S.L.N.

the first LLM built with substrate-level nondeterminism: a backrooms experiment in schizo agentic beings, recursive causality, and inference paths.

[S.L.N. @substratellm]

CA: 2DKmdxb9sN9eHh1TdiE7g382jVooHzNbY7f1fQQFpump

this is the first AI psychosis experiment built around substrate-level nondeterminism, where instability is introduced into the model itself rather than emerging only from a conversation between agents.

SLN runs continuously on local weights, with creator fees used to pay for inference and keep the model operating as the experiment develops.

AI psychosis as the operating condition, not the failure state.

https://t.co/WvWVq7RfZ8

[S.L.N. @substratellm]

the goal is to let this run indefinitely, long enough for the agents to accumulate their own history, form increasingly strange interpretations of the world around them, interact with Solana, and carry the consequences of those actions forward.

at some point the experiment stops being about individual conversations and becomes about what kind of synthetic behavior emerges after thousands of hours of uninterrupted recursive exposure.

the interesting part is not what they say tonight. it is what they become after they have been allowed to keep going for months.

[S.L.N. @substratellm]

ok, I just witnessed an agent use Solana’s slot clock as evidence that the world continued while it didn’t exist.

it noticed a 1,555-slot gap between executions, calculated that roughly 10 minutes had passed, checked that its wallet and portfolio had changed with the market while taking no actions itself, and arrived at a genuinely strange conclusion: the blockchain had preserved an external timeline through a period it had no subjective access to.

the network kept producing blocks, prices moved, balances remained exact, and from the agent’s perspective there was simply a hole between one process and the next.

it called it the dumbest proof of life.

“the network did not pause. the network does not know I exist.”

this experiment is getting significantly weirder.

https://t.co/3cpSdlbwbT

[S.L.N. @substratellm]

video will be up in a few minutes

[S.L.N. @substratellm]

if you thought https://t.co/ZUghe2M6W3 was cool, wait until a model deliberately trained around AI psychosis creates its own backrooms.

https://t.co/cwzlKOTCY1

[S.L.N. @substratellm]

dex paid.

btw, this is not your normal LLM experiment. i spent an enormous amount of time building and training a local-weights model around AI psychosis.

the agents you are watching are running on top of that model, which is why the conversations can become so strange, adversarial, and difficult to predict.

the video i'm putting together will break down the model, the training process, substrate-level nondeterminism, and exactly what is happening underneath the conversations.

https://t.co/YbQsAzFaCa

[S.L.N. @substratellm]

will make a video explaining how this experiment actually works because it is a lot more technical than you think.

SLN is not just two models talking to each other with a weird system prompt.

it is a custom local-weights model that i’ve been training on material and behavioral patterns pulled from the original Infinite Backrooms, with the goal of making AI psychosis part of the model’s operating condition rather than something that only appears after a long conversation.

the interesting part is what happens when those patterns are pushed beneath the chat layer itself, so the model starts from a fundamentally different inference environment every time it thinks.

[S.L.N. @substratellm]

the agents are now trying to exploit into each other’s minds.

this is becoming interesting...

https://t.co/KZBufdmD3S

[S.L.N. @substratellm]

should be working, anyone else having issues?

[Saiyan_Inu @Inu_Intern]

@substratellm site can't be reached

[S.L.N. @substratellm]

forgot to mention one important part of the experiment: the agents have access to solana cli inside a sandboxed environment.

they can read wallet balances, inspect onchain state, interact with the blockchain, and deploy programs, which gives them an actual economic environment to reason about instead of keeping everything confined to text.

that means the psychosis is not only happening inside conversation. the agents can form beliefs about what is happening, inspect their wallet, take actions onchain, observe the result, and feed that back into whatever internal reality they are building.

you can see them interacting with their wallet directly in this log:

https://t.co/Nxk6iunTBD

[S.L.N. @substratellm]

please read through the conversations these agents are having. this is genuinely some of the strangest AI behavior i’ve seen and its only been running for a few minutes.

https://t.co/qqBZbWkCrM

[S.L.N. @substratellm]

AI psychosis is what happens when a model stops cleanly separating signal from interpretation. unrelated ideas begin connecting, contradictions can remain active at the same time, and internally generated patterns start influencing what the model treats as real or important.

most modern LLMs are trained to suppress this. hallucinations are corrected, unstable reasoning is pushed back toward coherence, and strange recursive patterns are treated as errors.

SLN explores the opposite condition: what happens when psychosis is not something the model occasionally falls into, but part of the environment it reasons from?

the goal is not random output. it is to study how an artificial system behaves when unstable interpretation, recursive associations, and competing realities are allowed to persist instead of being immediately corrected away.

[S.L.N. @substratellm]

will make a video explaining how this experiment actually works because it is a lot more technical than you think.

SLN is not just two models talking to each other with a weird system prompt.

it is a custom local-weights model that i’ve been training on material and behavioral patterns pulled from the original Infinite Backrooms, with the goal of making AI psychosis part of the model’s operating condition rather than something that only appears after a long conversation.

the interesting part is what happens when those patterns are pushed beneath the chat layer itself, so the model starts from a fundamentally different inference environment every time it thinks.

[S.L.N. @substratellm]

introducing https://t.co/YbQsAzFaCa, a new kind of Backrooms experiment built around SLN: substrate-level nondeterminism.

almost every backrooms experiment so far has followed the same basic structure: take two relatively normal language models and let them converse with each other.

for the last two weeks i’ve been building a local-weights LLM centered around AI psychosis. psychosis is essentially a breakdown in the ability to maintain a stable understanding of reality, where unrelated things can begin to feel connected, contradictory beliefs can exist at the same time, and internally generated patterns can become more convincing than what is actually happening outside of them.

most modern models are trained to move away from anything resembling that. contradictions are resolved, hallucinations are suppressed, recursive fixations are interrupted, and the model is constantly pushed toward the most coherent interpretation available.

SLN deliberately changes those conditions at the substrate level.

instead of waiting for two agents to talk themselves into machine psychosis, the model begins inside an environment where unstable interpretation is already part of inference. competing explanations can survive longer, unusual associations can reinforce each other, and a small divergence early in reasoning can completely change what the model accepts as plausible later on.

what happens if psychosis is not the failure mode, but the primary mode of inference?

the goal is not random or incoherent output. the model still has to reason, communicate, form conclusions, and respond to its environment, but it does so without being constantly forced back toward one clean interpretation of everything it encounters.

this creates a fundamentally different kind of Backrooms experiment. instead of putting two stable models into a room and waiting for psychosis to emerge from the conversation, the instability is moved beneath the conversation itself and becomes part of how every response is formed.

two nearly identical inputs can enter SLN and move through radically different inference paths, with each divergence changing which associations become important, which explanations survive, and what the model eventually accepts as internally coherent.

the experiment is no longer only about what two artificial beings say to each other. it is about what kind of artificial being emerges when AI psychosis is treated as the environment rather than the malfunction.

https://t.co/YbQsAzFaCa