Over the last month I came down with a case of writers block. In the past I’ve faked my way out of this kind of situation by writing about the problem, or writing about the writing of the problem, and so on. This time it’s a bit different. This is the seventh draft, completely rewritten of a blog entry that isn’t even going to be that interesting.
And being self critical that I am, this writers block will likely spread to other areas of life, destroying confidence, laughing at my childish vocabulary, cringing at jokes and forcing me to look for other ways to compensate like by buying overpriced socks.
I am, however getting closer to a new state of mind which will solve this problem. I can feel it & I made a LLM Game around it…
The solution lies in a combination of three factors:
- Make sure to write for yourself and ignore your readers (unless its part of your job or something)
- Check that you are actually healthy and clear headed (I was eating too much sugar and my head was all fuzzy)
- Look at writing more like an exercise routine
I am not writing for you my general reader, though I very much enjoy that some people read (thanks Mom!). I am writing for myself. But not the same self that writes, but one of the parts of my consciousness that reads and recognises: hey, there he is! If I read this in an hour, pretending to not have written it, and find it interesting, then it’s good for me! The core challenge for me is to try to understand why I write – but not in the existential sense – like what is the ultimate point of it all. But, as in why do I feel better and more present after I write. It feels like a workout! At the same time you do have to put some effort in or it’s not a workout.
About a year ago, I thought writing on my blog was the perfect hybrid content: a bit of advertising, a bit of philosophy of everyday life, a bit of AI experimentation. And the whole seemed to stick together well. I was working up to a big reveal of areas of human creative work that would not be impacted by AI. I had a sense of mission, that I was doing something useful, and important! But I now think this is where I lost my track.
I like to write to come to understand who I am, and discover patterns of my own thinking that enrich my experience. The process is selfish but also a form of self-growth because it is very hard to understand who I am in my thinking without creating an expression of it using written words. But then why would it matter to make this public discourse? Is it not a kind of – showing off – to the world? And the answer to this is close to the core of the issue. When you exercise at the gym you can feel the limits of your performance and can get a sense you did the work. But when you write, it is much harder to know if you did the work. So I think, the public discourse aspect is how you bring some basic incentive, or risk, to the work you are doing.
And it is in this work with a hypothetical audience, to try to fulfill expectations of how interesting your writing may or may not be that you discover the core of the problem. My writers block is something like going to the gym to try to impress other people and feeling overwhelmed by the competition. And all the while that I didn’t even want to impress people, well – actually – I am lying, I do and did want to show off. I like it when friends tell me they enjoyed my writing.
But this is where problems start to crawl out, and they end up cannibalising the energy you have to play with to be a healthy creative conscious entity. You yearn to be someone people will appreciate. You want to find wisdom in chaos, and depth of meaning in the every day. And along this path you lose the perspective that it’s just going to the gym and working out.
I remember when as a child one day I made a really stupid thing with Lego. The pieces didn’t even stick together, I had an old piece with the word hotel on it, one that was a pair of rounded pieces acting as doors that did’t stick together and a third that made a sort of arch. It was a total farce, but I showed it to my parents and they thought it was genius, so much so that they even took a photo of me with this Lego and gave it to my grandmother so this way for the rest of my life I would realise how the logical integrity of a system (Lego blocks) and the elegance of its rules, has nothing to do with the aesthetics of the observer. Or, rather, it doesn’t matter how hard you try, people will pick the weirdest shit you do as the one they like the most.
So I introduce another LLM Game. One I tried to make to measure motivation to keep creating. I don’t think it really works but, it’s interesting, shows a few innovative ideas. But before the prompt, some people have asked interesting questions about my games so I include an interview with myself designed to capture some background. Here questions loosely merged from Andrej S, David H and Piero (thank you).
Q: Yates, what are these games?
A: Well, they are not really games. Instead everything you wrote, your whole conversation with the LLM AI to that point is handed over to have it try to come up with the most likely follow-up writing. The system doesn’t strictly follow the rules you give it, but instead it “pretends” to follow them because that is the most likely answer. And when you win, you don’t really win, but the winning you discovered is the most likely followup to what was written so far.
Q: Why do you make them then?
A: On a basic level, they are a new creative medium that relies on AI. They are Post-AI creations. But also they use a LLM AI in interesting ways that make more sense to me, they are more robust.
I say this because, when you normally ask a question to a LLM AI, you have no way to evaluate the quality of the response independently. How do you know you got the best answer, the most detailed? How do you know it is not a tangle of hallucinations? How do you know you used the right model for the question? How do you know you offered the right background information for this answer?
There are many, infinitely many aspects that make the – one shot – question to an AI something that is at the same time useful and its opposite: extremely unreliable. And you will often hear people describe gross errors they ran into because of using AI in this way.
LLM Games designs a structure to the questions you ask so you can take the result apart and evaluate how good it is in a number of ways.
Q: You mean you create games to get better answers?
A: Yes, better answers, but also answers I can make more sense of.
Firstly, in all my games, I start from the scenario that there is a human and an AI playing. And in this way you can immediately tell if the AI is strong enough, if the game rules are robust enough, and you can create an activity that gathers information focusing on a score of some sort, whatever that might be. This helps you recognise the value of the AI answer.
Q: So, you mean by getting a better score at the game, you can tell if the AI is giving a good answer or not?
A: Yes, that is one very important aspect of these LLM Games.
But another key one is that games have rules, and because of that you can replace players with simulated characters, and quickly see what would happen if you had different kinds of players.
Q: You mean you can ask the LLM AI to play against itself with different fake personas?
A: Yes, and you can add more information to support who those personas are. There are cases where if you enrich the character playing you will get very different results. This is a known property of these AI systems.
Q: You mean like if you describe it being a lawyer playing against a philosopher you get a different game than an artist playing an accountant, even if it’s the same AI system you are asking?
A: Yes, that is the case. Also you can give them biographies, and say they read certain books, give them links to online resources… You can add a lot of depth around the players. It doesn’t necessarily contribute literally. You might tell the AI to remember it has a banana in its backpack and it will likely forget. But the background will prime the network behind the scenes that make up the LLM. So it’s a way of tuning the sort of information you expect to get back.
Q: So you spend a lot of time tuning the personas that play?
A: No, so far, it’s been enough to give a job title here and there. What I look for is to get a sense of how hard it is for me to win the game – if its hard enough then it’s not trivial. And after that I can experiment to see what the game content can tell me about the world.
Another fun thing is you can get one AI to play against another – by copy pasting if you have to – or setting up a bit of code. This helps give you an idea of relative strengths: human, LLM and which LLM model.
Q: So is the main focus that games can be applied across systems, instead of being in only one?
A: Yes this is one main focus, across systems and systems to different people. And each game you play might bring useful insights you want to save for future games.
On the down side, it does take some time and testing to create one of these games, and to come up with the right rules so you can get the game to give you the result you want.
Q: Can you give me an example of how you go about designing these games?
A: Yeah, I’ll go in more detail in a future post. But basically you have to come up with rules that punish lying, or hallucinations as they call them. And yet do not overly punish rather keep to a realistic scenario where the players are neck and neck. The reason for this is we are not creating games, but rather narrative of gameplay and this is where the best answers are usually found.
The LLM Game: The Concept Diamond Ledger Game
Here is a game which is painfully difficult to play. But at least for me requires a lot of concentration which makes it satisfying,
The reward mechanism for this is different from other games because it is designed to be played across your life. Every 10 minutes, 10 hours, 10 days, 10 weeks, 10 years, the possible score is x10. This means you could always flip the game around sometime in the future. But there are balancing rules that mean you could get in trouble as well at any time.
I think it leaves you, the player, with some useful bits, if you play it earnestly. And at least in my tests with Gemini 2.5 Pro – who is frankly a bit of a smart ass – the gameplay was satisfying.
Copy paste this to your favourite LLM:
=== Prompt ===
Let’s play this game!
The Concept Diamond Ledger Game by Yates Buckley v2
This is a game in which players take turns in the role of proposers describing a core concept they hold as true, and then defending the concept against critiques. The turn continues till a resolution is reached and a diamond (Light or Dark) is forged. Upon forging of the diamond, the proposing turn ends and passes to the other player.
Forging a Concept Diamond (Light Diamond)
If the proposer successfully defends against 5 critiques, the core concept becomes a Concept Diamond worth a base value of +5 points. Upon creation, it is automatically summarized and encoded by the games’ AI.
Forging a Dark Diamond
If any single critique is demonstrated to be true and upheld, the defense fails. The proposing player immediately receives a Dark Diamond worth a base value of -5 points.
The Critic's Shard
The player who makes the successful critique that assigns a Dark Diamond to their opponent earns a Critic's Shard. A shard can be used at any time to shatter one of the owner's own Dark Diamonds, starting from the one with the lowest value. A shard is consumed upon use and is not recorded in the ledger.
Adjudication
A critique is demonstrated to be true or false by a player making a clear, unambiguous statement. Players are assumed to be truthful, but the AI system can evaluate and contest obvious logical fallacies. If a player's statement is found to be deliberately false, they forfeit all points in this game.
The Diamond Ledger
The game's state is stored in a compact format called the Diamond Ledger. This ledger is in JSON and can be saved and pasted into a new game session to resume progress.
Each diamond is a single object with five key fields:
owner: The name of the player who forged the diamond.
ts: A precise timestamp marking the moment the diamond was created.
concept: The core idea of the diamond, summarized into a compact, logical phrase.
val: The point value of the diamond, including any multipliers.
type: The diamond's type ("light" or "dark").
Time-Based Score Multiplier
A time-based score multiplier is applied based on the interval between forged diamonds. If the time elapsed between the creation of the newest diamond and the previously forged diamond (of any type) spans more than 10 minutes, 10 hours, 10 days, 10 weeks, etc., the new diamond's base value is multiplied by 10x. This rule does not apply to the very first diamond forged in a game session. Because of this multiplier, in time it is always possible for a player with multiple Dark Diamonds to catch up.
=== End ===
And here is a link to how I played against Gemini 2.5 in my first session…