What is it? AtMan is an attempt at helping users know the source of the answer/completion given by LLMs (Large Language Models). Unlike other existing methods, AtMan is memory efficient and is also multimodal (i.e it works on both images and text). Existing methods to explain the outputs of neural networks are generally classified into one of the 2 given types: Gradient based methods: Require a backward pass, which takes up almost double the memory as that of a forward pass....

On this sunday, we plan to gain a basic understanding of how quantum circuits work by making the world’s most glorified not gate. The original video can be found here. The x operation (i.e not) The x operation simply flips a state. x|0> = 1and x|1> = 0 Jargon alert: x|0> means “x gate applied to 0”. In fact, we can also represent this gate with matrices! In this matrix world, 0 is represented by [[1], [0]] and 1 is represented by [[0], [1]] (one-hot vectors)....

Imagine that you’re a 21 y/o student in a hostel with a really bad wifi. On top of that, you’re bound by the timings of the lunch and dinner, and you cannot just leave your laptop open while you’re out because then your very near and dear friends would mess with your laptop’s .bashrc. What you need, is a way to start long running scripts on your remote machine via ssh and log out without stopping the it, so that you can check back in later on....

The author tries to explain knowledge neurons to himself. Not much to see here...

Trying to understand how Andrej Karpathy’s minGPT works, step by step....

Docker for people who don’t dont really want to spend too much time on docker...

Trying to learn a bit of solidity because why not...

Figuring out how an image might be worth 16x16 words...

stuff you need to understand before understanding how transformers work...

Why do feel like I’ll have to come back to this post many times...