Here’re some more pics/video from SLOP FIGHTER. Up top is a real good go at battle flow, having straightened out much of the semantic graphing, finally. There are still some weird grammatical errors, but I am sure once I dig into those I will find it is the LLM making them.
The PVP screens:
So like you see how it works yeah? Here’s it running on Raspberry Pi 5 in Gameboy mode on my dirty-ass computer desk:
And here’s it connected to the fully compiled, basically ready-to-go game build on my computer:
I’m getting pretty fucking bored of LLMs though. I have other ideas for them which I probably just won’t bother with for a while. I’ve got something else to write about tech, Australia, and handling robot overlords, but after that, I’m just going to get back to writing some fucking books.
It’s a narrative battle simulator. In the game, random animals from all across the animal kingdom are mutated by one of eight special types, granted powers that befit their types, and instructed to fight each other. You give the commands and your mutated lil fella carries them out for you. It’s based on text. It’s a text-based game.
There are types ranging from elemental types (FIRE, WATER), to more unusual types (COSMIC, SHADOW), to aberrant failures (MUTANT). Each personality type is built around deep semantic graphs of animal heritage and mutation variety, and each type has its own grammatical quirks, its own manners of speech, and its own strengths and weaknesses. I have finely-tuned eight LoRA adapters for one small Qwen LLM that are each packed with useful words that help guide your monster through communicating actions, movements, and announcements that relate to their character. Birds will talk about their wings. Wolves will talk about their fangs. Snakes slither, tigers pounce, but more, they will talk about their mutation, too. An EARTH horse will kick with the weight of ages. A COSMIC gorilla will break reality with a wave of its arm. A FIRE mamba will spit flaming venom. There is a massive degree of uniqueness and variety to the responses. There is actually too much for the LLM to handle.
These are LLM-generated moves. They don’t always ‘strike’, but I guess falcons are more inclined to fight like that.
Your monsters will take damage and respond, they will mock the opponent, they will react to status effects, to the amount of damage taken, to victory and defeat. You can even feed them between battles (still in beta). I have developed versatile syntactic patterns (lots of sentence templates) to make all the words fit, then the LLM spins them together, adds its own flavour, and spins them back out.
There are two modes, CPU and PVP. CPU is entirely local. PVP takes place entirely over Bluetooth. Yes, you can play with your friends!
It’s even the most modern, up-to-date application of Bluetooth I could manage. It uses modern Bluetooth Low Energy as a primary and classic RFCOMM to make connections if BLE fails. It uses dbus-fast, a modern Bluetooth library, to improve connectivity, and I’ve implemented cryptographic handling of the messages between to prevent interference. It should even work on Windows and between Linux/Windows machines (still in development). At battle start, a VERY rough distance measurement is taken, and that sets the game environment. Your monsters have room to move around the battlefield, and you can direct their movements about it. You can MELEE attack from CLOSE, or RANGED attack from FAR. You can MELEE attack from FAR, too, and do less damage!
There is a lot of variety to the combat. I’m talking things like type effectiveness, status effects, misses, critical hits, and animal advantages (like predator/prey weighting), that all affect how much damage is given and received. For example, SHADOW types are weak against FIRE, but relatively strong against everything else, and don’t take damage from PHYSICAL non-fire attacks. They suffer from low HP to compensate. The CUSTOM COMMAND option also incorporates calculations like creativity weighting, so use that to your advantage.
Oh, and there are fully developed expressions for each monster type. Watch as your new pet gets hurt and fights back!
The game engine fundamentally relies on the English language, and the vagaries of a Chinese-made large language model, to function. Not every sentence produced by SLOP FIGHTER is accurate, correct, or even complete. The LLM works hard to chop and mix words, but the Qwen model I’ve used is a quantised 1.7B version. It’s just a lil dummy. It knows not what it really does. I wanted to build this whole project on top of the smallest LLM I could conceivably use for the job. As such, SLOP FIGHTER will run on almost any Linux machine quite well. It will even run on a Raspberry Pi 5. In fact, it runs BEST on a Raspberry Pi 5, due to their suitably advanced Bluetooth chips. It is slow on the Pi 5, but that adds to the old Gameboy-style charm.
Here’s a demo of it running on my computer:
I’m still pruning the datasets for grammar and contextual accuracy, but it’s p much done. One factor I can’t account for is interpolation from the LLM.
To me, this project kind of demonstrates how I see the world. I have, in the past, visualised words sliding into place based on factors like semantics and statistical likelihood just in time for my mouth to say them, or my hands to write them. I wouldn’t at all suggest that is my whole approach, but I understand it. Also with SLOP FIGHTER are smooth, fluid animations that help the words slot into place and the game carry on. The whole thing is one careful, intricate balancing act threatening to spill over into chaos and madness. Just like the world we live in.
This game will be released on Steam and itch.io in the coming weeks. I am bad at marketing so we’ll see how she goes tbh. I also pretty firmly believe this is the sort of thing Steam would do with Half-Life 3.
I’ve made an LLM-driven web browser. I’ve named it Zenbot. It’s just a mathematical word-generator with a set of word-tools, let loose on the internet. It’s sort of built to translate commands into actions, directly, using the Python coding language, which I’ve noticed is basically English.
I’ve described it like this: “Synchronous communication becomes asynchronous communication in an elegant double-helix of English language-powered Python interpretation driven by you, the user.” as I noticed two types of Python functions, sync and async, could be utilised together when an LLM was involved in the mix.
It’s sort of like if a whole restaurant shared one brain, which was passed from customer to waiter to kitchen and back again, and at the end of the chain a customer has their order. If you ask Zenbot to ‘search for pizza’, it will search for pizza.
There are, of course, lots of advantages to doing it this way, which I focused on, like:
Efficiency
Speed
Capability
Versatility
Security
As I don’t use any traditional method. I believe the LLM should be modelled on the task, not vice-versa.
I am of the opinion that Zenbot demonstrates how mainstream approaches to LLMs are changing. Small, tailored models are the future for operating untold new and old technologies. I still do not know if they should be writing words that mean things to humans. But it’s a brave new world!
I’ve got lots more ideas for projects, and Zenbot is getting some improvements as we speak. likecommentsubscribe
Pretty proud of this one. AI slop does have a purpose, if used effectively, and in consideration of its limitations. I guess it’s a medium, and one that equalises the availability of resources for anyone.
If anyone’s willing to invest, I’d set up a studio and produce all kinds of wild shit.
