To produce this implementation from the LLaMA paper we started from github.com/karpathy/nanoGPT, the LLaMA architecture is really similar to GPT. For instance we added rotary positional encoding starting from the original RoPE repo published with the paper.
We finally ran the original model to make sure the two models were numerically.
Shen is pretty amazing, it's a full fledged functional language with lots of features and lots of compilation backends. I don't think there's a lot of ground for a comparison with ki, which is much more limited in scope.
I have ideas around providing CSP in ki, but so far I haven't implemented them. All there is in ki right now is the (letc ...) form (see http://ki-lang.org/api.html#misc), which provides very similar functionality to LiveScript's backcalls. It is no CSP but it allows to straighten up nested async calls.
The main difference is structural sharing: I can update a Clojure map with a new value, which will create another map that shares the vast majority of the data with the former map. With plain JavaScript, you'd have to clone your object, update it and freeze it again. In addition Clojure/ClojureScript/mori offer a whole API designed to leverage immutable data structures.
I agree, LispyScript is very neat, closer to a full-fledged language than ki. On the other hand, ki is focused on ClojureScript's immutable data structures and related idioms - in fact it's just a thin layer around those.
[Hi, author here] Yes, that is basically the idea. ki functions are js functions, same goes for vars, which makes it easy to jump from js to ki and back, without giving up on immutability, etc. Note that ki is not really intended as a ClojureScript alternative at all. It's more a way to take advantage of its data structures and some of its expressiveness from a pure JS ecosystem.
To produce this implementation from the LLaMA paper we started from github.com/karpathy/nanoGPT, the LLaMA architecture is really similar to GPT. For instance we added rotary positional encoding starting from the original RoPE repo published with the paper.
We finally ran the original model to make sure the two models were numerically.