This Changes Everything - Going Over Gemini 1.5 Pro Technical Report
Вставка
- Опубліковано 16 чер 2024
- This is an overview of Gemini 1.5 Pro's technical report. If you are interested in the future of AI, this is not to be missed!
I'm Logan, an engineer and creator who's committed to making AI accessible for everyone.
A lot more to come! Stay tuned!
🙋♂️ Follow Me
- Twitter/X: / logancyang
- Github: github.com/logancyang
🙇♂️ Sponsor to Support My Work
- Buy me a coffee at: www.buymeacoffee.com/logancyang
- Sponsor my GitHub projects: github.com/sponsors/logancyang
📖 Chapters
0:00 Intro
0:41 The Report
13:40 Summary
#geminipro #openai #gpt4 - Наука та технологія
Just found your channel, great video, subscribed! Keep ‘em coming!
Thanks for subscribing! Will do!
Thx.
This makes a lot of sense, but doesn't this require the system to maintain the entire injected tokens/context in memory?
By system do you mean the model or do you mean something on top? The model's API is stateless so it won't maintain the context across calls, but there are options researchers are looking into where you can have a kv-cache for the model to rely on, it essentially becomes RAG more or less.
The way I think these models are going to be used is, call it with a super long context for a few times that cover as much ground as possible, then cache the result and serve similar requests from the cache.
@@loganhallucinates I was thinking of this more inline if these size context windows become normal for local open source models. When running RAG on those we can still just retrieve what we want into memory/standalone message. If we load an entire potentially huge knowledge document into context we need to maintain that in system memory. or is this oversimplifying it?
@@stephenthumb2912 For future long context window local models it would be similar, since the chat completion API is still stateless, each call is still independent and require you to send over your chat history + long context for every call, you can tell that's incredibly wasteful especially for long context models.
But for sure there will be system on top tailored to these long context models to avoid that. It could essentially be a stateful API of sorts, like the OpenAI Assistants API. But I can think of other solutions too.
Long story short, engineers are going to make it efficient soon.
I tried using my Gemini Advanced API with your Obsidian plugin but it just keeps saying "can't connect to Internet"? (Cohere API for embedding)
My vault has grown to almost 2.5 million words now, across 1500 md files, all geared towards one project, I need this 1.5 pro more than the usual case use.
what exact error message is it? Does it at least work in Chat mode for you?
@@loganhallucinates thanks for responding... it's been in default ChatMode, and then when I ask it to specifically address the note open, with [[title]], it says "I apologize, but I do not have access to your Obsidian notes"🤷♂️
@@101RealTalker oh i see, you need to use the little up button to pass the active note to it, or use the command “set note context” to set it using title or path or tags. My next video will be running a mini Needle-in-a-haystack test using this method on local models, stay tuned!
@@loganhallucinates I of course did that already and all it says back to me is, when I give it a basic benchmark test with the cliche "please summarize", bullet returns is "sure, here is a concise version of my introduction:
I am obsidian co-pilot, an AI powered assistant for obsidian, I can help you with-"... Yada yada 🤷♂️🤷♂️
thub nail is dead on.
buy a good mic
what’s the issue now?
seems fine to me
@@mattiawatchingvideos-bg1ok thanks! I guess it’s my position of the mic. It’s a bit too close so it gets plosives. Will try moving it away in my next video