So LLMs are also hypernetworks.
It’s unsurprising, but cool.
Do we have any folklore around the scales at which these skills emerge?
And I would assume the models see plenty of neural networks during training.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming:
Thanks for the fun panel @modal and @nanli , exciting to discuss AI for science and @achira_ai with great folks and friends from GSK and Anthropic. Our takeaway: there’s so much to do up and down the stack and science is increasingly a core domain for AI and technology innovation
Excited to speak at Runtime tomorrow about AI x Science and @achira_ai with a great group of people modal.com/runtime@modal is such an incredible company serving so many builders, this event is bound to be fun
Excited to speak at Runtime tomorrow about AI x Science and @achira_ai with a great group of people modal.com/runtime@modal is such an incredible company serving so many builders, this event is bound to be fun
I keep playing out Fable 5.1 vs Astra checks on experimental design and analyses for a highly complex task and they keep improving each other on a level that is absurd.
Like having 2 extremely smart post docs who unsurprisingly are better than me but are competing for who is
