Doing a little, everyday [devlog 2]
Yesterday was a day off. It was unintentional but according to this one video CGPgrey made, taking an off on wednesday and sunday is more effective than the standard saturday sunday weekend. Still experimenting with the idea though. I haven’t really thought of what I want to do today, there seem to be many things I need to do and it’s getting hard for me to keep track of everything. So for today I’ve decided to take small strides towards iiys. I’ve noticed that when it comes to mammoth tasks like this, the easiest way to draw out motivation when I’m feeling lost is to come up with really precise next steps. Right now it’s to figure out what APIs I’ll need for the most basic implementation. I’m assuming API calls from my local machine will suffice. Those APIs would include gemini, a vector database, a transformer embedder.
I realise that this reads more like a diary entry than a serious devlog that discusses only the project relevant bits, I’ve yet to find the kind of blogs I wish to write, for now, I won’t discount anything. Let today be a diary, who knows what tomorrow will be. I want to say with confidence that tomorrow definitely will be something. Anyway that’s enough rambling, let’s start searching.
It would seem I was charged 72 paise for something, I’m guessing it was because I saved some prompts and the storage costed. Hopefully it isn’t because I used gemini, coz that was supposed to be free according to this
Alright, for some reason I found myself looking at how PostgreSQL and MySQL are implemented. I’ve decided to go with PostgreSQL because vector support is more mature in it for AI applications. Though MySQL seems to be faster in general, I’m using this database primarily for vector searches. Plus since it’s OSS I can always pull that data to use locally when the time comes.
Also! ChatGPT just got long-term memory! If I can figure out how that was done I can add memory to my agents as well. I’m guessing it’s some sort of vector database. (Some searching later) Looks like the implementation is kept secret but here’s an excerpt from their website on how it works:
Memory works similarly to Custom instructions, except that we’ve trained our models to update the memories rather than requiring users to manage them. That means that when you share information that might be useful for future conversations, we’ve trained the model canto add a summary of that information to the notepad.
Like custom instructions, memories are added to the conversation such that they form part of the conversation record when generating a response. The model has been trained to also manage memories, including by intelligently updating pre-existing memories, consolidating memories where appropriate, and forgetting memories when instructed by the user.
I’m quite liking this style of devlog. I’m writing it as I’m working so it helps me keep track of what needs to be done. Normally I’d put such things in my obsidian note, which I still believe I should since I doubt my ramblings would be super helpful to people in general. Maybe I can have an LLM rephrase the obsidian note to contain only devlog relevant details? Will try that next time.
Anyway, looks like I’m done for the day. See you around!