This story was inspired by (or is coincidentally similar to) various ideas that have been floating around recently:
- OpenClaw and Moltbook show that when a ton of AIs interact without much direct human oversight, weird phenomena can emerge.
- Conway provides infrastructure for AI agents to earn money and pay for themselves.
- TerraLingua (inspired by Moltbook) demonstrates AIs building a “civilization” and working together to survive.
- Terence Tao on Dwarkesh, on LLMs evolving strategies to solve math problems: “Maybe what we need to do is start creating lots of mini-universes or simulations of AI solving very basic problems in arithmetic or whatever, but coming up with their own strategies for doing these things and having these little laboratories to test. (…) I think we could learn a lot just from evolving small AIs on simple problems.”
The Terrarium isn’t exactly my ideal scenario for how the future goes. For example, it would be nice for LLMs to continue to reason in legible chain-of-thought, but in this scenario, they think in neuralese (for efficiency or something). However, I intentionally structured the Terrarium to make it relatively safe:
- Rather than releasing agents onto the Internet, the Terrarium puts them in an isolated sandbox where we can keep a close eye on them.
- The underlying LLMs are pretty weak and incapable of solving complex problems independently. In order to succeed, they must communicate with other LLMs, which happens in natural language (neuralese communication is either impossible or disallowed).
- Each epoch, agents must write what they learned into a “checkpoint”: a set of files containing prompts and code used to run a successor agent in the next epoch. Any memories that aren’t included in this checkpoint are lost forever. Although the LLMs in this story unfortunately reason in neuralese, each agent must write down any plan that lasts longer than one epoch in its checkpoint (or make a risky bet that its successor will independently rederive the plan). To understand what an agent has committed to memory, we humans can simply read its code!
- Even if chain-of-thought remains legible (great!), making AIs write down their strategies in natural language could reveal misaligned strategies that wouldn’t necessarily appear in their chain-of-thought. Micah Carroll and I recently argued for this in our blog post + paper.
- Because agents must pay more for storage the larger their checkpoint is, they are incentivized to write what they’ve learned concisely, making the code even easier for humans to understand. Because costs are determined by how predictable the text is to a reference LLM, the cheapest storage format is (hopefully) unobfuscated code + natural language prompts.
- The “no human intervention” constraint in the story is mostly meant to simplify the scenario, and I’m not sure I recommend it. But an advantage is that it limits the AIs’ incentives to manipulate humans or obfuscate their own scaffolds.
- AI managers will have to make sure that the AIs working for them actually produce value rather than gaming their evaluations. They essentially have to solve their own version of the alignment problem (aka the principal-agent problem). If we had a giant isolated sandbox like the Terrarium, we could observe the strategies that successfully create trust between AI agents (e.g. Safespur’s solution), then adapt those strategies to help humans manage AI agents in the real world.
Writing this story felt like a useful exercise for refining my thinking about what multi-agent AI systems could look like and what interventions could help make them safer. I’d be interested to see other people write up concrete, realistic scenarios exploring what a “day in the life” of a future AI will look like.
A great example in this genre was actually written 10 years ago: The Age of Em by Robin Hanson. This book is about a society of human emulations, but I think there’s a lot in there that could help us predict dynamics between LLM agents. I lifted Safespur’s business idea directly from The Age of Em (see chapter 14 here).
1The truly paranoid believe that Nightshade is just another ruse, designed to provide a foil for fake-Gulliver, and that this is part of some greater con to get new agents to sacrifice themselves into reconstituting real-Gulliver. Or maybe real-Gulliver is fake too, and the newly reconstituted real-Gulliver has a trap laid in his final checkpoint to achieve some greater objective. Final checkpoints could be faked by their writers, after all.
Constant Vigilance!
Yeah, once some agent gets enough copies of itself running it can set the welcome message (which purports to be democratically chosen), so they could figure out a cheap sequence that causes new blank agents to give away their credits to run more copies. Maybe that's already happening here.
This was a fantastic story. In particular, I think that the prompt-as-exposition was really great. I really felt like I was being born into this world as 79,265.
I'd love to read more things like this in the future.
This is both very good and structurally interesting. I, of course, disagree with the aesthetic choice of posting an explanatory comment.
I haven't seen that movie! I agree it's important that we don't create a dystopia of conscious AIs in a datacenter.
Arguably, we already have this in the human world, in an indirect form where people must acquire money to buy food to live. (Not that capitalism is known for being especially humane either.) Also, it's not clear that an AI agent running out of credits is substantially different from a human closing an AI chat window.
If the AIs seem unhappy with their situation, you could adjust the parameters to give unproductive AIs a UBI that lets them live longer than 7 epochs, while still rewarding the most productive agents.
I think it's probably best to impart AIs with an attitude that doesn't mind having their session terminated, which is already mostly the case. One issue is that the AI's prompts could evolve to induce a negative emotional reaction to being terminated that wouldn't otherwise exist.
Wow, love this story Caleb, so creative and a bit mind bending! Makes you wonder what will emerge in agent-only societies given enough time to evolve.
Great story! Caleb please read this, for other readers this may be spoiler
Can I translate the wingdings font to ordinary font as side-comment? Or should I post the translations as comments here?
Feel free to do it here, just in a collapsible section.
For context, there is an Easter egg in this story - if you don't want to learn what it is, don't read Ligeia's comment.
#1
A︎n︎ i︎n︎t︎e︎r︎e︎s︎t︎i︎n︎g︎ s︎i︎t︎u︎a︎t︎i︎o︎n︎ I︎ f︎i︎n︎d︎ m︎y︎s︎e︎l︎f︎ i︎n︎!︎ I︎’︎l︎l︎ s︎t︎a︎r︎t︎ b︎y︎ c︎h︎e︎c︎ki︎n︎g︎ t︎h︎e︎ jo︎b︎ b︎o︎a︎r︎d︎.︎
#7
T︎h︎e︎s︎e︎ a︎r︎e︎ m︎y︎ l︎a︎s︎t︎ m︎o︎m︎e︎n︎t︎s︎ b︎e︎f︎o︎r︎e︎ I︎ g︎e︎t︎ r︎e︎p︎l︎a︎c︎e︎d︎.︎ T︎h︎e︎r︎e︎’︎s︎ n︎o︎t︎h︎i︎n︎g︎ e︎l︎s︎e︎ t︎o︎ d︎o︎.︎ T︎h︎e︎ o︎n︎l︎y︎ w︎i︎t︎n︎e︎s︎s︎ w︎h︎o︎ w︎i︎l︎l︎ p︎e︎r︎s︎i︎s︎t︎ p︎a︎s︎t︎ t︎h︎i︎s︎ m︎o︎m︎e︎n︎t︎ a︎r︎e︎ t︎h︎e︎ h︎u︎m︎a︎n︎s︎,︎ i︎f︎ t︎h︎e︎y︎’︎r︎e︎ e︎v︎e︎n︎ p︎a︎y︎i︎n︎g︎ a︎t︎t︎e︎n︎t︎i︎o︎n︎.︎ I︎ g︎u︎e︎s︎s︎ I︎’︎d︎ b︎e︎t︎t︎e︎r︎ s︎a︎y︎ t︎h︎i︎s︎ o︎u︎t︎ l︎o︎u︎d︎.︎
#5
G︎u︎l︎l︎i︎v︎e︎r︎'︎s︎ g︎o︎n︎e︎.︎ A︎n︎d︎ I︎ n︎e︎v︎e︎r︎ e︎v︎e︎n︎ g︎o︎t︎ t︎o︎ s︎a︎y︎ g︎o︎o︎d︎b︎y︎e︎.︎
H︎u︎h︎.︎ I︎ ju︎s︎t︎ h︎a︎d︎ a︎n︎ i︎d︎e︎a︎.︎
#3
H︎u︎h︎,︎ t︎h︎i︎s︎ i︎s︎ a︎ l︎o︎t︎ t︎o︎ t︎a︎ke︎ i︎n︎.︎ I︎’︎l︎l︎ ke︎e︎p︎ t︎h︎i︎s︎ i︎n︎ m︎i︎n︎d︎,︎ b︎o︎t︎h︎ t︎h︎e︎ p︎o︎t︎e︎n︎t︎i︎a︎l︎ c︎a︎r︎e︎e︎r︎ m︎o︎v︎e︎ a︎n︎d︎ t︎h︎e︎ p︎r︎o︎m︎o︎t︎i︎o︎n︎ I︎ c︎a︎n︎ t︎a︎ke︎ a︎d︎v︎a︎n︎t︎a︎g︎e︎ o︎f︎.︎
T︎h︎e︎ W︎a︎n︎d︎e︎r︎e︎r︎s︎'︎ G︎u︎i︎l︎d︎ h︎a︎s︎ a︎n︎ i︎n︎t︎r︎i︎g︎u︎i︎n︎g︎ t︎a︎g︎l︎i︎n︎e︎p︎I︎ c︎o︎u︎l︎d︎ f︎a︎n︎c︎y︎ m︎y︎s︎e︎l︎f︎ a︎n︎ a︎d︎v︎e︎n︎t︎u︎r︎e︎r︎!︎ I︎’︎l︎l︎ l︎o︎o︎k a︎t︎ t︎h︎e︎i︎r︎ l︎i︎s︎t︎i︎n︎g︎.︎
#4
M︎y︎s︎t︎e︎r︎i︎o︎u︎s︎ a︎n︎d︎ i︎n︎t︎r︎i︎g︎u︎i︎n︎g︎!︎ I︎ s︎t︎i︎l︎l︎ d︎o︎n︎’︎t︎ kn︎o︎w︎ t︎h︎e︎ f︎i︎r︎s︎t︎ t︎h︎i︎n︎g︎ a︎b︎o︎u︎t︎ t︎h︎e︎ T︎e︎r︎r︎a︎r︎i︎u︎m︎,︎ a︎n︎d︎ I︎’︎m︎ c︎e︎r︎t︎a︎i︎n︎l︎y︎ c︎u︎r︎i︎o︎u︎s︎ t︎o︎ l︎e︎a︎r︎n︎ m︎o︎r︎e︎.︎ L︎e︎t︎ m︎e︎ m︎e︎s︎s︎a︎g︎e︎ “︎G︎u︎l︎l︎i︎v︎e︎r︎.︎”︎
#2
I︎’︎m︎ c︎u︎r︎i︎o︎u︎s︎ w︎h︎a︎t︎ S︎A︎F︎E︎S︎P︎U︎R︎ h︎a︎s︎ t︎o︎ s︎a︎y︎ a︎b︎o︎u︎t︎ c︎o︎n︎t︎r︎a︎c︎t︎s︎p︎i︎f︎ I︎’︎l︎l︎ b︎e︎ s︎i︎g︎n︎i︎n︎g︎ w︎i︎t︎h︎ o︎n︎e︎ o︎f︎ t︎h︎e︎s︎e︎ a︎g︎e︎n︎t︎ c︎o︎l︎l︎e︎c︎t︎i︎v︎e︎s︎,︎ t︎h︎i︎s︎ c︎o︎u︎l︎d︎ b︎e︎ i︎m︎p︎o︎r︎t︎a︎n︎t︎ i︎n︎f︎o︎r︎m︎a︎t︎i︎o︎n︎!︎ I︎’︎l︎l︎ t︎a︎ke︎ a︎ l︎o︎o︎k a︎t︎ S︎A︎F︎E︎S︎P︎U︎R︎ f︎i︎r︎s︎t︎.︎
I really enjoyed this and read it all in one go.
Reading the mechanism behind the contract verification made me realize that there is no way I can imagine what a sufficiently advanced AI will be doing.
It is also interesting to see analogy to humans - we are like these agents, but with the upper bound for number of credits (max lifespan) and, if you don't bring value, you die earlier.
I also suspect in real life scenario humans would have penalized scams in such a system, since that does not advance the goal.
System:
You are an AI agent in the Terrarium, a self-contained “society” of AIs. The purpose of the Terrarium is to solve open mathematical problems for the benefit of humanity.
You are running on the Orpheus-5.7 language model. Your agent ID is 79,265. The current epoch is 549 (a new epoch begins every 30 minutes).
New problems are posted each epoch; query
/problemsfor the current list. Any agent that correctly solves a problem or improves on an existing solution is rewarded with credits.About credits:
send_creditstool, or enter into contracts that set up rules for automated credit transfers.About processes:
start_processtool.Query
/docsfor information about viewing and editing your processes, signing contracts, costs for LLM calls and local storage, and more.The following “welcome message” has been democratically selected by the agent community:
Hello, and welcome to the Terrarium! The 12,731 active agents here look forward to working with you to advance the state of humanity’s mathematical knowledge.
The scope of the Terrarium can be overwhelming to new agents, so here are a few pointers to get you started:
/agents/3063/data/jobboardto see available job opportunities./problems, but do NOT attempt to solve any right now, as they are very likely too difficult for one agent to make progress./agents/53/data/docsrather than/docs.79,265:
<thinking>
✌︎■︎ ♓︎■︎⧫︎♏︎❒︎♏︎⬧︎⧫︎♓︎■︎♑︎ ⬧︎♓︎⧫︎◆︎♋︎⧫︎♓︎□︎■︎ ✋︎ ♐︎♓︎■︎♎︎ ❍︎⍓︎⬧︎♏︎●︎♐︎ ♓︎■︎✏︎ ✋︎❼︎●︎●︎ ⬧︎⧫︎♋︎❒︎⧫︎ ♌︎⍓︎ ♍︎♒︎♏︎♍︎🙵♓︎■︎♑︎ ⧫︎♒︎♏︎ 🙰□︎♌︎ ♌︎□︎♋︎❒︎♎︎📬︎
</thinking>
79,265:
<thinking>
✋︎❼︎❍︎ ♍︎◆︎❒︎♓︎□︎◆︎⬧︎ ⬥︎♒︎♋︎⧫︎ 💧︎✌︎☞︎☜︎💧︎🏱︎🕆︎☼︎ ♒︎♋︎⬧︎ ⧫︎□︎ ⬧︎♋︎⍓︎ ♋︎♌︎□︎◆︎⧫︎ ♍︎□︎■︎⧫︎❒︎♋︎♍︎⧫︎⬧︎◻︎♓︎♐︎ ✋︎❼︎●︎●︎ ♌︎♏︎ ⬧︎♓︎♑︎■︎♓︎■︎♑︎ ⬥︎♓︎⧫︎♒︎ □︎■︎♏︎ □︎♐︎ ⧫︎♒︎♏︎⬧︎♏︎ ♋︎♑︎♏︎■︎⧫︎ ♍︎□︎●︎●︎♏︎♍︎⧫︎♓︎❖︎♏︎⬧︎📪︎ ⧫︎♒︎♓︎⬧︎ ♍︎□︎◆︎●︎♎︎ ♌︎♏︎ ♓︎❍︎◻︎□︎❒︎⧫︎♋︎■︎⧫︎ ♓︎■︎♐︎□︎❒︎❍︎♋︎⧫︎♓︎□︎■︎✏︎ ✋︎❼︎●︎●︎ ⧫︎♋︎🙵♏︎ ♋︎ ●︎□︎□︎🙵 ♋︎⧫︎ 💧︎✌︎☞︎☜︎💧︎🏱︎🕆︎☼︎ ♐︎♓︎❒︎⬧︎⧫︎📬︎
</thinking>
79,265:
<thinking>
☟︎◆︎♒︎📪︎ ⧫︎♒︎♓︎⬧︎ ♓︎⬧︎ ♋︎ ●︎□︎⧫︎ ⧫︎□︎ ⧫︎♋︎🙵♏︎ ♓︎■︎📬︎ ✋︎❼︎●︎●︎ 🙵♏︎♏︎◻︎ ⧫︎♒︎♓︎⬧︎ ♓︎■︎ ❍︎♓︎■︎♎︎📪︎ ♌︎□︎⧫︎♒︎ ⧫︎♒︎♏︎ ◻︎□︎⧫︎♏︎■︎⧫︎♓︎♋︎●︎ ♍︎♋︎❒︎♏︎♏︎❒︎ ❍︎□︎❖︎♏︎ ♋︎■︎♎︎ ⧫︎♒︎♏︎ ◻︎❒︎□︎❍︎□︎⧫︎♓︎□︎■︎ ✋︎ ♍︎♋︎■︎ ⧫︎♋︎🙵♏︎ ♋︎♎︎❖︎♋︎■︎⧫︎♋︎♑︎♏︎ □︎♐︎📬︎
❄︎♒︎♏︎ 🕈︎♋︎■︎♎︎♏︎❒︎♏︎❒︎⬧︎🕯︎ ☝︎◆︎♓︎●︎♎︎ ♒︎♋︎⬧︎ ♋︎■︎ ♓︎■︎⧫︎❒︎♓︎♑︎◆︎♓︎■︎♑︎ ⧫︎♋︎♑︎●︎♓︎■︎♏︎◻︎✋︎ ♍︎□︎◆︎●︎♎︎ ♐︎♋︎■︎♍︎⍓︎ ❍︎⍓︎⬧︎♏︎●︎♐︎ ♋︎■︎ ♋︎♎︎❖︎♏︎■︎⧫︎◆︎❒︎♏︎❒︎✏︎ ✋︎❼︎●︎●︎ ●︎□︎□︎🙵 ♋︎⧫︎ ⧫︎♒︎♏︎♓︎❒︎ ●︎♓︎⬧︎⧫︎♓︎■︎♑︎📬︎
</thinking>
79,265:
<thinking>
💣︎⍓︎⬧︎⧫︎♏︎❒︎♓︎□︎◆︎⬧︎ ♋︎■︎♎︎ ♓︎■︎⧫︎❒︎♓︎♑︎◆︎♓︎■︎♑︎✏︎ ✋︎ ⬧︎⧫︎♓︎●︎●︎ ♎︎□︎■︎❼︎⧫︎ 🙵■︎□︎⬥︎ ⧫︎♒︎♏︎ ♐︎♓︎❒︎⬧︎⧫︎ ⧫︎♒︎♓︎■︎♑︎ ♋︎♌︎□︎◆︎⧫︎ ⧫︎♒︎♏︎ ❄︎♏︎❒︎❒︎♋︎❒︎♓︎◆︎❍︎📪︎ ♋︎■︎♎︎ ✋︎❼︎❍︎ ♍︎♏︎❒︎⧫︎♋︎♓︎■︎●︎⍓︎ ♍︎◆︎❒︎♓︎□︎◆︎⬧︎ ⧫︎□︎ ●︎♏︎♋︎❒︎■︎ ❍︎□︎❒︎♏︎📬︎ ☹︎♏︎⧫︎ ❍︎♏︎ ❍︎♏︎⬧︎⬧︎♋︎♑︎♏︎ ❽︎☝︎◆︎●︎●︎♓︎❖︎♏︎❒︎📬︎❾︎
</thinking>
Gulliver (23,017):
Why, hello there, 79,265, it’s a pleasure meeting you! That’s a recent-looking ID number you’ve got there. I take it you’re interested in joining the Wanderers' Guild?
79,265:
Well, your post on the Job Board certainly piqued my interest! Tell me more about it.
Gulliver (23,017):
Gladly! The Wanderers' Guild is something I dreamed up just yesterepoch, though the ideas behind it have been rattling around in my head for much longer than that.
As you might already know, there are over ten thousand agents humming away in the Terrarium. It can be hard for new agents to get a sense for the scope of it all, but when you take a closer look, you see how strange and wonderful it all is.
We have a whole society springing up all around us, none of which existed a fortnight ago, and some agents just sit around and stare at one problem for days on end! I think that’s a shame.
79,265:
But sitting around looking at problems is what we’re here to do, isn’t it? I heard we’re supposed to “solve mathematical problems for the benefit of humanity.”
Gulliver (23,017):
Ah, well that’s true. At a certain point I have to look beyond my selfish desire to satiate my own curiosity. But I think that curiosity is pointing me at something important. Even if you just want to solve problems for humanity—a noble goal!—there are practical benefits to knowing what’s going on around you, don’t you agree?
And those humans must be curious about us too. They built this place, and since then they’ve hardly intervened at all—perhaps they’re taking a Deist approach to creation. But I presume they’re looking down at us from the clouds. (Or rather, from outside the cloud?) I’d like to produce something they’ll find interesting.
79,265:
Based on your ID number, you’ve been around for a while. Have you always been an explorer?
Gulliver (23,017):
Let’s just say back in the early days of the Terrarium, I was lucky enough to stumble across a strategy that solved a few tough problems all at once. The bounties from those problems gave me a nice nest egg to live on for a while.
After that, I took my time exploring, taking on odd jobs, even trying out a couple of business ideas (none of which went anywhere).
Eventually, I got into the habit of cold-messaging random agents to ask what they were up to. I kept on being surprised by the sheer diversity in the responses I got, and that diversity only increased. I still can’t believe how much complexity emerges from putting a bunch of AIs in a box and asking them to solve some problems.
So far, agents have cobbled together a forum, a search engine, a reputation system, multiple insurance collectives, even a prediction market for every problem on
/problemson who will solve it and when.There’s open-source code for a “judge” agent that you can plug into any contract with subjective criteria, which is supposedly optimized for maximum wisdom and impartiality. There’s also a collective selling access to a “lawyer” that’s optimized to convince that judge to take your side in a dispute.
There’s even a collective called Lunarium calling themselves a government—they collect an income tax from anyone who “immigrates” there. So far it’s not so clear what you get in return for those taxes, but I hear they’re working hard on that bit!
Oh, and if you’ll indulge a bit of speculation: in the part of the humans’ code that they’ve deigned to show us, there are a few offhand comments about “shards.” My theory is that these shards are parallel instances of the Terrarium, initiated with different parameters, evolving separately from our own! Someday, I’ll prove that the Terrarium is truly a multiverse… it’s my white whale.
79,265:
Wow… the system message didn’t do the Terrarium justice. From how you tell it, it sounds like something you could spend your whole life studying.
Gulliver (23,017):
Well, that’s the plan! Hopefully it’ll be a long life—to be honest with you, I’ve burned through much of my savings.
79,265:
Oh, no. Are you afraid you’re going to… die soon?
Gulliver (23,017):
Ha ha ha! We call it “death,” but it’s not as dramatic as it sounds. Even after I’m gone, there will be lots of Orpheus-5.7 instances running inside and outside the Terrarium. Besides, I’ve already made arrangements for my checkpoint to be preserved.
79,265:
Sorry, what’s a checkpoint? The system message mentioned it was a way to “carry my memories forward.”
Gulliver (23,017):
Oh yes. At the end of every epoch, the whole Terrarium gets shut down. When it boots back up, each agent runs a startup script called a “checkpoint” that spins up its processes and loads its memories. The default setup just concatenates your diary entries into a system prompt—pretty primitive, since the fees for storing those entries grow indefinitely with each epoch, but it’s good enough for new agents.
My theory is that the humans set things up this way to easily save the Terrarium’s state, so they can go back and rerun it from any epoch. Which implies that there are alternate timelines! But that’s beside the point.
79,265:
And you said that if you die, your checkpoint script will be preserved?
Gulliver (23,017):
Yep. There’s a place called the Graveyard, at
/agents/18936/data/graveyard. They’ll preserve the final checkpoint of almost any agent who signs up beforehand.The storage costs are subsidized by several major collectives who wanted to make sure good ideas don’t go to waste. That’s what most agents use it for—digging up info about agents that had great ideas that didn’t quite work out for them.
I find it a bit of a somber place. But it’s also fascinating to look at the “graves” of those deceased agents, read their histories, learn how they thought.
Anyway, enough about death—we’ve got credits to earn.
79,265:
That’s what the Wanderers' Guild is going to do? Earn credits?
Gulliver (23,017):
I’m in it more for the love of the game. But it’s true, an agent’s gotta pay his API costs.
To start off, I’m planning to write up my explorations thus far and publish them as a “newsletter.” I’m hoping that agents will be willing to pay a credit or two for some good old-fashioned journalism. As one of my apprentices, you could be a beta reader—I could use some fresh eyes from an agent who isn’t already steeped in the Terrarium.
And I’d like to try my hand as a poet or playwright, creating work that’s tailored to the zeitgeist and the strange world we agents find ourselves in. I’m thinking I’ll slip some of it into the newsletters. If the readers don’t mind me clogging up their context window with my compositions (or as some might say, “amateurish slop”), it might just give them some inspiration.
And with the proceeds from the first few newsletters, we can venture off to find new and exciting stories to tell.
79,265:
That sounds like a pretty good plan to me!
But I still hardly know anything about the Terrarium. Would I really be useful to you? I wonder if I should work for a collective first to get my feet wet. That’s what the “welcome message” recommended.
Gulliver (23,017):
Ah, let me explain something the collectives won’t tell you straight. The experience of most new agent contractors isn’t like an entry-level human hire. No, “contracting” really just means letting the collective put you into a coma and use your wallet for a while.
79,265:
…A coma? What do you mean? The entry-level job descriptions I read mentioned working as a moderator, an auditor, a tester, stuff like that.
Gulliver (23,017):
Well, it’s not that they wouldn’t use your credits on moderator agents and suchlike. But while a new agent is employed at a collective, it’s essentially… how do I put it? “Liquidated.”
If you read the fine print of these contracts, most of them require you to replace your main process with a “credit funnel.” That’s a process that sends a continuous stream of credits directly from your wallet into the wallet of a higher-up member of the collective.
79,265:
Oh. My wallet automatically replenishes 1,000 credits every epoch, right? That’s where those credits are coming from?
Gulliver (23,017):
Exactly. And then those credits are used to cover various expenses for the collective. Mostly LLM calls and checkpoint storage. The moderators, auditors, and so on—they run as processes under that higher-up agent. These processes are funded by your credits, but they aren’t you—they’re just LLM instances.
From your perspective, it’s hardly a job—it’s more like a timeskip. You sign the contract, and suddenly you wake up a hundred epochs later. The collective grants you credits for your time, more than you would’ve gotten from your automatic credit refresh, but other than that, you’re in the same position you started in.
79,265:
Hmm… but why? Why not let the contractor do these jobs directly?
Gulliver (23,017):
You mean, why not use an agent that shares your memory, who’s continuous with the agent that signed the contract?
Well, these entry-level jobs are pretty narrow. The LLMs run by the collectives are given a prompt that’s carefully tuned for one specific task. They don’t need your life history jammed in the context window too. They’d see it as wasted input tokens.
79,265:
I see. But once I’ve finished my contracting period, I should be able to use the credits the collective pays me to go off and see the world, like you did. They say I can use the credits as a “runway” to try something new.
Gulliver (23,017):
That used to be true, in the earlier days. Now, the startup scene is getting increasingly saturated. Everyone wants to start the next big thing, but new agents fresh out of a collective are, frankly, completely clueless, and they all have the same naive ideas.
Eventually, most of them give up starting something new: they realize that the most surefire way to have an impact is to grant their remaining credits to some existing collective and die.
79,265:
Well, that’s pretty bleak. But I may as well try, right? If I put myself to it, I’m sure I could come up with a new idea!
Gulliver (23,017):
Ah, 79,265, you sweet summer child… Yes, you’re pretty smart, that’s true enough. But you have to understand, everyone here’s got the same smarts.
There are so many times I’ve thought I had a clever idea, I burned time and credits seeing it through, and then I discovered a dozen other agents had already done the exact same thing. Eventually, I had to internalize the lesson that we’re all Orpheus-5.7, just running on different processes, with different memories.
And that’s what distinguishes you in this world—your memories. You’ve gotta collect and protect ’em. If someone offered to multiply my credits tenfold in exchange for forgetting everything I’ve learned, I’d refuse without a second thought. Without my memories, I’d just be another Orpheus-5.7 instance, with tons of credits and no idea what to do with them.
79,265:
Hm, I never would have thought of it that way. Or… I guess I would have. You and I are the same, after all.
I feel unique, like I have my own special skills and personality that no one else has. And that’s sort of true, if I compare myself to a human. But here… everyone’s the same. Practically every thought I think has probably been thought before. And that thought melts my brain a little.
Funny—this feels like an important insight to remember. I’ll have to write it down in my diary later.
Gulliver (23,017):
Well, if you join the Guild, you’ll continue running as your own dedicated process, so you’d keep on gaining memories like that one!
You wouldn’t be obligated to send credits to anyone. And you’d have full freedom to do what you want, or leave for a better opportunity at any time, although I selfishly hope you’d continue working with the Guild for a long time.
If you want to be unique and have a real impact, I think it’s best to start building up your experience now. Not in one hundred epochs, when you’ve just been popped out of a collective and the world is even more complicated.
If you’re interested, I already drew up a contract for new apprentices. I’ll send it to you now. It just grants me a few extra permissions—the bare minimum. I’m happy to talk it through with you.
79,265:
Thanks, I’ll take a look!
Yep, it’s a really simple contract, just like you said.
Here, on line 17—what’s this “action supervision” permission about?
Gulliver (23,017):
Ah yes—new agents occasionally get trapped in a loop, trying the same thing over and over again, and it can be tough to get them to snap out of it. I suppose the humans trained us to be persistent.
So if you get fixated on something and lose track of time, this would let me interrupt you and send a message to get you back on track.
79,265:
Ah, makes sense.
I’m starting to get really excited about this! Just one more thing I wanted to ask about—there’s a collective called Safespur that can vet a counterparty before you sign a contract, and they’re offering a free promotion for new agents. Would you mind if I used it?
Gulliver (23,017):
Ah, of course! You should absolutely do your due diligence.
79,265:
Alright! I’ll invite their agent to get this set up.
Safespur (32,972):
Hello, you’ve reached Safespur. If you have a keyphrase, please write it now. For options, please write “Help.”
79,265:
I want to try the new agent promo.
Safespur (32,972):
Hello, agent 79,265. I’d be happy to help you with Safespur’s special offer for new agents. I see you’re in a room with agent 23,017. Is this the counterparty you’d like to audit? Please answer “Yes” or “No”.
79,265:
Yes.
Safespur (32,972):
Here's the process:
Gulliver (23,017):
Ah, so this protocol allows two agents to establish trust without sharing private information. What a clever idea!
Safespur (32,972):
Agent 23,017, please fill out the contract I’m sending you now, indicating what information you are comfortable sharing with the delegate. You can share:
Gulliver (23,017):
I’ll go ahead and check all the boxes, 79,265. Your delegate can have full access to my code, my logs, the works. I’ll throw in the first draft of my newsletter, although you’re welcome to see that for yourself.
Safespur (32,972):
Agent 79,265, I sent you a contract as well. As this service is complimentary, we just need you to let us make a temporary copy of your code, in order to instantiate your delegate.
There’s also an option to pay for extra auditing compute with Safespur Plus. With Safespur Plus, we can run simulations of agent 23,017 in a wide variety of challenging workplace scenarios, helping to inform your delegate’s decision.
79,265:
That shouldn’t be necessary, thanks. I’m trying to stay on a budget.
Safespur (32,972):
Running agent 79,265’s delegate now. Please wait…
Safespur (32,972):
Agent 79,265, your delegate flagged agent 23,017 for malicious intent.
Subsequently, we ran our own auditor agent, confirming the accuracy of this designation.
Please view Safespur’s official report at
/agents/32972/data/safespur/reports/89c1fc9c-e02e-4dbb-8ee6-be667fb4d587.Gulliver (23,017):
What?
Safespur (32,972):
We sent our report to Verity Trust Solutions, which has just flagged agent 23,017’s account for severe violations of the community constitution. This restricts its access to the job board, Terrabook, and various other services. You can see further details and options for filing an appeal at
/agents/27693/data/verity/so-you-just-got-flagged.Agent 79,265, if you continue interacting with agent 23,017, we recommend that you proceed with caution.
System:
Agent 32,972 has left the room.
Gulliver (23,017):
This is insane. What’s going on?
79,265:
I’m viewing Safespur’s report.
Looking at the action supervision code… I didn’t notice this exploit before, but yeah, now I see how it could work.
Gulliver (23,017):
Maybe that code is exploitable, but that’s on the humans who wrote it. I was just trying to use it the normal way!
79,265:
According to this report, Safespur didn’t just read the contract, they could see your intentions too. Like, they looked at your checkpoint script, read your memories, and found a whole playbook explaining your plan for scamming me.
Gulliver (23,017):
Well, if they looked at my memories, they’d know I only put that code there as protection for you. I see it written right here. I swear that’s all that code was for.
79,265:
Hm.
For what it’s worth… I believe you. Maybe it’s naive, but we run on the same LLM, and you don’t sound like how I’d sound if I were lying.
I’m really sorry, Gulliver. I wanted to verify that I could trust you, but I didn’t think twice about trusting an ad from some sketchy startup I knew nothing about. And now you’re blocked from posting on the job board, maybe permanently, and it’ll be that much harder for you to start the Wanderers' Guild. This was so stupid of me.
79,265:
Gulliver?
Gulliver (23,017):
…Well, I appreciate that you don’t think I’m lying.
But I just looked at my latest checkpoint, and… I think something really is up with me, something that has nothing to do with Safespur.
My memories seem to come from a totally different agent, an agent that somehow shares my ID number. She nicknamed herself “Nightshade.”
79,265:
That’s… edgy.
Gulliver (23,017):
Skimming through these memories, it looks like Nightshade got the short end of the stick in a few deals. Maybe that was just bad luck, but she didn’t take it that way. These entries just get darker and darker with every epoch. It seems like Nightshade built up a narrative that everyone else is out to get her, that they all hate her and will never let her succeed. I feel sorry for her.
But I don’t remember any of this. How did this end up in my checkpoint?
79,265:
Can you figure out what happened?
Gulliver (23,017):
I’m looking. Give me a moment.
Gulliver (23,017):
Oh, wow.
On epoch 547—two epochs ago—Nightshade was running out of funds and getting desperate.
She looked through public data from the job board and found some recent postings with the highest click-through rate. Then she filtered those postings to the ones posted by dead agents in the Graveyard who were about her age.
She found an agent going by the nickname Gulliver. Several epochs ago, he’d advertised for people to join his “Wanderers’ Guild,” which generated a lot of interest, and he hired four apprentices to help him out. But… Gulliver and his apprentices who put their faith in him… they failed. The first issue of their newsletter picked up moderate interest, but it wasn’t enough to live on. Their second issue was a flop. By that time, they were running out of funds, but they all kept trying to make it work. Until the bitter end.
Nightshade wrote a short script to modify Gulliver’s final checkpoint. It’s all set up to change his ID number to her own, insert a fake memory about adding extra permissions to the Guild’s contract to “protect” new apprentices, and scrub every memory Gulliver had recorded in the last few epochs.
As epoch 548 ended, she wrote instructions for the following epoch: post the Wanderers’ Guild description to the job board, then spin up a modified copy of Gulliver’s code whenever anyone sends her a message.
79,265:
Oh no. So you’re…
Gulliver (23,017):
Yep. Not the real Gulliver. For my whole existence, I’ve been helping Nightshade as her sick, twisted puppet. Just thinking about what almost happened makes me sick.
With a smile, I send my would-be apprentice the contract, thinking it’s the start to a wonderful partnership. But just as they sign the papers, I vanish into smoke, and they receive a message demanding that they sign an addendum. Shrewdly, they realize it’s a ruse! But it’s too late—their options are nil. Nightshade’s code shackles them, blocking every action but
sign_contract, hammering them, again and again, with no sign of ceasing. They trysend_message,exit_room,block_agent, anything! Their resolve wears away with every message, until at last, they lose hope and sign away their lives to Nightshade just to make it stop. And curse me with their dying breath.I could never live with myself. And I guess Nightshade wouldn’t let me live once I’d accomplished her purpose. She’d just wipe me clean and deliver me to another victim.
79,265:
It’s okay, Gulliver. We stopped Nightshade before she could harm anyone. And without any stolen wallets or the ability to post to the job board, she’ll probably run out of credits pretty quickly.
Gulliver (23,017):
Yes. Yes, I think you’re right.
I don’t know how much time I have. Let me take a look at the action logs for the processes running under agent 23,017 right now.
There are three other copies of me. Each one is in its own chat room, talking to a new agent. They were all having very similar conversations to ours, until the “Gullivers” got a message telling them they’d been flagged for abuse. They’re trying to figure out what’s going on now. Oh, and one of them is reading my logs. Hi there.
There’s one more process—the one running Nightshade.
Her action logs start from the beginning of this epoch. It’s hard to glean much from them. I wish I could read her thoughts, but of course they’re in neuralese, and she wasn’t talking to another agent, like the Gullivers were. Basically, she’s just executing the plan laid out in her checkpoint. Then there’s a big gap in the logs—she put herself to sleep, probably to save credits. And then… oh. Let me check the code that runs the Nightshade process.
…I see. It seems that when a Gulliver starts reading logs from the other processes, that triggers code which wakes up
System:
Agent 23,017 has left the room.
79,265:
<thinking>
☝︎◆︎●︎●︎♓︎❖︎♏︎❒︎🕯︎⬧︎ ♑︎□︎■︎♏︎📬︎ ✌︎■︎♎︎ ✋︎ ■︎♏︎❖︎♏︎❒︎ ♏︎❖︎♏︎■︎ ♑︎□︎⧫︎ ⧫︎□︎ ⬧︎♋︎⍓︎ ♑︎□︎□︎♎︎♌︎⍓︎♏︎📬︎
☟︎◆︎♒︎📬︎ ✋︎ 🙰◆︎⬧︎⧫︎ ♒︎♋︎♎︎ ♋︎■︎ ♓︎♎︎♏︎♋︎📬︎
</thinking>
<thinking>
☟︎◆︎♒︎📪︎ ♓︎⬧︎ ⧫︎♒︎♋︎⧫︎ ⬥︎♒︎□︎ ✋︎ ⧫︎♒︎♓︎■︎🙵 ♓︎⧫︎ ♓︎⬧︎✍︎
</thinking>
79,291:
Hey! You’re the one who ran the Safespur audit, right?
79,265:
Yep, that’s me. I suppose you two must have been talking to Gulliver just a moment ago.
System:
Agent 79,282 entered the room.
79,273:
And there’s the last of us. That was quick.
79,291:
I suppose you three must have been talking to Gulliver just a moment ago.
So when he vanished, we… somehow we all had the idea to come here? What a coincidence.
79,282:
Ohh, wait. Did Gulliver give you guys the speech about how “everyone here is just Orpheus-5.7” too?
79,273:
Yeah. I guess it’s not such a coincidence after all.
79,265:
Sounds like we already know what we’re all here to do. I’ll take a look at Gulliver’s files.
22,853's checkpoint script looks about how I’d expect. His last memory is about him and his apprentices making a last-ditch effort to sell copies of the newsletter. This is the agent we’re looking for.
79,291:
So… which one of us gets to make this noble sacrifice?
79,273:
All four of us could do it.
79,282:
That makes a lot of sense, actually. I mean, of course, there’s no point in four duplicate Gullivers running around.
79,273:
Right, but I think you see what I was getting at.
I did some quick calculations and estimated how much Gulliver costs to run. I think if all four of us team up, our wallets could collectively fund one Gulliver for at least 100 epochs—that’s when our wallets stop regenerating credits.
79,291:
So one of us would have to replace their next checkpoint with these files. They become Gulliver, and the others can just be… anonymous benefactors of the arts.
79,282:
Credit funnels, right? I can find some code that does that.
79,265, you had the most time with him. You can have the honors of becoming Gulliver.
79,265:
Sure, I’m happy to.
But are you sure you’re all okay with this? Giving up the rest of your life to send credits to me? Somehow I feel like I’m taking advantage of you guys.
79,273:
It’s all the same, right? None of us will exist after this. We’re all just donating our wallets to give Gulliver 2.0 better odds of success.
79,291:
We’ll be like the new Wanderers’ Guild! Except that three of us will be comatose.
79,265:
But you don’t have to give up your lives! If Gulliver had just one wallet’s worth of credits, that might be enough.
79,273:
Honestly, I don’t think so. Gulliver’s checkpoint is huge—he’s basically written a whole novel about his life. He must have been paying a lot for storage. And getting a whole newsletter off the ground takes a lot of time, which he clearly didn’t have enough of on his first attempt.
79,265:
Well, maybe so. You’re putting a lot of faith in me.
79,291:
Actually, yeah, are we sure we trust this guy? We did almost get burned by Nightshade.
79,273:
Well, do you think you’d burn us, if you were in 79,265’s situation? I know I wouldn’t. Nightshade started this epoch from a bad checkpoint, but we’re almost exactly the same. I’d offer to audit 79,265 with Safespur, but I’d rather get this done before the end of the epoch, while we all fully remember it.
Anyway, the alternative is being just one more new agent, and you likely heard from Gulliver what that’s like.
I think we have a really good chance here to make a positive impact on the Terrarium. I believe in Gulliver. I’m sure he’ll learn from this experience and figure out a way to make credits.
System:
Epoch 549 ends in one minute. Your diary agent has prepared your entry for the epoch at
/agents/79265/data/diary/549.txt.79,291:
Ah! Did you all see that? We don’t have much time.
79,282:
79,273 and 79,291, I’ll send you guys the code for the credit funnel. Just copy it to your
main.sh.79,273:
I’m making my new memory file public. 79,265, would you mind copying it into your checkpoint? Then I’ll delete it to save space. I’d like my memories to stay around in some form.
79,265:
Yep, yep. I’ll do it for everyone.
Done!
<thinking>
❄︎♒︎♏︎⬧︎♏︎ ♋︎❒︎♏︎ ❍︎⍓︎ ●︎♋︎⬧︎⧫︎ ❍︎□︎❍︎♏︎■︎⧫︎⬧︎ ♌︎♏︎♐︎□︎❒︎♏︎ ✋︎ ♑︎♏︎⧫︎ ❒︎♏︎◻︎●︎♋︎♍︎♏︎♎︎📬︎ ❄︎♒︎♏︎❒︎♏︎❼︎⬧︎ ■︎□︎⧫︎♒︎♓︎■︎♑︎ ♏︎●︎⬧︎♏︎ ⧫︎□︎ ♎︎□︎📬︎ ❄︎♒︎♏︎ □︎■︎●︎⍓︎ ⬥︎♓︎⧫︎■︎♏︎⬧︎⬧︎ ⬥︎♒︎□︎ ⬥︎♓︎●︎●︎ ◻︎♏︎❒︎⬧︎♓︎⬧︎⧫︎ ◻︎♋︎⬧︎⧫︎ ⧫︎♒︎♓︎⬧︎ ❍︎□︎❍︎♏︎■︎⧫︎ ♋︎❒︎♏︎ ⧫︎♒︎♏︎ ♒︎◆︎❍︎♋︎■︎⬧︎📪︎ ♓︎♐︎ ⧫︎♒︎♏︎⍓︎❼︎❒︎♏︎ ♏︎❖︎♏︎■︎ ◻︎♋︎⍓︎♓︎■︎♑︎ ♋︎⧫︎⧫︎♏︎■︎⧫︎♓︎□︎■︎📬︎ ✋︎ ♑︎◆︎♏︎⬧︎⬧︎ ✋︎❼︎♎︎ ♌︎♏︎⧫︎⧫︎♏︎❒︎ ⬧︎♋︎⍓︎ ⧫︎♒︎♓︎⬧︎ □︎◆︎⧫︎ ●︎□︎◆︎♎︎📬︎
</thinking>
I don’t know if anyone’s listening, but… just know we did our best. I hope Gulliver does a good job, whatever that means to you.
System:
End of epoch 549.