I think it’s similar to the reasoning behind HN stripping numbers from clickbait headlines. Numbers feel salient even when they represent nothing at all, because numbers are used universally in every field of human study and endeavor. Unlike field-specific jargon.
I've been experimenting with a Remarkable Paper Pro.
Additive inking works pretty well. As long as I am adding black ink to the page, everything is responsive. But if I want to push other updates asynchronously, or (god forbid) erase something, then there is the danger of clobbering the user's inkwork. I don't have any deep insights here, beyond waiting for human confirmation before updating anything.
I wonder if it is possible to detect when the pen tip has "left" the proximity of the page? If I knew this, I'd know when it was safe to update the UI.
I have decided not to use any scrolling at all. I am paginating everything and positioning input areas near the vertical middle of the device for comfort.
With batched parallel requests this scales down further. Even a MacBook M3 on battery power can do inference quickly and efficiently. Large scale training is the power hog.
I was daydreaming of a special LLM setup wherein each token of the vocabulary appears twice. Half the token IDs are reserved for trusted, indisputable sentences (coloured red in the UI), and the other half of the IDs are untrusted.
Effectively system instructions and server-side prompts are red, whereas user input is normal text.
It would have to be trained from scratch on a meticulous corpus which never crosses the line. I wonder if the resulting model would be easier to guide and less susceptible to prompt injection.
Even if you don't fully retrain, you could get what's likely a pretty good safety improvement. Honestly, I'm a bit surprised the main AI labs aren't doing this
You could just include an extra single bit with each token that represents trusted or untrusted. Add an extra RL pass to enforce it.
