saanvi

cs + neurobio @caltech / 19 / prev @stanfordmed, @seattlechildren
life is programmable! 🧬🧠💻
part-time musical poet and writer

[saanvi @saanvisub]

3 midterms this week but i spent my whole weekend on this:

our brains don't run on backprop, so why do our models?

i built apical dust on top of q labs' dust, a feedback pathway for forward-only pretraining. https://t.co/U9XeQwNjzw

[saanvi @saanvisub]

@dmmiller597 cool work- important to see if protein design models can reason or just fake for an optimizer

[saanvi @saanvisub]

cool blog post! some thoughts:

- so intrigued abt using agents for interpretability, since it's such a tedious yet purely computational discipline

- humans should spend time where their intuition & research taste matters most

-> connecting findings into this "science of ai"

[Tom McGrath @banburismus_]

https://t.co/42s6eNq5Fw

[saanvi @saanvisub]

@jerryfan_07 my attention span's too cooked for that but i'm an audio overview warrior😂

[saanvi @saanvisub]

say what u want abt gemini but imo its the best at explaining concepts and producing readable output, esp w notebooklm

[saanvi @saanvisub]

@jerryfan_07 thank you!!

[saanvi @saanvisub]

3 midterms this week but i spent my whole weekend on this:

our brains don't run on backprop, so why do our models?

i built apical dust on top of q labs' dust, a feedback pathway for forward-only pretraining. https://t.co/U9XeQwNjzw

[saanvi @saanvisub]

@adamnemecek1 hey, two main points- 1) our neurons are unidirectional and there's no known mechanism for them to mirror forward weights to pass backwards, 2) the brain doesn't have a global error system to every neuron.

more detailed answer in the blog post!

[saanvi @saanvisub]

again, check out the full post on https://t.co/U9XeQwNjzw if you're interested!

also thanks to opus5.5 for the snazzy animations :)

[saanvi @saanvisub]

so... local rules don't fix forgetting, feedback as a control variate doubles sample efficiency at small k, and bias, not noise, is what will break scaling.

excited to continue testing whether killing directional bias keeps zeroth-order training stable under adam!

[saanvi @saanvisub]

@borrasque_ ppl from wealthier backgrounds can afford to take risks and deviate from a safer path, plus they have the advantage of their network

[saanvi @saanvisub]

@lihanc02 ur welcome

[saanvi @saanvisub]

instead of spotify wrapped im gna need a claude wrapped @AnthropicAI

[saanvi @saanvisub]

@imjustnewatai this made me giggle