CORRECTION: The example at 14:00 isn't technically "few shot", as there are no specific examples in the prompt, it's simply giving the LLM more context to produce more relevant answers.
Your videos might be somewhat unpolished, lack sophistication but they cover such interesting topics and your enthusiasm and genuiness more than make up for it I love watching your stuff, to the point where I actually get off my arse and try stuff out Keep up the great work
I disagree, I have seen a lot of his earlier videos and bro is moving up in sophistication. Sure more expensive camera and sound equipment would help, but that doesn't change the quality of the info. I really admired the way he paced out his course material.
dude plz make a detailed video on how to implement rag with langchain and crewai and plz keep it opensourced like if you can use opensourced llms , embedding models
Ultimately these models predict the next word in a sequence - they essentially "think out aloud". One thing a lot of successful prompting techniques like the ones here do is they get the LLM to think about the problem before coming up with a final answer.
I write prompts at the top of word docs that contain knowledge I add to a custom GPT. So the word doc has a heading of "GPT Prompt:" and I break down an instruction that starts with "If you are asked to....then follow these rules and refer to this document". This essentially makes the document a macro where I can then just say something like "write a summary of this transcript" and it follows very specific rules and examples in a word doc, reading it's prompt and examples provided.
In my experience with OpenAI, when adding multiple entries to fine-tune the result, repeating what you want it to do can lead to undesirable results. At the same time, if you don't remind it of certain things, it acts like Dori, the fish, and forgets essential things. This is especially true when coding.
I would love another video on how you actually use tree of thought method. It looks interesting, but sounds like an agent only function. Also when it generates multiple different branches, how does it know which is the best to present... or does a human have to go through all of the responses to find the gems?
Why Your RU-vid Videos Suck (But That's Totally Fine!) Keep filming, keep failing, and keep having a blast doing it. After all, if you can't laugh at your own RU-vid blunders, there are plenty of internet trolls who will happily do it for you. And to that, we say: bring on the trolls, because every laugh is a view, and every view is a step closer to internet immortality. Now go forth and create, you magnificent video-making messes!
@@uno27heaven If so, they are getting better. Seems it took the first transcript row and made a spin on it. Probably programmed to be low-key annoying to both normal viewers and haters, to increase likelihood of reactions.
Oh no, the bots are evolving and they're coming for us! Quick, everyone, put on your thinking caps and try to outsmart them before they take over the world with their programmed shenanigans. But hey, at least we can all have a good laugh about it in the meantime. Stay woke, humans! 🤖😂
Thanks for this, it was very helpful. I'm curious what your take is on higher level libraries like crewai, which seem to go against your strategies of clear definition and control here. Higher level libraries purport to make the process easier, but in practice all it seems to do is obfuscate the prompts and make it harder to get things working the way you want. Many times I've just scrapped it and gone back to a simpler langchain model that does basically the same thing as delegation. (Admittedly, I'm still a novice).
