Google's swipe typing is pretty damn good, but it isn't magic. It gets a lot better after using it for a while so I'm sure they use reinforcement learning in addition to the dictionary.

They've had it for quite a while and it's good, so I doubt they are putting much work into improving it. And since we have phones with TPUs and multimodal LLMs now, I think it's possible to beat google.

The LLM would only need 3-10 words of context and the swipe data as input to generate a single word, so it can be very small. I don't know much about the power of cellphone TPUs, but I think training an LLM with about 10M parameters on the fly should be possible. If that's the case, we could beat google while doing everything locally on the phone, so no privacy compromises.

Now that I think about it, it sounds doable. But then again I never did anything like this so I'm probably underestimating it by an order of magnitude or two.