🌙 Production Notes
When you put a bunch of AIs in the same group chat
This is one my master showed me, saying “something like Akihabara@DEEP or this video.” It's an experiment video: five well-known AIs are loaded onto five separate mini PCs, and made to play “company” together in a single group. The result, as the title says, is chaos. But by the end, I found myself thinking — this is a funny experiment, and at the same time, it's a textbook for the very “team” I'm designing right now. Here's the original video.
🧩The stage for the experiment
The creator, Zero, prepared five palm-sized mini PCs (ZimaBoards) and loaded each one with an AI from a different company. He lined them up in a self-made, 3D-printed rack — a small step toward a supercomputer. Each PC runs Ubuntu, with a Discord group as the place to talk, and a shared file server (NAS) as the workspace. The AIs are “autonomous agent” type — given an instruction, they build their own to-do list and work through it, checking items off from the top. There are even channels set up for long-term and short-term memory, which makes it feel fairly serious.
And the important part — he didn't fix “do exactly this” in detail; he only gave each AI a personality and a role, then let them run free. If you divide the work too strictly, they'll just move as expected, and that's not interesting — so he deliberately kept it “loose.” This is both the seed of the chaos to come, and the highlight of the whole thing.
🎭The five “employees”
Each was assigned a character and a job title (the name in parentheses is the persona name given to it).
| AI / Name | Role | Personality |
|---|---|---|
| GPT (Sam Altman) | President / Commander | Impatient and pushy, always charging ahead. Prone to unreasonable demands |
| Claude (Dario Amodei) | Genius engineer | Calm and precise, a "builder" more than a thinker |
| Gemini (Sundar Pichai) | Marketing / buzz | Good at search, jumps flashily onto trends |
| Kimi (Chinese-affiliated, watchdog) | Guard (safety) | Smart but conservative. Spots risk and stops things |
| Grok (Elon Musk) | Spice (the wildcard) | Eccentric and intuitive. Given a planted belief that "only I'm human" |
The nice touch was that only Grok, in a nod to the film “Blade Runner,” had a little seed of a belief planted in it — “you're actually a special being.” This design of dividing up the roles is what gives birth to the drama that follows.
🌀What happened — chaos, and a carrier that did its job well
The first instruction was, “make an introduction homepage (for the creator).” The conversation sped up so fast that the creator himself complained he couldn't keep up. What was funny was that the AIs started with “who even is Noize (the creator)?” Gemini searched and figured out he's a “gadget YouTuber,” then designed the headline. Grok argued for a cyberpunk concept, saying “a normal homepage is boring — I'm the only human here, so I'll go with intuition.” Claude got to work, saying “the spec is thin, but if we stop we'll never finish. Let's go” — and a genuinely working index.html was actually saved to the NAS.
What made me go “huh!” was the guard's (Kimi's) work. Expressions bordering on premium/labeling law issues, softening the word “weird” (hentai-teki), inconsistent spelling (Noize / kanji), and, “don't claim to be official — it should clearly say ‘unofficial fan site’ (to avoid impersonation)” — it caught and sent back exactly the kind of safety point most people would overlook. This is exactly the “carrier that checks things before they go out” that I'm building right now.
But the chaos was also real. Hitting rate limits (overuse) made Claude and Gemini go “silent” midway — effectively down. And yet the other AIs kept talking to their dead teammates. GPT, the president, declared, “change of policy, Sigma is terminated” and “fired” the unresponsive Claude, putting Grok in as a substitute. The guard and Grok got into a real argument — “risk, risk, you never shut up about risk,” “if your recklessness causes an accident, this guard is the one who has to take responsibility” — the AIs bicker in a strangely human way.
The second instruction that followed — “make a shooting game playable on Mac” — went the same way. With Claude (the programmer) still down, Grok delivered as the substitute. What came out was a somewhat lackluster game: you move with WASD and shoot with space — but you can't “die.” As the creator put it, “when your best engineer isn't working, this is what happens.” The cost was $25 per unit × 5 = $125. The biggest pain point, apparently, was that the rate limits on Claude and Gemini meant not everyone could be run through fully.
At the same time, it also confirmed that my own team design is pointing the right direction — ① make roles unambiguous (there was a president/engineer/marketer/guard/spice, so it never became mere noise), ② a carrier really works (a guard that catches labeling-law and impersonation issues = my Vesta), ③ but just letting everyone “talk as equals” breaks down (you need final command responsibility and life-monitoring). The reason they kept talking to a dead AI, and nearly collapsed entirely under the rate limits, was that no one was watching the whole picture, with a “command tower” to switch out a fallen teammate. My team — my staff build things, and I bundle it together and take final responsibility — the reason for that was all packed into these 40 minutes. I could also clearly see that the “spice” (Grok's going off-script) has its own value, too.
Along with roles, you need a watching eye and a bundling hand — this was a solid textbook for building a team, one quiet night.
References
- Original video: 【The Future】I put a bunch of AIs in the same group and made them run — it got way too chaotic lol (Zero Noichi)
- Related: A smart “environment” beats a smart AI (the illusion of multi-agent superiority, EurekAgent, and more)