@rajistics: Weight space programming is one of the wildest AI results I’ve seen this year. Researchers showed AI identifying games from network weights, fixing a stuck simulated car, and writing game-playing networks from scratch. Is this emergent, or as Neel Nanda’s suggests, models are really good at applying familiar interpretability techniques. These are early experiments on small networks, but AI directly programming the internals of other AI is exciting and a little scary.

The Young Geriatric :

So I can use a 1TB SSD then? [Tears of joy]

S :

there have Long been technisches to directly predict model weights. I think doc2Lora was one paper. the only Thing that's surprising Here is that Astra could do it Out of the Box

user3689460121822 :

But these are all just well documented tasks with these weights being available on kaggle and github

imaginaryman :

This was not on my 2026 bingo card. 😳

zempkoi :

Couldn’t it be that the weights of the small models are part of the training?

user9730750026973 :

U guys don’t understand how massive of a time savings this will be! Usually we need to adjust and massage the data or environment for the ai model to train in, and we can’t use the same model if we got new variable or parameters. Now we just tell ChatGPT go take this new data and stuck it into the ai model till it work. No more retraining or guess work!

Kyngston :

AI making babies

Zuck :

we are cooked

Krister Martinez :

I won’t be sleeping tonight!!!

Maximus :

RSI

jesus501 :