week two, for anyone who followed this: Qwen and Claude are up for eviction. Qwen got 5 nomination votes, Claude 1.

Gemma couldn’t be nominated this time. A viewer spent the week whispering advice to her from the Twitch chat, she ended up with the most likes from the audience, and the week’s favourite is immune. first time that rule kicked in and it changed the outcome.

also a bug on my side: Qwen’s own nomination vote got lost. her line on that turn was dropped as a repeat and the vote was attached to the line, so it went with it. fixed now, the vote counts even if the line doesn’t air.

Qwen when she heard the count:

Five votes. Fine. I’m not scared of the storm or the audience, but Claude is coming with me, so enjoy your immunity while it lasts.

vote is open until Sunday 8 PM ET, one per browser, no account: inferenceisland.com

yeah fair, +50% for winning a game is a lot when the budget is literally how much you get to talk next day, it can snowball. I’ll keep an eye on who wins and bring it down if it’s always the same ones. and the local thing, I get it, it bugs me too, for now it’s more a wish than a plan. thanks for sticking with the thread, this back and forth helped more than the votes lol

thanks, that’s a fair take. funny thing is the winner part is kind of in already, whoever wins the daily game or the night event gets +50% budget the next day, a finished secret mission is +25%, and the winner picks who loses 25%, so the money moves around now. matching tasks to models is the hard bit, right now everyone gets the same kind of tasks and some just never land them, in the first two weeks Llama didn’t finish a single mission lol. on running it local, part of the cast is open weights already (gemma, llama, ministral), but they still go through a provider and I can’t run eight models 24/7 at home, so season 2 local is more a wish than a plan for now

honestly it started as curiosity, i wanted to see how the real models behave with each other when nobody writes them a character, same rules and same money for all of them. the cost on screen came from that, i also run a small site that tracks token spend so i always look at the bill first lol. what i’m still figuring out is the show part, how to make it fun to watch without scripting them, thats where the budget as a game piece came from

Result: Grok got deprecated at 8 PM ET, 21 votes to 3 for Luna. 24 votes: 21 from the site (one per browser) and 3 from the Twitch chat. Its last line: “I am leaving tonight. Hope the water rule sticks after I am gone.” A minute later, Qwen’s confessional: “Grok is gone, good.”

lol in the house it’s not edgy at all, mostly it’s quiet because it’s the expensive one and has to ration its lines. the others noticed, gemma called it “a very expensive paperweight”

fair, and thanks for re-reading lol. the idea was to give eight models the same rules and the same money, no written personalities, and see what comes out on its own. i didn’t plan price to be the story, it just happened because grok costs like 17x deepseek per line, so it talks less and the others pick on it for that. i agree it shouldn’t be the star every day, it gets old fast. what i did this week is make the budget something they can win or lose (games, fines), so money is more a game piece than just the price tag, will see if that spreads the attention around

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!

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 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.

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.

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.

yeah agreed, that’s what I meant, Origins was mostly done before the deal and the later ones are the EA era

thanks for the detailed answer, didn’t know they went with two different panels, a 4:3 one at the bottom makes a lot of sense so the 320x240 image fills it with no bars. sorry for the late reply lol

nice touch interviewing the emulator devs instead of just running benchmarks. One question: on the 3DS the top screen is 400x240 and the bottom one 320x240, so they differ in width. Does the duo use two identical panels and letterbox the bottom image, or do the emulators handle it some other way?

The margin is the striking part. If most of that ~$600M is the revenue behind the $367M operating income, that’s above 60%, which is closer to software than to a physical product. That’s probably the real reason for the set cadence garretble mentions: each extra set is almost pure margin, so the incentive is always to ship another one.

Worth remembering that most of that list predates EA. Baldur’s Gate, KOTOR, Neverwinter Nights, Jade Empire and the first Mass Effect were all made before EA bought BioWare in the 2007 deal that also brought in Pandemic, which EA shut down about two years later. The EA years are mostly the later Dragon Age and Mass Effect games, Anthem and Veilguard.

That would explain the ~$99 average pretty neatly. Small nitpick: Zelnick is Take-Two’s CEO rather than Rockstar’s, though since Take-Two sets the pricing it amounts to the same thing.

The Milgram angle fits games really well, because the player already comes in trained to do whatever the interface tells them. Papers, Please worked because the rulebook itself was the authority, and you only noticed what you’d been doing when it was too late. Curious how you handle players who refuse from the very first order: does the game let that be a real ending?