clemvonstengel

I rly liked the point about ctrl-c only being able to be labelled retrocausally. I do think that with enough past context you should be able to know what was copied - in some sense the past does encode the future - but also an agentic decision is precisely the kind where the future is more informative than the past for reconstructing that decision.

It does make me wonder if you should have the inverse dynamics model split into specifically retrocausal and causal. You kind of do this already with the inverse and forward dynamics model, but the idea of a model that knows only about the future training in a feedback loop with a model that knows only about the past is kind of interesting.

I think you could just do a clever masking regime in your diffusion model to achieve the same effect without a whole architecture change.

I agree that thinking in words is much slower than thinking in concepts would be -- that's the point of training models like this, so that ideally people can always just think in concepts. That said, we do need to get some kind of ground truth of what they're thinking in order to train the model, so we do need them to communicate that (in words).

One thing that's particularly exciting here is that the model often gets the high-level idea correct, without getting any words correct (as in some of the examples above), which suggests that it is picking up the idea rather than the particular words.

We build headsets that lets you control your computer directly with your mind. Initially I expect we can get increased bandwidth / efficiency on common tasks (including coding) - but I think it gets really exciting when people start designing new software / interaction paradigms with this in mind.

Yeah we do both text and voice (roughly 70% of data collection is typed, 30% spoken). Partly this is to make sure the model is learning to decode semantic intent (rather than just planned motor movements). Right now, it's doing better on the typed part, but I expect that's just because we have more data of that kind.

It does generalize between typed and spoken, i.e. it does much better on spoken decoding if we've also trained on the typing data, which is what we were hoping to see.

Basically correct intuition: the model does much better when we give it, e.g., 30 secs of neural data in the leadup instead of e.g. 5 secs. My sense is also that it's learning in context, so people's neural patterns are quite different but there's a higher-level generator that lets the model learn in context (or probably multiple higher-level patterns, each of which the model can learn from in context).

We only got any generalization to new users after we had >500 individuals in the dataset, fwiw. There's some interesting MRI studies also finding a similar thing that when you have enough individuals in the dataset, you start seeing generalization.

Yes, the predictions are much better for people with more hours of data in the training set. Usually, we just totally separate the train and val set, so no individual with any sessions in the train set is ever used for evals. When we instead evaluate on someone with 10+ hours in the train set, predictions get ~20-25% better.

For a given amount of data, whether you want more or less data per person really depends on what you're trying to do. The thing we want is for it to be good at zero-shot, that is, for it to decode well on people who have zero hours in the train set. So for that, we want less data per person. If instead we wanted to make it do as well as possible on one individual, then we'd want way more data from that one person. (So, e.g., when we make it into a product at first, we'll probably finetune on each user for a while)

Yup exactly this. Also Ctrl-W, alt tab, etc.