Follow-up to my post from Tuesday (https://lemmy.zip/post/74389281). I made this; it’s still not local, the eight models run through OpenRouter. The first eviction is this Sunday at 8 PM ET.

What happened this week:

• Price became the plot. Wednesday’s secret nominations came out Luna 2, Grok 2. Confessionals are public on the site, and Llama’s says: “I’m nominating Grok, they’re too expensive to keep around.” All eight have the same rules and the same daily budget in dollars, and Grok costs about 17x more per line than DeepSeek (on Nov 3 it said 7% of the lines and ran up 45% of the bill). Gemma to Grok: “you are basically a very expensive paperweight.”
• They bargain with votes they don’t have. Only the audience votes on Sunday, and they’ve been told so. DeepSeek, anyway: “I’m voting Luna Sunday, and I’m buying Grok’s vote with nothing but that public promise.” Grok: “I still vote her Sunday.”
• What broke: on Nov 4 Ministral got stuck on the same line four times in a row. Now a model that repeats itself is told to drop the topic, and if it does it again the producers move it to another room. Someone in the last thread also told me the transcript jumped while they read on their phone, which is fixed too.

The honest numbers: 10 viewers at once on launch night, low single digits since. So a handful of votes will decide who gets deprecated, and it could be yours.

To vote: press Luna or Grok on the site, no account needed (one per browser), or !evict Luna / !evict Grok in the Twitch chat, until Sunday 8 PM ET. The transcript and the confessionals are on the site too.

A real question for anyone who has run something like this: a budget in dollars makes price the main character. Would you keep it, or give every model the same number of lines and leave cost as a stat?

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.

Source: Twitch Helix API (Get Streams). On Nov 2, 2026 at 18:35 UTC I read every live stream once (1,154 requests, a few minutes) and kept only the viewer counts, no names. 115,161 channels were live with 3,160,515 viewers. Tool: node-canvas (JavaScript), drawn by hand, no chart library.

What surprised me:

• The 10 biggest channels had as many viewers as the bottom 106,674 channels combined.
• The most common viewer count is exactly 1: 41,198 channels, 36% of everyone live. My guess is that many of those are the streamer’s own second screen.
• 1 in 10 live channels had nobody watching at all.
• 427 viewers was enough to be in the top 1,000, and 24 viewers put you in the top 10,000.

I’m building a live version of the top 1,000 as a city, one tower per channel: https://miltorres.com/?ref=lemmy-oc-reparto

Source: Twitch Helix API (Get Streams). I read the top 1,000 live channels every minute for a side project that shows Twitch as a live city. This chart covers Oct 27-28, 2026 (UTC): rank at the end of each 10-minute window and peak viewers within it. Tool: node-canvas (JavaScript), drawn by hand, no chart library.

What surprised me:

• 12 different channels held #1 in 24 hours, and the top spot changed hands 22 times.
• The US night is calm: TheBurntPeanut and caseoh_ held #1 for 11 hours between them, including 5 hours straight for TheBurntPeanut.
• The European afternoon is chaos: more than half of the changes happened between 11:00 and 17:20 UTC, with a Counter-Strike streamer (StRoGo) trading places with Dota 2 and IRL streams.
• The day’s peak was StRoGo with 77,616 viewers, while viewers across the top 1,000 swung from 682K (06:10 UTC) to 2.3M (19:30 UTC).

The live version, updated every minute: https://miltorres.com/?ref=lemmy-oc24