Видео

oui oui ;D

thank you :)

声音好好听👀

sure, you will also get the indian accent along with it 😂

​@@meigustas depends what part of india 😅 ,mother tongue influence btw I tried. Typical accent i as indian i couldn't speak probably old generation hv that since they were not that good in english

Do you happen to know any cheaper alternatives?

@@KevinxJin If you're on a tight timeline for product delivery, just build out the working solution first with openai. The cost will likely come down in future release cycles, as this is only the beta version. Here are other alternatives: - VAPI (this is probably the easiest way to get started with building a prototype) - Telnyx as phone call orchestrator + LLM (I built this out last year, a lot of hassle and lots of latency) - Amazon Lex (good for on prem deployments for data paranoid companies but from the demos it sounds very robotic) - Deepgram (there's a waitlist for this, latency is still present from what I've seen)

hey! small world :D

@@mahakleung6992 test test

@@meigustas I will do this in sections. Let's find out what YT's big problem.

@@meigustas 妳是什麼 "Mei"; 美國的美? 妳的視頻不錯! 1 of 6

@@meigustas First video. Well done and informative. I haven't played around with Llama 3.2 yet. My three main models were ORCA1 13B, ORCA2 13B, and MIXTRAL 8x7B. I am pretty happy with the MIXTRAL as it is well suited to NVidia RTX 4090 ... 46B in all with a 32K context. Given the MOE architecture, there is very low latency to the first token with a throughput of around 30-40 T/S. I constructed a memory the MIXTRAL using ChromaDB. I need to post process the MIXTRAL's memories before ingestion; which I do very simply now. I have considered using Llama 3.2 for post processing given its small foot print and very large context. I believe the best way to improve VDB model memories is by improving the quality of the memories themselves. I am planning to give Llama 3.2 a try for that. 2 of 6

First video. Well done and informative. I haven't played around with Llama 3.2 yet. My three main models were ORCA1 13B, ORCA2 13B, and MIXTRAL 8x7B. I am pretty happy with the MIXTRAL as it is well suited to NVidia RTX 4090 ... 46B in all with a 32K context. Given the MOE architecture, there is very low latency to the first token with a throughput of around 30-40 T/S. I constructed a memory the MIXTRAL using ChromaDB. I need to post process the MIXTRAL's memories before ingestion; which I do very simply now. I have considered using Llama 3.2 for post processing given its small foot print and very large context. I believe the best way to improve VDB model memories is by improving the quality of the memories themselves. I am planning to give Llama 3.2 a try for that. 2 of 6

@@brabbbus got it thanks for the feedback

I don't mind it much actually​@@meigustas