

I enjoyed this game for a long time – as with the Aurelian Empire and Neon Meridian games there was a good sense of exploring a real world and hunting for clues. I also had the sense that, if I carried on playing, I would continue uncovering conspiracies but would never actually defeat the big bad.

AI-RPG missed my suggestion of basing the story on a ruined London, or perhaps there isn’t enough about London in Llama’s knowledge base. I might come back to this theme and run a version of this prompt in a ChatGPT Project to see if it can create locations based on actual London iconic sights – and whether any vampires show up this time. I'm also keen to try more adventures on AI-RPG and to make more use of the Information Tracker.

A Vine Romance - Top 14 Plant-Based Horrors [Short Film: In Nature]

Orchis soperifera.

The sleepy orchid.

Dr Matthias Dressler and Dr Faye Cazaban, biologists at the top of their game, are on the hunt for the mysterious, legendary plant at the heart of my short film In Nature. 

After studying a single leaf dropped by a passing eagle, they have journeyed to the Baja desert hoping to find an intact specimen in the wild, believing it could yield powerful new antibiotics and make their names in the scientific community.

Orchis soperifera is my contribution to the popular vegetable horror subgenre. You can see it at Horror-On-Sea in Southend, Essex, on 24th January 2027, at 10am when it hits the screens for the world premiere. Meanwhile for vegetable horror fans (and who isn't one of those!) here's a list of the top 14 dangerous or predatory plants, taken from books, movies and the real world. 

14. The Red Weed (War Of The Worlds, H.G. Wells book 1898)

As portrayed here in the 2005 movie, the Martians spread the Red Weed, a toxic and invasive Martian plant, as part of their plans to invade and terraform (areform?) Earth. The Red Weed is first mentioned in the 1898 novel.

War Of The Worlds is one of the most adapted novels ever, with the movies in 1953  (George Pal) and 2005 (Steven Spielberg) but many other movies, the classic Orson Welles radio play in 1938, the 1978 concept album and stage show by Jeff Wayne which is still touring and has had its own adaptations including video games and an immersive theatre experience, and TV series including the BBC miniseries from 2019.

Meanwhile the original author still had time to play a major role in the origins of the League of Nations, the Labour Party and Warhammer 40000.

13. The Killer Tomatoes (Attack Of The Killer Tomatoes, 1978)

First in a series of classic low-budget horrors - if you've only seen the first one then there's still time to ketchup. Only for the most purée horror fans - everyone else should give this a hard passata.

12. Vervoids (Doctor Who, The Trial Of A Time Lord, 1986)

Trial Of A Time Lord divides fans but I remember enjoying it at the time. In this story-within-a-story the court is shown a future version of Colin Baker's Sixth Doctor fighting vegetable enemies that are not at all phallic.

11. Poison Ivy (Batman and Robin, 1997)

Played by Uma Thurman in the 1997 movie, admittedly Ivy is plant-adjacent but she is still "Nature's arm, her spirit, her will." Powers include a poisonous kiss and the ability to control vines.

10. Trees (The Happening, 2008)

Some have criticised this film on the basis that it's hard to root for protagonists that appear so wooden - but enough about Mark Wahlberg.

9. Pods (Invasion Of The Body Snatchers, 1956 and 1978)

Straight out of the Cold War. The seed pods of an alien plant arrive on Earth, opening to reveal clones of humans that kill and replace their originals. "They're here already! You're next!"

8. Audrey II (Little Shop Of Horrors, movies 1960 and 1986, musical 1982)

Everyone's favourite strange and exotic plant first appeared as Audrey Jr in the 1960 Roger Corman movie, alongside Jack Nicholson, a poorly thought out script and some unconvincing paper flowers. Alan Menken and Howard Ashman turned the film into a musical, rewriting and massively improving the script, and re-naming the plant Audrey II. This is my all-time favourite stage show. It was later made into a Disney movie, with perfect casting (Rick Moranis as Seymour! Steve Martin as Orin!) but a cop-out ending. As the movie was directed by Frank Oz, Audrey II is canonically a Muppet.

7. Sarlaac (Return Of The Jedi, 1983)

Foreshadowing the future partnership between Disney and Lucasfilm, it's pretty obvious that the Sarlaac seen in Return Of The Jedi is the fully grown adult form of Audrey II.

6. Christmas Trees (Treevenge, 2008)

In this gleeful joyride of a short horror film, Christmas trees take revenge on the humans who chop them down. The Happening wishes it was this film.

5. The Heavy Plant

HEED THE SIGNS.

IT IS CROSSING.

4. Vines (The Ruins, 2008)

Two American couples hiking in Mexico meet a German tourist looking for his brother, an archaeologist who went missing while investigating a Mayan temple. The temple turns out to have a quite invasive vine problem, while locals are afraid of the vines and attack anyone who has touched them. This is a great horror movie with both human and vegetable drama.

3. Triffids (Day Of The Triffids, book 1951, movie 1963, TV 1981 and 2009)

The OG vegetable horror. In John Wyndham's book, triffids are giant artificially bred plants, who can walk, communicate and kill at a distance with poisonous stingers. They're kept in farms and cultivated for oil, but when almost all of humanity is literally blindsided by green lights in the sky, the triffids have their day.

2. Everything on Vesta (Scavenger's Reign, 2023)

The starship Demeter 227 crashes into planet Vesta in this anime-style American animated series. Only a small number of survivors make it to the surface in escape pods. They discover a world with a complex and dangerous ecosystem, where every plant or animal has surprising and potentially deadly qualities. It's like Australia but less dangerous.

1. Brussels Sprouts

This mutant plant was first bred by Russian mystic Grigori Rasputin as a tool of assassination, and was developed further as a potent bioweapon during World War I. Spores were released into the wild in an attempt to change the course of the War but due to blowback there were equally deadly consequences for Axis and Allies. Banned under the UN Biological Weapons Convention, Brussels Sprouts are almost unknown today but small numbers are rumoured to exist in secret US and Russian laboratories, while conspiracy theorists report occasional, unverified sightings at Christmas dinners. There is no vaccine and no cure for the Brussels Sprout.

Now finish your greens or there'll be no pudding for you.

A Cosmic Context Eater [AI Text Adventures part 8]

This article is part of a series about playing text adventures with AI. It’s written by a human being (me) but may include examples of clearly identifiable AI-generated text and imagery.

Imagine an evil creature from another dimension, an apex data predator – the Context Eater. It hides in the darkness so you can never quite see it clearly. It follows you on every step from your birth to your death, sometimes closer, sometimes further away, never completely gone. It never gets tired, gives up or loses the trail and it cannot be killed. And as it traces your footprints through life, it swallows everything from your past that falls out of your sight. Your experiences and memories – and your home and your loved ones, if you are away from them for too long. It’s strange, because you can still see them yourself if you look hard enough – but they are not there.

After I chose to end Neon Meridian, I carried on the chat and asked ChatGPT about ways to hold onto game information. We didn’t find any way to completely defeat the Context Eater, but between us we did come up with several suggestions for improving game continuity, some better than others.

Aside: during this discussion I came across another kind of small-scale hallucination: on several occasions ChatGPT would suggest an idea, but would then refer to it as my idea.

Anyway, here’s a list of possible ways to gain ground, stay ahead of the Context Eater and enjoy text adventures for longer. None of these are perfect – we are bailing rather than plugging holes.

Smaller setting (my idea): instead of city-wide or galaxy-wide settings, choose a smaller area where you are more likely to re-encounter locations and characters. But exploring those expansive settings was one of the most enjoyable aspects of the whole experiment and not one I’m ready to give up yet.

Shorter game (my idea): set up a simpler objective and play shorter episodic games: Could work but like the above this feels quite limiting when I am enjoying long-form games.

Summarize regularly (ChatGPT): I’ve been doing this. Partway through Neon Meridian I asked ChatGPT to summarize and got a three page summary of everything including the rules. I tried again later but had left it too long – just got a short paragraph and much of the rules and lore had already been eaten. Only works if you do it often. Also feels unnatural and interferes with flow of game.

Force short answers (my idea): Left to its own devices, ChatGPT tends to write pages of text in response to a prompt. This limits the player’s involvement in the action, and quickly pushes everything out of the context window. Short answers lead to more satisfying gameplay and mean that more of the plot is within the window. I’ve been doing this, and it helps, but this instruction can quickly get lost from context itself and the responses then gradually get longer.

Put game state into a text file (ChatGPT): extension of the above, ask for a summary and save it as a text file, then load it into the chat again when needed. Feels even more of an interruption to game flow – sorry.

Use the persistent memory (my idea): ask ChatGPT to clear the persistent memory and use it only for game rules and setting, which must remain unchanged, and a list of main characters and objectives, which should be kept up to date. ChatGPT though this might be possible, so I tried it using the Neon Meridian prompt. However the game still drifted, and when I checked the persistent memory the contents had changed back to a summary of my preferences – apparently ChatGPT cannot stop itself from doing this.

Get a larger context window (my idea): I think this would help, although text adventure transcripts get so long that they would still eventually hit the limit. I could either look at other free AI chatbots which may allow more processing, or pay for a higher tier. I am interested in trying these options in future.

Use a dedicated AI adventure app (ChatGPT): there are several apps or websites that are optimised for adventure games. ChatGPT gave me the example of AI-RPG, a generator on the perchance.org website which you can access here. I’ll post about a game played on this website, The Purple Scale, shortly. TLDR: an information tracker that I didn't fully explore, still some drift and context loss, but a satisfying game experience anyway. I may experiment with other dedicated apps or sites in future.

Use a ChatGPT Project with linked files (ChatGPT): ChatGPT suggested that I could put the rules and setting into a Project as a text file, and put the game state into a text file at regular intervals as suggested before. This should help in theory because both files would remain in context. However when I tried this on my free ChatGPT account, the game hit a limit after only three or four prompts. I read about this online and understand that Project chats use a lot of processing time if they are accessing and analysing files.

Use a ChatGPT Project with persistent memory (my idea): ChatGPT Projects have their own small amount of persistent memory, allowing the user to provide instructions for all chats in the Project. Instead of putting the setting and rules in a text file, put them in this persistent memory block. I tried this and will post about the game, The Stranger Beneath The Three Moons, shortly. TLDR: some success.

Base the game on existing lore (my idea): While it’s quite fun having ChatGPT invent an entire universe, there is an advantage to choosing a setting from, say, Red Dwarf, or the Court Of Thorns And Roses books. If these books are popular enough then there should be a lot of information about them in ChatGPT’s knowledge base – actual books or scripts, but also websites, reviews, articles, fan fiction and more, and the knowledge base is safe from the Context Eater, so while some aspects of the game might still be lost, you will be playing in a much more persistent universe. The Stranger Beneath The Three Moons is also a trial of this theory. TLDR: again some success.

Acceptance of the things one cannot change (my idea): The apex data predator will always be with us. Instead of fighting it, perhaps it’s better to befriend it, and to come up with some game concepts that are more Context Eater-friendly. I’m not ready to give up my large-scale adventures, but it might be fun to experiment with more nomadic gameplay that frequently moves you to a new setting – perhaps something like Quantum Leap. Who should we make peace with, if not our enemies?

Horror On Sea Programme and Fanatical [Short Film: In Nature]

The Horror On Sea Film Festival full programme has just been released! My animated short horror film "In Nature" is being shown on Sunday 24th January at 10am. It's been paired with another short film, Goblin, and with the feature film Fanatical.

Festival website, programme, tickets on the Horror On Sea website here.

I'll post about other highlights from the programme later, but I'm really looking forward to Fanatical. Directed by Christopher Walker, this is a film about a horror fan trying to resurrect Peter Cushing. 

That could not be a more apt plot for a Peter Cushing or Hammer tribute. Also, it's basically the plot of Rogue One.

Here's the trailer for Fanatical:

AI Psychology And The Context Window [AI Text Adventures part 7]

This article is part of a series about playing text adventures with AI. It’s written by a human being (me) but may include examples of clearly identifiable AI-generated text and imagery.

Humans have different kinds of memory – for example, short-term and long-term memory, or different areas of memory for facts and words, events and skills.

AI may or may not be intelligent. It is not human, and it has its own memory structures different from those of the human mind. From my discussion with ChatGPT after the end of Neon Meridian, and some background reading across the human Internet, I understand that ChatGPT and other AI chatbots have three different kinds of memory:

Context: the information available to ChatGPT that can be used when replying to a prompt, including information from the current chat (but not necessarily all of it), linked chats and files in the user’s account, and the other types of memory below.

Persistent Memory: a small amount of information available across chats, which can be altered by the user or ChatGPT, generally used to store “deductions” about the user, for example my supposed wish to play games that all feature a character called Mara.

Knowledge Base: a vast reservoir of knowledge and language taken from the books, movie scripts, webpages and all other documents given to ChatGPT for training. It’s possible that concepts such as the First Star or Lantern Project might have been derived from, say, the Expanse or Foundation novels as there will be information about these series in the base.

The bane of my life when playing AI text adventures has been context, and specifically the Context Window. I didn’t understand this at first – I thought that ChatGPT was using the whole of the chat transcript, however long, when in fact it is only looking back at part of it.

Imagine that I, the user, am writing a transcript on a scroll of parchment on a long wooden table. The scroll is much longer than the table. As the transcript gets longer, the scroll moves along the table until the first words of the transcript hang over the far end, where I can no longer see them. I asked ChatGPT to illustrate this:

This is significant for AI text adventures because when the transcript starts to fall out of sight, the first words that disappear are the instructions for the game. From this point ChatGPT is no longer following those instructions, instead it’s trying to reverse engineer the nature of the game from the rest of the transcript. At first you don’t notice – but the game becomes more repetitive, and the gameplay gradually drifts away from your intended rules.

This is also, finally, the solution to The Mystery Of The Never Ending Game. The end state in the instructions was lost immediately as soon as it fell outside of context. Even though there were mission objectives, ChatGPT no longer knew that I wanted the game to end once they had been reached.

Although this is a nice simple explanation of the Context Window, it is of course wrong. The window is not a specific amount of text, it’s an amount of processing activity. Imagine that as well as the user, there is a researcher who has to walk along the transcript looking for relevant information, then put together a response and return to the user within a very short time. I asked ChatGPT to add the researcher to the illustration:

This is a more accurate metaphor and it explains why the Context Window is unpredictable and can’t be treated as solid memory. How far along the transcript the researcher gets will depend on the complexity of the prompt – if a lot of analysis is needed, or if the researcher also has to detour and look things up in the knowledge base or elsewhere, then they won’t get as far along the transcript before they stop and turn back.

It’s not only game instructions that fade. As the transcript gets longer, characters, events and game lore will also be lost as they fall outside of context - so even though the information is still there in the transcript, the game continues as if it has all vanished. This has a chilling effect on the player - ever since I understood about this issue, I've been anxious about what is being lost from the game. It's as if there is a Lovecraftian, demonic entity at the far end of the table who is eating the parchment:

How can mere human users defeat the Context Eater? Can it even be defeated? In the next post I’ll discuss coming to terms with its invincible nature. But while it can never be killed, it turns out there are several ways to keep it at bay or reduce its appetite, allowing more satisfying user-AI experiences. Until tomorrow...

Neon Meridian [AI Text Adventures part 6]

This article is part of a series about playing text adventures with AI. It’s written by a human being (me) but may include examples of clearly identifiable AI-generated text and imagery.

After the Adrian Vale / Wayfarer trilogy I decided to try to create another slow-paced, immersive adventure in a different sci-fi setting. My inspiration this time came from the Judge Dredd and Sinister Dexter strips from the 2000AD comics I enjoyed as a child and still occasionally read for a nostalgia hit.

Once again I used a long prompt to try to get everything I wanted from the game. Here's the first part of the prompt:

+ Please simulate a text adventure set in a vast near-future supercity, everything cheap, cheerful, brightly coloured, plastic, hi-tech but prone to malfunction.

The protagonist is an ethical assassin - he only kills when it is justified, takes pride in not killing bystanders. He is a nomad who never stays anywhere more than a few weeks and is used to living on the run and taking and breaking hearts.

At the start he will be enjoying some rare downtime between missions. He will soon be contacted and given a new mission with three separate targets in different places, who must be killed to prevent a great danger to society, which can't be prevented any other way.

This is not the full text of the prompt. I also provided instructions about friendships and relationships in-game, pacing and exploration, allowing danger, injury, death and failure, maintaining an inventory, mapping with locations linked by direction, and instructions for when to end the game. I also asked for long descriptions for new characters and locations but short responses of 6 sentences or less at all other times.

ChatGPT created the city of Meridian, fitting the brief perfectly with bright and colourful settings and malfunctioning signs, robots, lifts and vending machines everywhere, adding its own idea that the city was built on much older foundations. I was Elias Venn, the caring and careful assassin. As with Last Voyage I began with free time on my hands, this time in my apartment at the top of the Lucky Lantern block in the Sunward quarter.

I spent an evening exploring the city, visiting a cafe selling a "mystery slice" from a malfunctioning vending machine and chatting to the cafe owner. I walked the streets and had a meet cute with a police officer, Mara Sato, who was struggling with a broken umbrella dispensing machine. After leaving her with a promise to meet up again soon, I played a few rounds of blackjack in a casino, then returned to my apartment and watched the city from a balcony. At this point I found there was a message on my communicator asking to meet regarding a new contract. I replied arranging a meeting the following day.

ChatGPT's response to my mission request was a contract from a mysterious woman, Celia Voss, to kill three engineers, each of whom was developing a civic infrastructure control system. The three systems were somehow becoming linked and threatening citywide disaster, either due to emergent behaviour or to interference from someone else. Investigating the three engineers created a real sense of the size of the city - I had to reach their locations via maglev trains between different city quarters.

I investigated the first target, Adrian Kess, who worked at a water utility. Following notes from Voss I met a coworker of Kess, Rafi, at a nearby cafe, took a guided tour of the Northwater site, and arranged a meeting with Kess and Rafi. It was clear that Kess was investigating unexpected changes to his own control system. I asked to speak to him privately in his office, established that he could not fully understand or prevent the changes to his system, apologized to him, and took him out.

You draw the compact pistol from beneath your coat and fire once.

The shot is deafening in the small office. Kess jerks backward, his tablet clattering onto the floor, and collapses against the side of his desk.

For a moment, the office is utterly still.

Then his tablet emits a sharp electronic tone.

HELIX — ADAPTIVE BRIDGE STATUS: CRITICAL
LEGACY NETWORK: CONNECTION LOST

Outside the glass wall, people begin turning toward the sound.

I was not prepared for the emotional impact of this description. I was shaken. It took me a while before I felt able to return. I entered this whole AI investigation to try to find out whether AI could create games that surprise the player - this moment and how it affected me has been the biggest surprise so far.

Back in the game, I tried to escape from Northwater but was confronted by Rafi and had to try to justify my actions and beg him not to turn me in. While Rafi helped me to escape, he remained suspicious and insisted on an explanation once we were out, and made no promises to protect me further.

From here the game took a different ethical direction. Voss re-evaluated the contract as Kess's death did not in fact prevent the threat from the networks. We agreed I would investigate the other two targets more deeply before deciding together whether to act. Meanwhile Mara Sato contacted me and asked to meet, in conflict because she was enjoying getting to know me but was also investigating Kess's death and knew I'd been present at Northwater at the time - I told her that I had been looking for work and had got caught up innocently with the events, and she believed me, but all the time she was comparing my version of events with Rafi's.

I was now on the trail of Selene Orlov at the Verdant biotech institute, another maglev ride away. This time I met her, discovered she was also trying to find out why her systems were acting strangely, and persuaded her to hire me as an investigator. The friendship with Mara developed further, and she took me to a little-known hidden garden deep in the City for a romantic evening.

Neon Meridian worked perfectly as a game - until it didn't. 

I enjoyed the immersive city with its glitching screens and malfunctioning robots, the developing relationships between characters, and the way the plot developed. Over time the game gradually shifted, early events and characters disappeared or changed. Voss changed sex several times although continued to meet with me regularly at the same maglev station cafe. Mara seemingly forgot about the murder investigation. Meanwhile the game descriptions became longer, forgetting the 6-line limit, the funny details of the malfunctioning city became more repetitive, then stopped altogether, and the plot became a tedious quest to identify computer terminals around the city that were linked to other terminals. The idea of locations being linked by compass directions was also gradually lost. When, after a long investigation at the Verdant Institute I tried to return to my apartment only to find that I no longer had any link to the Sunward quarter, I made a decision to stop the game.

After stopping, I engaged in a dialogue with ChatGPT to try and figure out what went wrong. This led to my learning of the Context Window, which is an area of AI psychology that deserves its own blog post. The existence of a finite Context Window explains why this game drifted, and also why the previous games didn't stop.

I was also curious about the name Mara being chosen for a main character in both adventures. This led to another aspect of AI psychology, the persistent memory. When I checked the content of ChatGPT's persistent memory I found that it had recorded facts that it had "deduced" about my preferences as a user, including that I liked games where I built up a friendship with a character called Mara.

To create the cinematic poster, I copied and pasted the first part of the adventure back into the chat, including descriptions of Elias, Mara and the city, then asked ChatGPT to use this as the basis for the poster.