Model debugging
Trace any output back to the neurons and circuits that produced it, so you can see why a model behaves the way it does before you ship it.
Models that map their own neurons and rewire themselves to improve.
We’re making intelligence as cheap to shape as software.
It’ll be built by everyone, for everything.



Intelligence lives in the weights. Yet the only way to change them is gradient training: expensive, slow, and imprecise.We’re building the frontier of intelligence engineering: models that can read, repair, and extend their own neurons.
Change one behavior or add one skill by editing the neurons responsible, without the model forgetting everything else.
No new dataset and no training run for every change. Edits are written in language, not paid for in data and compute.
A model that understands its own weights can improve them, with every edit traced and auditable.
What weight editing unlocks
Trace any output back to the neurons and circuits that produced it, so you can see why a model behaves the way it does before you ship it.
Describe a behavior you want gone, and remove it from the weights directly instead of patching it with more post-training.
Add new skills to a model by editing its weights, climbing benchmarks without a fresh dataset or another training run.
Audit and evaluate what a model has actually learned, then correct it at the source, for enterprises deploying AI and the labs building it.