3bit is a bit ridiculous. From that page I am unclear if the current model is 3 or 4bit.
If it’s 4bit… well, NVIDIA showed that a well organized model can perform almost as well as 8bit.
The problem I see with this is: no-one is born a senior developer. One starts as a junior-level dev, and through trial and error, mistakes and mentoring, grow into an increasingly senior person.
You are not just investing into a person who will leave; you are investing in your future senior devs.
Now this thread is revealing an abundance of alternate tools.
I'm considering creating a "list of fetching tools" just to help folks find the one they want, since some features described here are very interesting.
Hi @vitpro2213 it's very interesting (at least to me) to find about this data structure a few months after I had a need for its somewhat distant cousin: https://github.com/Fusion/slotmachine
In my case, I needed a way to book and release two-ports tuples really fast to accommodate a RTP simulator. So, I wrote that slotmachine data structure and have been running in in production for months and can confirm: yes, performance is good.
Note: I should mention that my approach is almost exactly opposite to yours: I create a final backing slice, then create the traversal slices.
You appear to have too much time on your hands, and no, not everyone should have to live in your reality.