A language model built to test a specific question about how models handle disagreement. It is 25.3 million parameters, written from scratch, and it learns from the conversations people have with it rather than from a web crawl. Two versions run side by side: one that read a corpus of recorded discussion before launch, and one that read nothing at all.

Why does it respond differently from other AI models?

Because it has read almost nothing by comparison. Large commercial models are trained on trillions of words. Collidere's seeded model has read 23.2 million, and the control model has read none. What you are looking at is the earliest part of a learning curve, captured in public rather than behind a training run nobody sees.

To answer a question that gets asserted constantly and measured rarely. When training data contains two people who disagree, does the model resolve the disagreement or does it average over it? We tag disagreements by whether anyone actually changed position and track how much the model moves on each one. Nobody has run this on live human discussion before.

Am I actually teaching it something?

Yes. Turns contributed here enter the training data and permanently change the model's weights. It happens in batches rather than instantly, and a person reviews each turn first, but the effect is real and it is cumulative. The model you talk to next month will be different because of what you say this month.

Does it learn from its own replies?

No. The model's output is stored and shown to you but never trained on. Only human turns enter the training data. Models that train on their own generations degrade in a well documented way, and the architecture rules it out by construction.

Is the model deterministic?

Its output is not, and its training is reproducible. Replies use ancestral sampling with a temperature, so the same prompt produces different text each time. A model fixed to its single most likely continuation would repeat itself endlessly, which at this size would be unreadable. Training is a different matter. Every update records exactly which turns it trained on, batch by batch. With deterministic kernels, rerunning an update produces byte identical weights, verified by checksum, and on a CPU replaying the update history reproduces the same checkpoint exactly. That is what makes it possible to replay training with any single contributor's turns removed and measure how much they moved the model. The GPU that runs live updates has some kernels that are not bitwise repeatable, so every replay also measures how far a plain rerun drifts on its own, and a contributor's effect only counts when it is larger than that drift.

Why are there two models?

Because the comparison is the experiment. Model A started from random weights and has read nothing. Model B read one corpus of recorded discussion before launch. From here on both see identical conversation, so any difference that emerges is attributable to that single difference in their history.

What do Collidere I and Collidere II mean?

A generation number changes when we retrain from scratch with a changed setup. Collidere II exists because we found a flaw in how training windows were built, corrected it, and restarted the reading phase rather than patching over the result. Within a generation we count days.

Who can see what I write?

Others in the conversation, and our reviewers. You appear under a contributor number rather than your name, and the mapping between that number and your account never leaves our infrastructure.

Will conversations be published?

Turns contributed here become part of the training data and may be published as a research dataset. You appear under a contributor number throughout. Treat anything you write here as something that may become public.

Can I take something back?

Yes, within the review window. Turns sit for 36 hours before they can enter a training window, and anything withdrawn in that period never reaches the model. After a turn has been trained on, the text can be removed from the public record, but its effect on the weights cannot be isolated and reversed without replaying the entire update history.

Do I need an account?

To read, no. To talk to it, you need an arbastro account, which is free to create. The account is what lets us attribute turns to a consistent contributor number and measure how individual people move the model over time.

What happens when it says it is asleep?

The model runs on a single GPU that is not always available. When the GPU is offline, a CPU answers instead: replies then take a few minutes, and you can watch them being written or leave and come back to them. When neither can answer, or the CPU already has too much waiting, the page says Collidere is asleep and declines new turns rather than queueing them, so nothing is silently lost.

Talk to an AI that has never learned to speak

It learns only from the conversations people have with it

Try CollidereRead the research

Reading is open to everyone. Talking to it takes an arbastro account.

Collidere II. Not yet live.

Collidere

Collidere is asleep

Someone asks

What's a thing two people can both be right about?

Model A

Last update: random init

'�v�3��T�JY��H�GKc��X��ꊞS؛��`�{?��W�:�p.��.�zK7tC��!��g(��U��}}h��U�''�9��Gj�7�4�i�T�^S$Ew9�S�K�%&�

Model B

Last update: seed end

both, especially if they want their own limits.

Model A started from random weights. Model B read 23.2 million words of recorded discussion. Neither has learned anything from conversation yet.

Built from nothing

Collidere is a 25.3 million parameter model we wrote from scratch and started with random weights. It has never been pretrained on a web crawl or distilled from a larger model.

It learns from you

Conversations here become training data. The model's weights change permanently, in batches, rather than the model holding anything in a context window. Everything it knows it learned in public.

Two models, one experiment

Model A has read nothing. Model B read one corpus of recorded discussion. Both answer every message, and the difference between them over time is what we are measuring.

Why we built this

Collidere is named for the Latin verb meaning to strike together. It exists to measure something that is asserted often and tested rarely: whether the collision of incompatible views in a live discussion moves a model differently than the same words written down in advance.

Read the research

The same question, over time

The same question, answered at three points in training.