MIXTRAL 8x22B: The BEST MoE Just got Better

No video

MIXTRAL 8x22B: The BEST MoE Just got Better | RAG and Function Calling

Рет қаралды 4,415

Prompt Engineering

Күн бұрын

Пікірлер: 38

@engineerprompt 4 ай бұрын

Disclaimer: This video was pre-scheduled before the release of llama3 and was the best model at the moment :) So some of the info in this video is outdated by a week. Crazy how fast things are moving

@rusdad1304 4 ай бұрын

Fixed

@sankyuubigan 4 ай бұрын

это определенно хорошо. надеюсь скоро искусственный интеллект сможет создавать порноролики с красивыми девушками

@geniusxbyofejiroagbaduta8665 4 ай бұрын

Did it beat llama3 70b

@s0ckpupp3t 4 ай бұрын

Probably more uncensored which is a win

@electric_mind 4 ай бұрын

Nope, LLaMA 3 70B has beat Mixtral 8x22B on almost all benchmarks with decent margin

@elawchess 4 ай бұрын

@@electric_mind but presumably those benchmarks use the base version of Mixtral 8x22B

@dasistdiewahrheit9585 4 ай бұрын

@@s0ckpupp3t I don't think so. In my experience all uncensored flavors are worse than there censored versions. And since there exist working jailbreaks if one really needs it, I don't use uncensored models (means de-censored models) anymore.

@s0ckpupp3t 4 ай бұрын

@@dasistdiewahrheit9585 de-censored is probably worse than a censored base model yes, re-alignment is still a form of lobotomy and brute forcing the weights. But comparing base model to base model a less censored model is far more useful

@scitechtalktv9742 4 ай бұрын

Where is the link to the Colab notebook?

@AubzMan 4 ай бұрын

Where indeed it is. Unless the omission somehow act as a signpost for 'sign up for 'Advanced RAG' which has recently covered by LanceMartin of LangChain

@bertobertoberto3 4 ай бұрын

Excellent video!

@malikrumi1206 3 ай бұрын

Although I heard you say that you were imagining a scenario in which you had two different, specialized vector stores, it made me want to ask you:: Is there a capacity limit on vector stores that isn't present on traditional rdbms systems? Thx.

@kishoretvk 4 ай бұрын

Good one

@mohsenghafari7652 4 ай бұрын

Hi dear friend . Thank you for your efforts . How to use this tutorial in PDFs at other language (for example Persian ) What will the subject ? I made many efforts and tested different models, but the results in asking questions about pdfs are not good and accurate! Thank you for the explanation

@pawan3133 4 ай бұрын

When the LlamaIndex's LLM calls multiplication function then does it use python interpreter to actually run the function and get the answer? Or does LLM try to do the calculation by trying to figure out the function?

@engineerprompt 4 ай бұрын

I think can the LLM does the calculations. For simpler calculations like these its not going to be an issue for bigger models but for more complex operations you might want to have another tool which does the computation using python interpreter.

@pawan3133 4 ай бұрын

@@engineerprompt I checked, the model calls the function in python and passes the arguments to the function and then picks the output and display it. Quick way to check: define a function that uses np but don't import np and it will throw an error 😀

@barackobama4552 4 ай бұрын

Lama i have a question, for RAG is best gpt 4 or this model? And what about gpt3 vs this model? Appreciate your answer

@engineerprompt 4 ай бұрын

I think gpt4 will still give you the best results. This model will probably be better than gpt3.5. But for RAG, you need to consider the embedding model that you want to use. That will play a major role.

@barackobama4552 4 ай бұрын

@@engineerprompt Thank you i really appreciate your answer Lama

@Techonsapevole 4 ай бұрын

is Llama3 better in function calling?

@engineerprompt 4 ай бұрын

Haven't really seen function calling with llama3 yet. Will have to wait for that.

@adriangpuiu 4 ай бұрын

@@engineerprompt dolphin llama 3 on ollama model list

@mohamedkeddache4202 4 ай бұрын

Is the Mistral API free ?

@engineerprompt 4 ай бұрын

Mistral api is not free.

@8eck 4 ай бұрын

But still, Mistral/Mixtral was moving this industry forward for a long time. RIP

@pawan3133 4 ай бұрын

Why RIP? MoE is still great

@8eck 4 ай бұрын

Guess it is useless now after Llama 3 release.

@engineerprompt 4 ай бұрын

I wouldn't say that. llama3 is great but its only 8k context but this model supports upto 64k context. This alone limits the use of llama3 in a whole bunch of applications.