Ambix

7 ms·

1.
by Ambix 5mo ago
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.
2.
by Ambix 10mo ago
This [1] link is absolutely golden, thanks!
3.
by Ambix 1y ago
I've used this meta framework for LLM tuning, it really one of the best out there.
4.
by Ambix 1y ago
Still my default tool to install on any remote server. My secret weapon :)
5.
by Ambix 1y ago
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.
6.
by Ambix 1y ago
Wow, so cool retro designed blog there! For those who kinda liked window system of first Mac generations :)
7.
by Ambix 1y ago
This so cool! Would love to have some time to train myself with those exercises more
8.
by Ambix 1y ago
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 electric and then progress towards acoustic.
9.
by Ambix 2y ago
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, there are no any degradation compared to full FP16 models.
10.
by Ambix 2y ago
> LLMs are very bad at interpreting Markdown tables Which table format is better for LLMs? Do you have some insights there?
11.
by Ambix 2y ago
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
12.
by Ambix 2y ago
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 end. It's trickier than seems from m
13.
Show HN: Collider – the platform for local LLM debug and inference at warp speed(github.com)
3 points by Ambix 3y ago|1 comments
14.
by Ambix 3y ago
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 only 6 cores from 10 on my M1 Pro laptop.
15.
by Ambix 3y ago
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. Russian is complicated.
16.
by Ambix 3y ago
Is it possible to write llama.cpp in Futhark? Like do effective math manipulations on 4-bit vectors within GPU?
17.
by Ambix 3y ago
OMG you do not need this bloatload for just PG hosting. Just harden SSH, harden PG configs and voila :)
18.
by Ambix 3y ago
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.
19.
by Ambix 3y ago
Token is basically a number. Embedding is a vector consisting of those numbers (tokens).
20.
by Ambix 3y ago
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 language you use, each has each own terminolog
21.
by Ambix 3y ago
You might reuse simple LLaMA tokenizer right in your Go code, look there: https://github.com/gotzmann/llama.go/blob/8cc54ca81e6bfbce25...
22.
by Ambix 3y ago
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 "optimal" tokenizer might harm the performance of the m
23.
by Ambix 3y ago
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 to so many layers of neurons inside it.
24.
by Ambix 3y ago
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 doing crazy 200 / 400 / 800 GB/s on the consume
25.
Ask HN: C++ still dominates on GPUs, why not Rust?
11 points by Ambix 3y ago|2 comments
26.
by Ambix 3y ago
Wow, looks like fresh idea and will be useful for me! Will wait for new releases
27.
by Ambix 3y ago
From what I know nobody in China cares about English and that's a nation like 2-3x bigger than all population of English-speaking countries :) Can't see a demise of Chinese even after 1,000 years. As well other languages with big
28.
by Ambix 3y ago
Go see yourself :) https://github.com/ggerganov/llama.cpp/blob/master/examples/... There's too many schemes right now with 4_0 and 5_1 really popular between LLM geeks.
29.
by Ambix 3y ago
Not actually, you might see here how bigger models have much worse perplexity with 4bit due to the weight outliers: https://github.com/saharNooby/rwkv.cpp/issues/12 For LLaMA models - yeah, different story.
30.
by Ambix 3y ago
Not sure here. The LLaMA models - yes, all weights fit in the small range between -2.0 .. 2.0 And some other models have more crazy numbers with even more crazier outliers within them, like you might have a weight of 12.00 between long arra

More ›