462 karma · joined February 11, 2021
You have to add articles yourself and it does not have any special logic for HN threads yet.
If that’s not helpful, were you getting at having the model return some rich data about the attention weights that went into generating some token?
You can tell Andrew uses it himself from the practical docs and providing nice things like express middleware. I imagine it must be a bit mind bending to dogfood the service.
I think the main virtue of tagging systems is in the low friction to add info and multiple inclusion.
Tagging also has its downsides and I think we'll probably end up on some hybrid system in the long run.
These can be used to simulate physical systems like magnets and particle fields.
The Lagrange Constraints are used to constrain things like probability, energy, and particle numbers. What is constrained accounts for the difference between the ensembles listed above.
The derivation of the Support Vector Machine also involves Lagrange Constraints as well!
What does not go away is the need for a flexible and low friction way to collect and organize things like journal articles. Some sort of mental ontology about your field replaces the need for actually writing lots of stuff day to day.
[0] https://steamcdn-a.akamaihd.net/apps/valve/2007/NPAR07_Illus...
Our website got 1TB of traffic yesterday (which it turns out I have to pay for) and thats without hosting the binary on our site!