@AtharvaRNaik on the reading list. would love to understand the blog in detail. in case you are interested do check out this work https://t.co/staFiPDhkC
We (@VanditPatel002) just completed a "Hello-world" task for Weight-Space Programming and open-sourced it.
Can an LLM figure out what a neural network was trained to do, simply by inspecting its weights without running a single forward pass?
We tried this on a small MNIST classifier with random labelling. The results were fascinating.
A thread 🧵
@tfmlab@prathoshap So if we permute the weights itself, then I think llm answer will change because it explicitly sees which weights falls in the initialisation range and based on that figures out which pixels are such which are never changed.
We need to yet test it past mnist.
@JohnKocdr@prathoshap I am feeding raw weights as text. but i have mentioned how the weight looks like. you can check the github repo . we have weight.txt and prompt.txt.
Here we are interested to see how llm are able to understand the network.
@TensorSlay@prathoshap the motive behind the experiment was also to understand how LLM understands other networks. If it is able to understand it is no longer a black box. and if we by some means learn to understand it , Networks will no longer be black box and we will be able to better use them
@SriSan68@prathoshap Actually i dount that it replaces probes . weights can tell what the network could do. inputs tell us how the weights are actually fired up. so like a useful thing to do is weights proposes mechanism and probes are used to test it.
@SriSan68@prathoshap this is particularly a good experiment. digits with loops. but i think agent will figure that out. just a wild guess. if you note they were always moving here and there around a similar point. but a nice exercise to do
@SriSan68@prathoshap they were not guessing , just that their hypothesis were different. we had provided just one data point, and the weights thats all. so in such cases there is a high chance that majority of hypothesis might be wrong and hence averaging out is not a good idea.
@harrykesh067@prathoshap we cant. we want to see if LLM is able to figure out on its own. does it have the knowledge to do so. Such a study helps us in understanding some unknown network.
@SynderesisAI@prathoshap so Astra in one call was not able to name the task. however when we had 4 agents (with each of them having 3 subagents) , one of the agent exactly predicted the task upon providing a very minor hint. the other 3 were very close though(just one or two digits off)
The tech industry convinced an entire generation of developers that they can skip the math.
They are wrong. You cannot build foundational architecture with a "GenAI in 5 days" bootcamp. Real engineering requires staring at the equations until they make sense.
If you actually want to build state-of-the-art models, you cannot skip the math. No fluff, no quick-bytes. Just the mathematical foundations of Deep Generative Modeling.
Here is a course I designed to put everything you need to learn.
Namaste X!
This is my first post here - I'm humbled to release my app: VedaVaaNi.
It is an interactive learning platform designed to help you practice Rig Veda and Krishna Yajur Veda chanting with absolute phonetic and prosodic (Chandas) precision.