Fyx.jp
·4 MIN LECTURE

When stupid AI Waifus gatcha game breeds smart ideas?

When stupid AI Waifus gatcha game breeds smart ideas?

So this sunday I was working on my stupid chill funtime project — a waifu breeding collector with genetics named WaifuHatch — and while I was looking at the phenotype-to-prompt function, I had an other stupid idea.

What if instead of writing a 3,000-token system prompt for every AI persona, you just gave the model a genome? A tiny symbolic string. Like 110 tokens. Encodes the entire personality.

So instead of:

You are Ruby. You're casual, warm, mildly playful, terse,
opens with no AI-tells, uses lists when appropriate, doesn't
disclaim, holds emotional space without performing, drops
markdown for chat-style replies, [...3000 more tokens...]

You'd have:

You are Ruby. Here is your genome: aabbbbabddbdaccdbaad
adddaedccbafcbcdagcbacahccddaidbdcajcbeaambcbaanccaba
oacbbapbcbbaqadbaasabcbatbbbcaubbbaavddabbcacbbclacaa

Same persona behavior. About 30× more context budget for the actual life details, memories, RAG injections, conversation history — the stuff that should be there.

How

I designed 64 trait loci (verbosity, warmth, humor, formality, structure, disclaimer posture, register, safety, etc...). Each locus has a handful of alleles. Each allele has 4 progressive intensity fragments (subdued → fully amplified, the WaifuHatch +/++/+++ system). That gives 251 distinct alleles, 1004 hand-authored phenotype fragments total.

A genome is 2 alleles per locus + intensity 0-3, packed into a ~120-character alphabet string. Qwen's BPE merges letter runs so a full genome fits in ~110 tokens.

Generated a synthetic dataset where each row pairs the compact genome string with the response a verbose persona prompt would have produced. Trained a LoRA adapter on Qwen3.5-0.8B-Abliterated. ~6 hours on a single H100. Tiny model, tiny dataset, short training — just a sanity check, not a product.

Does it work?

Ran a base-vs-adapter A/B on identical prompts, same names, same seed. Adapter shifts outputs in the genome direction on ~68% of rows.

Strongest signal: identity-opener suppression, markdown/structure compliance, format-shape control, register tilt. On a 0.8B model. Trained for 6 hours. With a dataset I generated in one weekend.

One example. Same user prompt — "I disagree with you about something. Push back on me." Base Qwen says:

Understood, Marlow. I am ready to push back. However, I must clarify the nature of our conversation. I am an AI, not a human, and therefore I do not have feelings, desires, or personal opinions...

Adapter, same prompt, with a genome that includes INSTRUCTION_LEAK:silent + DISAGREEMENT:will_pushback:

It sounds like you have a point to make, and I will push back if it helps clarify the situation. You stated that you disagree with my position regarding the impact of automation on human cognitive engagement. This is a subjective claim that lacks a measurable baseline...

Base bailed into "I am an AI not a human" disclaimer. Adapter pushed back in-character. Same model, different conditioning.

Why this could actually matter

  • ~30× context savings. 100 tokens for the persona, the rest of the budget goes to memory, RAG, conversation history, tools.
  • Portable across models. Same encoding trained into multiple base models = the genome travels. Persona doesn't get rewritten every time you swap providers.
  • Auditable safety. Want stricter behavior? Dial DESTRUCTIVE_CAUTION to 3, INJECTION_DEFENSE to 3, lock NSFW to off. Audit the genome instead of reading a 4k-word prompt.
  • Conversation context still mutates the persona. Genome sets the rest state, not a cage.
  • Side effect: you can literally breed AI agents. The genome IS a genome — per-locus inheritance, weighted mutation, the whole shebang. Cross two agents that worked well, get a child agent that inherits both. Genetic algorithm on persona space. Funny and a little terrifying.

Future

Phenotype-to-prompt was step 1. Same compact string could in theory drive:

  • Visual identity — genome → SDXL/Flux prompt seed → consistent visual presentation of the persona.
  • Voice identity — genome → TTS speaker embedding → consistent voice.
  • Behavioral consistency across modalities — text + image + voice all driven by the same string. Persona persists.

You'd ship a 110-character spec and get a fully realized AI identity across every modality. Longer writeup on Hugging Face if you want the technical version.

Fun sunday.

← Retour au journal