Fabien Roger

Fermi es­ti­ma­tion of the im­pact you might have work­ing on AI safety

The im­pact you might have work­ing on AI safety

How To Know What the AI Knows—An ELK Distillation

A Mys­tery About High Di­men­sional Con­cept Encoding

By De­fault, GPTs Think In Plain Sight

Ex­tract­ing and Eval­u­at­ing Causal Direc­tion in LLMs’ Activations

The Translu­cent Thoughts Hy­pothe­ses and Their Implications

Some ML-Re­lated Math I Now Un­der­stand Better

What Dis­cov­er­ing La­tent Knowl­edge Did and Did Not Find

How Do In­duc­tion Heads Ac­tu­ally Work in Trans­form­ers With Finite Ca­pac­ity?

LLMs Some­times Gen­er­ate Purely Nega­tively-Re­in­forced Text

Sim­plified bio-an­chors for up­per bounds on AI timelines

Pass­word-locked mod­els: a stress case for ca­pa­bil­ities evaluation

When AI cri­tique works even with mis­al­igned models

If in­fluence func­tions are not ap­prox­i­mat­ing leave-one-out, how are they sup­posed to help?

Will early trans­for­ma­tive AIs pri­mar­ily use text? [Man­i­fold ques­tion]

Pro­gram­matic back­doors: DNNs can use SGD to run ar­bi­trary state­ful computation

Prevent­ing Lan­guage Models from hid­ing their reasoning

Coup probes: Catch­ing catas­tro­phes with probes trained off-policy

Some nega­tive steganog­ra­phy results