AT0M

What it is

Quick answers

How does AT0M differ from generative AI models?

AT0M is non-generative and scores supplied options rather than generating text. It returns one probability per option and cannot answer outside the candidate set, making malformed or off-list answers structurally impossible rather than just discouraged, unlike generative models.

Can AT0M be trained on your own data?

Yes, AT0M is trainable on your own decisions. Ten thousand of your own records with five questions each requires about two hours to train on a laptop, without needing datacentre GPU or cluster time.

What is the deployment model for v1?

v1 will be a self-contained Rust binary that runs on your own machine. You can either score a file of decisions or serve the same API yourself on your local network, with no external network dependency, provider, or per-decision fee.

What types of decisions can AT0M handle?

AT0M supports three decision types: choice, score, and yes-no (noul). Multiple questions can be answered together in a single forward pass, with the output being one probability per option.

What are AT0M's inference requirements and performance?

AT0M achieves 78.9% accuracy on the typed-decisions benchmark with a 16 millisecond response time on a T4. It runs as a deterministic Rust binary on CPU, Apple Metal, or CUDA with 60M trainable parameters and requires no external service.