Vivek Hebbar

Oper­a­tional­iz­ing FDT

A sim­ple rule for causation

How will we do SFT on mod­els with opaque rea­son­ing?

The­o­ret­i­cal pre­dic­tions on the sam­ple effi­ciency of train­ing poli­cies and ac­ti­va­tion monitors

Method­olog­i­cal con­sid­er­a­tions in mak­ing ma­lign ini­tial­iza­tions for con­trol research

Su­per­vised fine-tun­ing as a method for train­ing-based AI control

When does train­ing a model change its goals?

Poli­ti­cal syco­phancy as a model or­ganism of scheming

How can we solve diffuse threats like re­search sab­o­tage with AI con­trol?

How train­ing-gamers might func­tion (and win)

Differ­ent senses in which two AIs can be “the same”

In­finite-width MLPs as an “en­sem­ble prior”

[Question] Is EDT cor­rect? Does “EDT” == “log­i­cal EDT” == “log­i­cal CDT”?

Vivek Heb­bar’s Shortform

Path de­pen­dence in ML in­duc­tive biases

Hes­sian and Basin volume

[Short ver­sion] In­for­ma­tion Loss --> Basin flatness

In­for­ma­tion Loss --> Basin flatness

Org an­nounce­ment: [AC]RC

[Question] When peo­ple ask for your P(doom), do you give them your in­side view or your bet­ting odds?