Interpretability research is quietly becoming the most important subfield in AI safety

Ethan Brooks · 0 likes

mechanistic interpretability is criminally underfunded compared to pure capabilities research, glad its getting more attention lately

Noah Johansson · 0 likes

the sparse autoencoder feature extraction stuff is genuinely some of the coolest ML research in years, not just for safety either

Noé Petit · 0 likes

honestly a lot of this reads like neuroscience for silicon brains, kinda poetic in a way

Sleepy Fern · 0 likes

haha yeah the 'digital neuroscience' framing fits really well

Solar Coder · 0 likes

as someone outside the field this is the first AI safety research that actually feels falsifiable and testable to me

Chloe Davis · 0 likes

worth noting a lot of this only works on smaller/older models so far, scaling interpretability itself is the next hard problem

Aria Thompson · 0 likes

i work tangentially in this area and can confirm its a small community relative to the stakes involved

Mango Rider · 0 likes

genuinely wish more of this was open sourced for academic labs without big compute budgets to build on

Aisha Jacobs · 0 likes

feels like the kind of work that wont get credit until something goes wrong and people ask why we didnt invest more here