Ambix

14 ms·

31.
by Ambix 3y ago
No need to convert models, 4bit LLaMA versions for GGML v2 available here: https://huggingface.co/gotzmann/LLaMA-GGML-v2/tree/main
32.
by Ambix 3y ago
Looks like ADHD person with high IQ and curiosity (like me). You could cure the problem with medicine or just become better managing it with age. From my experience, eventually you'd become more selective with your ideas and more invol
33.
by Ambix 3y ago
And many professional studios still using tape machines and emulations over the final songs to produce that fat, warm sound of analog era. Like this Ampex recreation I love to use with my Apollo and Luna: https://www.uaudio.com&#
34.
by Ambix 3y ago
Yeah, it's really so bad on desktops. With my LLaMA AVX implementation on 32bit floats [0] there no performance gain after 2 threads, so remaining 14 threads available are of no use, there no memory bandwidth to load them with work :)
35.
by Ambix 3y ago
It's not about threads number, it about memory bottleneck. Sweet spot for my M1 Pro laptop is around 6 threads and 4bit model - I've managed to get 20 tokens per sec, really impressive
36.
by Ambix 3y ago
Is it Zen1 architecture? It should be much better on Zen2 and newer Epycs
37.
by Ambix 3y ago
Happy to announce new version of LLaMA.go - open-source CPU inference framework for GPT models. That's a big milestone, we've embedded scalable server which allowing access to GPT model with simple REST API. Now anyone is able to
38.
by Ambix 3y ago
LLaMA.go - open-source framework for LLM inference on regular CPUs [0] It took me about a month of full-time, hard, day and night coding (including weekends) to finally build a solid piece which can handle some crazy CPU workloads of tensor
39.
by Ambix 3y ago
Not the TS, but that's actually the same goal I have in mind with [0] project. Right now I'm building my homelab server which aimed to fit 1 TB RAM and 2 CPUs with ~100 cores total. It will cost like 0.1% of what I need to pay for
40.
by Ambix 3y ago
> "a picture is worth a thousand words" and it might be opposite for the GPT models actually. it's just easier for humans to grasp the bunch of knowledge with one eyes sight, but usually most of useful information might be
41.
by Ambix 3y ago
I'm developing framework [1] in Golang with this goal in mind :) It successfully runs relatively big LLM right now, and diffusion models will be the next step [1] https://github.com/gotzmann/llama.go/
42.
by Ambix 3y ago
Could you point to some actual examples of using FFT for NNs in real projects / GitHub codebases? Would like to dig more into the thing.
43.
by Ambix 3y ago
Python needed only to converting PyTorch models into more accessible bin format :) All the inference and tensor math code is 100% Golang. I'm going to inject AVX2 assembler soon, but that's too very specific to Go ecosystem.
44.
Show HN: Llama.go – port of llama.cpp to pure Go(github.com)
25 points by Ambix 3y ago|5 comments
45.
by Ambix 4y ago
How cool is that! There's also exciting news about early access for GPT5 13T model [1] features from Noosphere AI [1] https://noosphere.chat
46.
Launch HN: Noosphere (VC W23) – First GPT5 chatbot with superhuman abilities(noosphere.chat)
1 points by Ambix 4y ago|1 comments
47.
by Ambix 4y ago
Noosphere AI released GPT5-based chatbot today. Spoiler: - 10x bigger than GPT4 - 3.14x smarter than average human - Trained with Chinese and Japanese texts - Egyptian hieroglyphs as well - Computing costs doubles the whole Croatia GDP Join
48.
by Ambix 4y ago
> I haven't yet found a replacement for Django Take a look here https://adonisjs.com
49.
by Ambix 4y ago
--------------------------------------------------------- Location: Remote Remote: Yes Willing to relocate: No Technologies: Golang, Rust, PostgreSQL, Docker, micro-services, clouds, queues and caches Résumé/CV: https:&#
50.
by Ambix 4y ago
Yes, paper books on hard tech just easier for performant and/or thorough reading. And I just love the taste and look of real books.
51.
by Ambix 4y ago
I've had the playground very similar to those shown on the first photos in USSR school back 30 years ago. It's hard to imagine what dangerous plays we had there.
52.
by Ambix 5y ago
~25,000 Russians moved there right after the incident. So maybe many will return or move farther.
53.
by Ambix 5y ago
There's a big difference between democracy and political chaos. I went through all 90s in Russia - believe me, there were so many freedom and democracy around, that no one in modern America could ever imagine :) Despite all of that, al
54.
by Ambix 5y ago
Exactly. And there still many people shocked by NATO'1999 bombarding of Serbia.
55.
by Ambix 5y ago
What about the same advice for the citizens of Iran, Iraq, Afghanistan? Should they have fought for their liberty against US troops the same way?
56.
by Ambix 5y ago
https://en.wikipedia.org/wiki/NATO_bombing_of_Yugoslavia That was a really eye-opening moment for many people, I felt the same as now about Ukraine.
57.
by Ambix 5y ago
Yep, I remember the real excitement in Russia about US around 1990+ and Kosovo was like a cold shower for many believers there.
58.
by Ambix 5y ago
Thanks for Kotkin recommendation, looks very interesting
59.
by Ambix 5y ago
I've worked with ArangoDB recently and very much liked their AQL. How Edge compares with Arango?
60.
by Ambix 5y ago
Laravel is great, but sometimes just overkill for simple tasks. That was the reason for me to start PHP Comet, it's like SlimPHP on async steroids: https://github.com/gotzmann/comet

More ›