I appreciate you being transparent and also to the point when it comes to the testing. Not wasting anyone’s time, and delivering excellent info. Thanks 🎉
I'd love to see a model that I can give a task and leave it for a few hours to think about it. It could surf the web to find info and so on, just like a human would.
Merci c'est génial ce petit modèle tombe à point, il ne faut pas laisser l'avenir de l'ia entre les mains exclusives des gafam sinon nous serons leur esclaves.
I would like to see the AIs get a chance to correct themselves. Can they get it right after being told they got it wrong the first time? I don't expect people to get things right the first time every time, it's very important that they can correct their mistakes though. Small tangent: I really wish they would separate arithmetic from the rest of math AI. Have a tiny model that can run a calculator, have a big model commit just to algorithms and such. Seems like people AI I is more likely to make a mistake doing the math, than coming up with the correct math to do. /rant
I really believe if companies trained 3b models at the same compute as the big models ( meaning passing data multiple times ) they would be very competitive to the big ones. For example a 3b mixture of experts with 6-8 experts would be great. We need Mixture of Experts scaling laws.
@@CODE7X I don't need an agent to do phd physics or law. I need a fast cheap agent to propose code with minimum cost so it automatically does that 24/7
can you create a video on how to have a small agentic slm or llm running on 4gb 3050RTX or 40GB RAM with i5 10th gen cpu? I was trying to understand how one can create small agentic code generating model which can have only typescript, javascript, deno js and can work with bolt.diy. all your videos only have testing and info about different models. can we have something which can be practically implemented to learn more and customize? I don't see such videos. Like how to choose or create datasets and train own model and use it locally
I have constructive feedback for your next video. Most model reviewers on here do not take into account what the model was designed for and try to benchmark them with general purpose questions. I am assuming most of your audience use these models for other useful things instead of chatting. Here is what I would love to see: Test the model in a few benchmark categories like: -Creative writing -Instruction Following -Needle in the haystack -Function calling -Censorship -Coding -Logic -Reasoning -Multi-turn chat -RAG In addition noone talks about the hyper parameters like temperature and top p to name a few. It can effect the output you are looking for. Usually the benchmarks bring the temp down close to 0. A chart showing the models performance with 1-8bit quants will be amazing. Most people run a lower quant. I know it will require a lot of work but it will really boost the transparency, which right now everyone is trying to hype every model without honest benchmarks etc.
![I.N "HALLUCINATION" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/n5B5q1Hwt_U/mqdefault.jpg)
Having locally small models to think for as much as we want, is amazing.
Your channel makes me think I dont need the other ones tbh, great quality, quick explanations and up to date information the quickest.
Thanks for excellent quality videos which are consistently delivering cutting-edge AI coverage
Thanks a lot for the support!
I appreciate you being transparent and also to the point when it comes to the testing. Not wasting anyone’s time, and delivering excellent info.
Thanks 🎉
I'd love to see a model that I can give a task and leave it for a few hours to think about it. It could surf the web to find info and so on, just like a human would.
Agent zero ai will do that for you lol
Love your content i review so many but I your the best hands down 2minute papers is also a great one but thank you and I wish you have a beautiful day
Merci c'est génial ce petit modèle tombe à point, il ne faut pas laisser l'avenir de l'ia entre les mains exclusives des gafam sinon nous serons leur esclaves.
eh ben dis donc
what is system requerments for this to run it
3GB of ram if you use 8 bit quantization
@Ancient1341 I have 8gb vram and i7 and 16gb ram is good enough
@BBZ101 any modern phone is enough, so yeah.
he said m1 8 gig mac will do for this
Hey king👑! Can you tell me which ai tool are you using to create the super awesome animation intro you use these days
I would like to see the AIs get a chance to correct themselves.
Can they get it right after being told they got it wrong the first time?
I don't expect people to get things right the first time every time, it's very important that they can correct their mistakes though.
Small tangent: I really wish they would separate arithmetic from the rest of math AI. Have a tiny model that can run a calculator, have a big model commit just to algorithms and such. Seems like people AI I is more likely to make a mistake doing the math, than coming up with the correct math to do. /rant
Hi King, what interface are you using? Is that OpenWebui?
Yes.
I really believe if companies trained 3b models at the same compute as the big models ( meaning passing data multiple times ) they would be very competitive to the big ones.
For example a 3b mixture of experts with 6-8 experts would be great.
We need Mixture of Experts scaling laws.
idk bigger the model the more it can learn
Small models are good at making them do something particularly very good but not very generalised version of everything in the world
@@CODE7X I don't need an agent to do phd physics or law. I need a fast cheap agent to propose code with minimum cost so it automatically does that 24/7
@@Gamatoto2038 MOE have solved this, as well as scaling laws ( which are for old school llms)
dose anyone tell me that what he use for Local LLM GUI?.
It's OpenWebUI
@@AICodeKing thanks bro!. and I really like youre video concept.
Any idea how can we make use of this with deepseek in cline? Idea being the small thinker generates the (pseudo) algorithm and deepseek implements it
aider can do that
not useful for code
Please check Hermes Llama 3.2 3B model and compare with Small LMs
3b is not enough i'd prefer nestmind thinkflow that has 11b i think that's a great balance also deepthink is not that bad
Pass of fail dont means much. Analyzing the chain of thought capabilites would be more interesting with these thinking model.
Winner🎉
Make a video about how to make application for phone (ios,android) with no code as ai assistant base on my file as knowledge base
can you create a video on how to have a small agentic slm or llm running on 4gb 3050RTX or 40GB RAM with i5 10th gen cpu? I was trying to understand how one can create small agentic code generating model which can have only typescript, javascript, deno js and can work with bolt.diy. all your videos only have testing and info about different models. can we have something which can be practically implemented to learn more and customize? I don't see such videos. Like how to choose or create datasets and train own model and use it locally
I'm rocking trhe same thing. 3b models are usually pretty great
I have constructive feedback for your next video.
Most model reviewers on here do not
take into account what the model was designed for and try to benchmark them with general purpose questions. I am assuming most of your audience use these models for other useful things instead of chatting.
Here is what I would love to see:
Test the model in a few benchmark categories like:
-Creative writing
-Instruction Following
-Needle in the haystack
-Function calling
-Censorship
-Coding
-Logic
-Reasoning
-Multi-turn chat
-RAG
In addition noone talks about the hyper parameters like temperature and top p to name a few. It can effect the output you are looking for. Usually the benchmarks bring the temp down close to 0.
A chart showing the models performance with 1-8bit quants will be amazing. Most people run a lower quant.
I know it will require a lot of work but it will really boost the transparency, which right now everyone is trying to hype every model without honest benchmarks etc.
Thinks
Merci !
Thanks a lot for the support!
Hi
🤗
these small models are useless and pointless
100% not true. laptops with 3050 and mobile devices can run these. and they are really smart for thier size