how far away are more personalized AIs? For example github copilot but it'd know about your specific programming style and how you like to solve problems
is this possible with LLMs?
#How far away are personalized AIs?
prisma vortex
sry dont know why i marked with math
tender cypress
Not far at all. Just look at older recommendation systems (ie Netflix from 5 years ago). If you're asking in regards to personalized LLM experiences, it's also not far away. The problem is that you'd have to tailor the experience to each person which requires a lot of tweaking and work on the side of the company. It's easier (and more cost effective) to just make 1 general AI that "resembles" a custom experience than actually develop a framework so that everyone gets a custom experience.
Bringing it back to netflix and their recommendations, nowadays the recommended are 9/10s full of movies/shows that Netflix wants you to watch, rather than what aligns with your actual tastes. Those shows pushed on you are usually considered popular or well received based on the aggregated viewing times of everyone else who uses the platform (but netflix can also be pushing some less popular shows because they want to make a return on investment for those projects).
prisma vortex
Not far at all. Just look at older recommendation systems (ie Netflix from 5 yea...
thanks
what about something like programming on your computer and at the same time talking to an AI that sees what you are doing and is personalized so that it knows how you like to solve problems?
right now github copilot:
- can work as a programming buddy
- only a vscode extension
- not personalized at all, but still useful
what are the challenges when extending this to a personalized programming buddy that 1. talks how you want 2. voice 3. just understands you better
Is it like the algorithms, the data, compute power or something else?
tender cypress
Those models are quite a lot of power to use. Services like that are going to have to be paid APIs. That comes with latency issues depending on people’s network speeds and how busy the API resource is.
There is also the issue of having those services aware of your actions across different apps. It’s extremely invasive and requires the developers to account for a lot of edge cases.
Integration into people’s computers is going to be hit or miss depending on the OS. Windows is gonna want their own AIs integrated with their OS just like Apple would. People also may not want to opt into those services.
This is just to get these services integrated into peoples machines
To have a personalized experience, you’d require additional compute power to be expended just in finetune it on your data. Not worth it regardless of how many people want it
As people generate more data, you’ll have to finetune again
prisma vortex
To have a personalized experience, you’d require additional compute power to be ...
so are LLMs just not ideal for personalized AIs?
tender cypress
Kinda yeah
tender cypress
The only way a personalized AI is viable is someone found a way to template one (ie MemGPT) and people then put that open sourced template on their devices
prisma vortex
The only way a personalized AI is viable is someone found a way to template one ...
interesting, what resources should I study if I wanna know more about personalized AI?
If I want to get into personalized AI stuff in particular, is it critical to understand machine learning and the math used in AI, or is a high level understanding enough? (for example rn you can build AI applications with just a general understanding of AI)
right now im learning about things like fine-tuning, context length, embeddings etc
should i also be learning about gradient descent, eigenvectors etc? (ie how the models actually work)
tender cypress
it's not going to be hard to do personalized AI stuff in the context of using LLMs and existing models. There are also a lot of pretrained AIs for each of the components you want (text generations, speech to text, and text to speech) as well as relatively simple tutorials on how to train/finetune them.
That said, there is A LOT within AI that is good to learn about. #🚀・getting-started-in-ai-ml is a good resource to get you started with fundamentals. Topics like gradient descent and eigenvectors are not brought up too much in AI (especially Deep Learning, where you use neural networks).
I usually recommend that you get an understanding of fundamentals before diving deep into the various aspects of AI.