antirez

75 ms·

781.
by antirez 3y ago
Use llama.cpp for quantized model inference. It is simpler (no Docker nor Python required), faster (works well on CPUs), and supports many models. Also there are better models than the one suggested. Mistral for 7B parameters. Yi if you wan
782.
by antirez 3y ago
This one is very good, and will provide certain key insights on the way you should think at NNs. -> https://www.amazon.it/Deep-Learning-Python-Francois-Chollet/... This is a good explanation of the Transformer detai
783.
by antirez 3y ago
I agree with you, stavros. There is no transfer between C coding and ML topics. However the original question is a bit more in the business side IMHO. Anyway: I've some experience with machine learning: 20 years ago I wrote (my first n
784.
by antirez 3y ago
* Open source models: give you all the attention (pun intended) you can get, away from OpenAI. At the same time do a great service to the world. * Maybe in the future, bigger closed models? Make money with the state-of-art of what you can
785.
by antirez 3y ago
1. This is an open source model that can run on people's hardware at a fraction of the cost of GPT. No cloud services in the middle. 2. This model is not censored as GPT. 3. This model has a lot less latency than GPT. 4. In their endpo
786.
by antirez 3y ago
Tobi is the founder of Shopify.
787.
by antirez 3y ago
llama.cpp support needed ASAP, not just quantized, but I hope it will be albe to run in a way that keeps the same two experts for minimum N tokens generated and re-valuate the expert to select only then, so that it will be possible to do th
788.
by antirez 3y ago
I believe that the rationale is that if you can do an outstanding 7B model, it is likely that you are able to create, in the near future, something that may compete with OpenAI, and something that makes money, too.
789.
by antirez 3y ago
The comments were full of people downloading it and providing insights. The previous Mistral model was released in the same way
790.
by antirez 3y ago
I really hope there isn't some organized anti European AI mob flagging posts. Because this stuff was really weird.
791.
by antirez 3y ago
The first post accumulated 200 votes in 1h and got killed for no apparent reasons.
792.
by antirez 3y ago
Appreciate the words of support! Thanks. Very happy you managed regardless of scoliosis.
793.
by antirez 3y ago
Yes, I have this vibe about many library authors in the Python/Ruby/Javascript ecosystem at this point. I don't want to be part of that, nor as a user, nor as a coder.
794.
by antirez 3y ago
To the point that Linux, probably the most important C project in the world, was written in the GCC dialect for many many years. I think it's no longer the case.
795.
by antirez 3y ago
I believe the C++ way would be the right thing for C. Note that this does not mean in any way to inherit all the stuff from C++, nor even constructors and stuff like that. Simply a way to avoid doing this: list_add(&mylist,1);
796.
by antirez 3y ago
Sure, C could be much better right now. In some incredible way people that are able to drive the future of C made all the worst choices, undermining the language instead of improving it, for most parts (with undefined behaviours). C just ne
797.
by antirez 3y ago
The Thompson story is very telling about C and how semantically powerful it is. KT tried to rewrite UNIX in B at some time, but he failed. Assembly allows you to do many interesting things because it is a low level language, and yet you can
798.
by antirez 3y ago
Thank you Jay! It's a great thing to read in the morning :) Appreciated. I don't believe in God, but in a very "soft" way, that is, I'm happy other people enjoy their religious life.
799.
by antirez 3y ago
Author here, thank you for having this project on the front page! I want to explain why this bot is in C, while, for instance, another project of mine that implements a LoRa driver/protocol for the ESP32 (a much lower level project, in
800.
by antirez 3y ago
This depends on the quantization level. Sure, they lose part of the performance, but are still quite useful and cheap to run.
801.
by antirez 3y ago
Just use llama.ccp and run the quantized model like yi-34b-chat.Q4_K_M.gguf (you will need a 32GB Macbook, likely).
802.
by antirez 3y ago
Wow, this is really significantly better than Mistral 7B and other 13B models I tried. And runs on my M1 max at decent speed.
803.
by antirez 3y ago
Because of outstanding execution of OpenAI technical folks. An execution that has nothing to do with Altman. Similarly Mistral 7B model has much better performances than others. There is some smart engineering plus finding the magical param
804.
by antirez 3y ago
It's a lot better than that. OpenAI is just very good execution of publicly available ideas / research, with some novelty that is not crucial and can be replicated. Moreover, Altman himself contributed near zero to the AI part i
805.
by antirez 3y ago
Linkbait at its finest. But it's true that the Python AI stack sucks big times.
806.
by antirez 3y ago
I noticed that on HN lately calories-deficit driven weight loss is considered controversial. I wonder what do you think about creationism VS evolution, folks.
807.
by antirez 3y ago
Estimating calories spent with the Apple Watch is not the best way. Just calculate your TDEE as an average, and in a conservative way (if you go to the gym 3 times a week and sit for your daily work you are moderately moving at best). The
808.
by antirez 3y ago
The problem with idle scan isn't much scanning itself, but understanding trust relationships between hosts. This allowed to target the attacks much better. But I get what you mean in general.
809.
by antirez 3y ago
In Catania (Sicily) it is very common to eat "mauro", that is, Chondrachantus teedei,
810.
by antirez 3y ago
Great description. And most importantly, it was possible to choose C IP address to test if other machines on the network had special privileges and ports open. This allowed to penetrate C to gain access to B.

More ›