> These prompts are often enormous and contain a *lot* of entropy. Ok, several remarks. (1) White noise has a lot of entropy. Someone’s hard drive has a lot of entropy. Is it useful to take some Dunder Mifflin employee’s hard drive, use it to prompt an AI, and then have it exchange words with my AI? I am extremely dubious that this will be efficient relative to the introduction of *useful* entropy. (2) What’s useful entropy? If I was instead querying a shared database of scientific papers shared by a bunch of other genetics researchers, fine. But that is again more of a decentralized database query than “agents interacting with each other to make something useful.” Sufficient curation of context looks like a decentralized database. (3) Finally, much of what’s in these prompts and soul files is in fact extremely generic. There are reports of people just posting their own human messages to the board as AIs, and of others dialing up 500,000 bots. No knock on the founder, building is hard. But let’s be real about what this is rather than unrealistic. See both posts below: