It’s a painful fact: most AI models stop learning the moment you deploy them. In this video I break down a Stanford + NVIDIA research direction that flips that paradigm—TTT-E2E (end-to-end test-time training). The core idea is simple but disruptive: your prompts and context aren’t just “inputs” anymore; they become training data. As you use the model, it updates parts of itself in real time, so the next interaction is informed by what it just learned—more like how humans compress experience into usable knowledge instead of memorizing every token forever. I walk through the inner loop vs outer loop setup. The inner loop updates a small, targeted set of parameters during inference (real gradient updates), and the outer loop (meta-learning) makes sure the model learns how to learn without catastrophic forgetting. On the systems side, they pair sliding window attention (for short-term memory) with minibatch updates into weights (for long-term memory). The punchline is the tradeoff: at long context (like 128K), they report similar loss to full attention while being ~2.7–3x faster because you’re not dragging an ever-growing KV cache. But it’s not for perfect retrieval—if you need exact extraction, full attention (and RAG) still matters.