Viewing profile — Ambix

There were two popular models: Agat 7 and Agat 9. The second one was great machine, one of the best personal computers produced in USSR in large quantitites.

This [1] link is absolutely golden, thanks!

I've used this meta framework for LLM tuning, it really one of the best out there.

Still my default tool to install on any remote server. My secret weapon :)

Hehe, great post. Just use Hono or Hono X. It's so easier to reason about what's going on for me, mostly backend guy.

Wow, so cool retro designed blog there! For those who kinda liked window system of first Mac generations :)

This so cool! Would love to have some time to train myself with those exercises more

On acoustic with light or medium strings - it's OK. I used to flatten them with nail file from time to time. But it might be much easier on your finger tips just to start with elec…

I did my own experiments and it looks like (surprisingly) Q4KM models often outperforms Q6 and Q8 quantised models. For bigger models (in range of 8B - 70B) the Q4KM is very good, …

> LLMs are very bad at interpreting Markdown tables Which table format is better for LLMs? Do you have some insights there?

I did started learning 6502 assembler back in those days :) Recently I have been busy writing the emulator in Golang: https://github.com/gotzmann/6502

Yes, that's what I've seen from a lot of my experiments with fine-tuning. One should be really careful to not "lobotomize" already capable model and achieve better results at the e…

ChatGPT turns one today :) What a day to launch the project I'm tinkering with for more than half a year. Welcome new LLM platform suited both for individual research and scaling A…

Try to use less cores. RAM bandwidth is real limiting factor there, so there always some sweet spot between CPU cores and RAM bandwidth for individual system. For example, I use on…

Both 1 and 2 versions are good enough in Russian even for some real use in production environments. But you should tolerate some crazy / weird typos and mis-wording they'll produce…

Is it possible to write llama.cpp in Futhark? Like do effective math manipulations on 4-bit vectors within GPU?

OMG you do not need this bloatload for just PG hosting. Just harden SSH, harden PG configs and voila :)

Yep, I'd always prefer freedom and power of hosting my own PG instance upon a some robust VM offering to guiggling with clumsy AWS menus.

Token is basically a number. Embedding is a vector consisting of those numbers (tokens).

Tokens are just integer numbers, showing their position in the big vocabulary - it's that simple :) And vocabulary is just an array / vector / list - it depends which programming l…

You might reuse simple LLaMA tokenizer right in your Go code, look there: https://github.com/gotzmann/llama.go/blob/8cc54ca81e6bfbce25...

For those who interested, there some new researches in the field [0]. It usually possible to create more compact token representation from given text, but my guess the greedy "opti…

I think the same. LLMs are actually sort of "multi-linguas", able to transform source of any language to internal representation and then do output in some other language, thanks t…

Intel and AMD might be fast on the paper... but the RAM bandwidth matter much more now (thing of gen AI applications) than raw CPU power. So Apple Silicon is really the only CPU do…

From what I know, modern GPUs are still programmed with C++ exclusively. See CUDA [0] for Nvidia and ROCm [1] for AMD. Why is this? Why Rust is not loved there? [0] https://docs.nv…