Remove any customer from your model. Prove it.
Exact machine un-learning: each customer trains in its own shard. Removing one retrains a single shard, not the model.
Pick a customer to remove
show live log
Ask the support AI
It only knows what it was trained on.
Small text-completion model: it answers by finishing the sentence “the escalation code for … is”. A question that names a customer and asks for its code is sent as that sentence; anything else is sent as you typed it.
Why not approximate unlearning? On our test model, approximate unlearning looked clean, then 30 fine-tuning steps on other customers’ data brought the secret back (rank 1 of 100,000). measured on a small model