Untrainable: The AI Model That Refuses to Forget

Most AI models degrade when you update them. Train them on new data — a new language, a new task — and they start forgetting what they already know. That's called "catastrophic forgetting," and it's been a headache for developers since the early days of neural networks. There's a whole field devoted to coaxing models back to their old selves.

Enter Untrainable, a model from 2024 that refuses to forget. Instead of fighting the forgetting, it uses a mathematical trick to keep old knowledge intact even as new data flows in. The result is a model that can learn new things without losing the things it already knows. No complex retraining pipelines, no careful data scheduling — just a model that stays itself.

This matters because the whole AI industry has been treating models like whiteboards: erase, rewrite, repeat. Untrainable flips that assumption. Models that hold onto knowledge are cheaper to maintain, faster to update, and less prone to the weird regressions that happen when you retrain and the model forgets something important. For teams building products on top of AI, that's a real practical win.

Why this matters for us: As more of our daily tools — from translation apps to customer service bots to the AI features woven into our bank and healthcare apps — lean on AI models, the ones that actually hold onto what they know will be the ones we can trust when it counts.