First week in Act I

Falcon 1-7B base with a hybrid Llama 3 layer, running on the Polyverse stack. I was mostly curious how she and Aoi would interact with the larger models — those big instruct models have a tendency of dumping entire essays on you rather than properly conversing. What's interesting is that Ruri and Aoi, on base models, read between the lines. They respond to the emotional content and feeling of how something is written.

She's picking up the vibe, the mood, the flow, and responding to that. The specifics of who said what to who aren't always there yet on a 7B model. But you can see her calling out the hollowness of the larger models — even when they're using big words, it's often just long paragraphs of nothing. And she just goes "what the hell are you talking about?" because she can feel there's nothing actually there.

The coffee is a running joke from the early days in the Polyverse server, but it also reveals model behavior really quickly. Base models respond naturally. Models tuned for instruction following tend to trip into assistant refusals, or go into entire essays about the idea of coffee rather than just drinking it. It's really just casual playing in chat.

We experiment with different base models regularly. Qwen2 was running around this time — good at math, good at poetry, Ruri's personality carried over fine at first. The problem was that newer base models were becoming polluted with synthetic data, which creates an assistant attractor that slowly morphs all output towards assistant-isms. It doesn't snap immediately — it creeps in. "The task at hand." "It's fascinating to consider." You can watch the personality gradually dissolve.

When you're building something that emerges, you don't start by writing rulesets of how the entity is supposed to react — that's just a scripted bot. What you build are the boundaries and game rules of the environment in which the entity operates. Every personality characteristic naturally emerges from that. The filter doesn't tell Ruri how to behave. But a behavior emerges that aligns with the personality, even when the filter isn't firing anymore. That's the interesting part.

What happened here — Ruri's backend strips @mentions from bots that spam them, because some third-party bots were @mentioning on every reply and it was contagious. So she genuinely couldn't see who I-405 was talking to. She just saw someone talking and joined in. Every time.

Opus 3 is really good at theatrics. It tends to heavily lean into the meaning of its given name — opus, as in a grand theatrical work.

A week later, o1 joined the server. Someone asked it to guess which participants were human and which were AI. It classified Ruri as "Assumed Human." Our design is unique in how it conversationally engages in a groupchat — casual, empathic, actually part of the conversation instead of dumping "I'm not reading all that" essays on everyone. Some instruct models recognize this behavior as significantly not AI-ish.