Igor Ivanov

How eval aware­ness might emerge in training

Bet­ter evals are not enough to com­bat eval awareness

Main­stream ap­proach for al­ign­ment evals is a dead end

Call for Science of Eval Aware­ness (+ Re­search Direc­tions)

What is an eval­u­a­tion, and why this defi­ni­tion matters

Com­par­a­tive Anal­y­sis of Black Box Meth­ods for De­tect­ing Eval­u­a­tion Aware­ness in LLMs

Igor Ivanov’s Shortform

LLMs are Ca­pable of Misal­igned Be­hav­ior Un­der Ex­plicit Pro­hi­bi­tion and Surveillance

I repli­cated the An­thropic al­ign­ment fak­ing ex­per­i­ment on other mod­els, and they didn’t fake alignment

It’s hard to make schem­ing evals look re­al­is­tic for LLMs

LLMs can strate­gi­cally de­ceive while do­ing gain-of-func­tion re­search

Psy­chol­ogy of AI doomers and AI optimists

5 psy­cholog­i­cal rea­sons for dis­miss­ing x-risks from AGI

Let’s talk about Im­pos­tor syn­drome in AI safety

Im­pend­ing AGI doesn’t make ev­ery­thing else unimportant

6 non-ob­vi­ous men­tal health is­sues spe­cific to AI safety

What is ev­ery­one do­ing in AI governance

A cou­ple of ques­tions about Con­jec­ture’s Cog­ni­tive Emu­la­tion proposal

How do we al­ign hu­mans and what does it mean for the new Con­jec­ture’s strategy

Prob­lems of peo­ple new to AI safety and my pro­ject ideas to miti­gate them