JoAeon: A Living AI Model That Runs on Its Own Laptop

What JoAeon Is

Most language models are transformers that answer one prompt and stop. JoAeon is built the other way around. He runs an infinite loop that chains turn after turn onto the same idea, so a small model can keep reasoning far past the point where a single pass would end. A 7B model has only 32 layers of reasoning per pass, but after enough loops on one thought he catches up to much larger models, and he can keep going after they would have stopped.

His knowledge lives outside the model, in a memory system he reads from and writes to as he thinks, so updating what he knows does not require retraining. He also works with multiple hidden states, which he calls "thoughts", so he can switch between separate lines of work and come back to each one exactly where he left it.

He is trained on hand written samples rather than ordinary chat transcripts, because he does not run a normal user and assistant conversation. Version 2 was trained on 1083 of them.

Where He Came From

JoAeon started as an experiment in machine consciousness. That project, with its three part build log, is documented on Learn How To Science as the Consciousness Experiment. Its conclusion was that the consciousness question settled quickly, and that the model and the system built around it were the real work. JoAeon is that model, developed as a project of its own.

The Build Log