Zvi

AI #138 Part 1: The Peo­ple De­mand Erotic Sycophants

Monthly Roundup #35: Oc­to­ber 2025

Trade Es­ca­la­tion, Sup­ply Chain Vuln­er­a­bil­ities and Rare Earth Metals

OpenAI #15: More on OpenAI’s Para­noid Law­fare Against Ad­vo­cates of SB 53

2025 State of AI Re­port and Predictions

• Zvi 9 Oct 2025 21:34 UTC

  6 points

  0

  on: Real­is­tic Re­ward Hack­ing In­duces Differ­ent and Deeper Misalignment

  Would a reasonable way to summarize this be that if you train on pretend reward hacking you get emergent misalignment that takes the form of pretending (playacting) misbehaving and being evil, whereas if you here train on realistic reward hacking examples it starts realistically (and in some ways strategically) misbehaving and doing other forms of essentially reward hacking instead?

• Zvi 9 Oct 2025 15:28 UTC

  2 points

  0

  in reply to: Daniel Kokotajlo’s comment on: Bend­ing The Curve

  Yes.

• Zvi 9 Oct 2025 15:28 UTC

  2 points

  0

  in reply to: cousin_it’s comment on: Bend­ing The Curve

  Knowing that, hopefully you wouldn’t?

• Zvi 9 Oct 2025 15:25 UTC

  8 points

  2

  in reply to: nostalgebraist’s comment on: AI #137: An OpenAI App For That

  Oh, of course, how silly of me!

AI #137: An OpenAI App For That

NEPA, Per­mit­ting and En­ergy Roundup #2

Bend­ing The Curve

Med­i­cal Roundup #5

• Zvi 3 Oct 2025 12:08 UTC

  2 points

  0

  in reply to: aphyer’s comment on: AI #136: A Song and Dance

  I was not aware of this at the time.

Sora and The Big Bright Screen Slop Machine

AI #136: A Song and Dance

Claude Son­net 4.5 Is A Very Good Model

Claude Son­net 4.5: Sys­tem Card and Alignment

On Dwarkesh Pa­tel’s Pod­cast With Richard Sutton

• Zvi 29 Sep 2025 11:53 UTC

  2 points

  0

  in reply to: james oofou’s comment on: More Re­ac­tions to If Any­one Builds It, Every­one Dies

  My guess is that on the margin more time should be spent improving the core messaging versus saturating the dialogue tree, on many AI questions, if you combine effort across everyone.