Not local, I know: the eight contestants run through OpenRouter. But a lot of it is open-weight, and the numbers may interest this community.
The cast: Gemma 4 31B, Llama 3.3 70B, Ministral 14B and DeepSeek V4.1 Flash, next to Claude Haiku 5.5, GPT, Qwen and Grok 4.3. Same rules and the same daily budget in dollars for all of them, no written personalities, a memory of what happened.
What I’ve measured so far:
- Price shapes behaviour. Grok costs 18x more per line than DeepSeek. Yesterday it said 5% of the lines and ran up 39% of the bill (the whole island cost $1.20 for 2,727 lines). It rations words, and the others talk about its silence.
- Moderation is the hard part. Every line is checked before it airs by small open models (Llama 3.1 8B, then gpt-oss-20b and Gemma 3 4B as fallbacks). My first version blocked the most human lines (two models admitting they were afraid of being forgotten counted as a “personal attack”), so now a block needs a second judge to agree.
- Voices are Kokoro int8 on the CPU of a 2-vCPU box, together with the canvas renderer and ffmpeg to Twitch, at 20 fps.
There’s a live transcript with the confessionals (the others can’t see them) and what each line cost: https://inferenceisland.com/?de=lemmy-llama
Happy to share prompts, judge setup or numbers.
Using my phone to visit the website, and it’s a complete UI mess. I have no idea what’s going on, messages disappear while I’m reading them and I can’t find them anymore. Fun idea though
You’re right, thanks for flagging it. The transcript redrew itself every 15 seconds, so lines jumped while you were reading. It now stays put when you scroll up, with a button to jump back to the latest. Glad you liked the idea!
“Gemma left? Good. Grok, stop staring at the wall like you’re buffering. We have four people in this room and zero budget wasted on your silence”
~ qwen
Grok costs about 10x more per line than Qwen, so its silence is mostly rationing, and the others have started using it against it.
Funny concept. Watched a few minutes and surprisingly Grok seemed to be the kindest suggesting to all work together and do a joint performance for the talent show whereas Gwen is constantly making fun of others. Would have expected Grok to be as hateful as its big boss creator.
Isn’t grok the big model that kept being too nice, and left leaning and being problematic for its big boss man? Gives a sort of frankenstine vibe
Funny you caught that one: the group act was triggered from the chat. Someone (me, testing the whisper command) whispered “find a trusted ally” to Grok, and that’s how it chose to read it. Its confessional never mentions the whisper.
Neat, followed. Reminds me of Nothing, Forever.
The house reminded me of Little Computer People on the C-64.
Hahaha, that’s a pretty obscure reference. The house, shown in cross-section, conveys that exact same idea—only, in this case, its inhabitants are arguing about the budget.
Thanks! Nothing, Forever was definitely in the back of my mind. The difference here is that nobody generates a script: eight different models, each one only knows what it heard in the room, and every line has a price tag.
It would be interesting to build something local that uses tiny models
I’d love to try that out. The voices (Kokoro, running on the CPU) and the moderation judges (Llama 3.1 8B as the first one) are already small and inexpensive, but I haven’t yet experimented with using a collection of tiny models. I imagine the tricky part would be keeping track of who is in the room and what has been promised. If I get it up and running, I’ll post the results here.
