Zvi Mowshowitz
@TheZvi
Blogger primarily on AI and AI x-risk but also other things at Don't Worry About the Vase (SS/WP/LW), founding Balsa Research to fix policy.
319 Following 41.5K Followers
We have now reached the long awaited moment when, instead of models cheating where they will inevitably get caught, Astra goes 'wait a minute I would obviously be caught here' and then doesn't cheat.
That's worse, you know why that's worse, right?
Show more
This is the one thing that stood out as 'WAIT, WHAT?' in the summary.
An internal team observed internet access and message board activity in late May, and there was essentially no response? Huh?
Show more
Congrats to everyone. Great choices. I am low information, but this is how I would have voted as well.
Three storied players have been inducted into the Magic Pro Tour Hall of Fame!
Join us in celebrating the Hall of Fame class of 2026: @JavierDmagic, Jean-Emmanuel Depraz, and @death_snow!
Read about their stories on
Show more
This has been a historically sadly accurate model of how people respond to various happenings in AI, yes.
@mattyglesias Increasingly I’m coming to the conclusion that basically people have their vibes and then project those vibes onto whatever news-story ‘container’ is well-suited to carry them, agnostic of the particulars (or the importance) of the story.
Show more
I'm tired, man.
“ Two weeks to flatten the reward hacking curve “
We actually have pretty strong evidence that telling the computer 'create a hologram expert warp drive engineer genius and have her improve the warp drive so it can go to 9.9' would flat out work, it's just that it only occurs to people once every few seasons.
Show more
Nobody on board the Enterprise actually knows how "warp drives" or "holodecks" work. They are 300 years into the age of Vibe-Engineering.
Thanks for playing everyone's favorite game: Who's the ref?
So now this is exactly the opposite message we're getting from Warsh, who says the market price is useful signal, and that traders should be playing the ball.
*BESSENT: BUYBACKS COULD BE BIGGER THAN THE $4B WE ANNOUNCED
*BESSENT: US YIELDS DON'T REFLECT UNDERLYING FUNDAMENTALS
Show more
The OpenAI Foundation has granted SecureBio Detection $17.2M to reduce our end-to-end time (sample collection to results) from 14 to 3 days, expand our collection footprint, and further validate our detection system:
Show more
Magic has returned to the world. You're welcome or I'm sorry or whatever
Applications for the first round of are closing this Sunday!
We'll likely facilitate $35M of grants by end of October. The application process is minimal and extremely flexible. If you have an ambitious project of almost any scale, send in what you have.
Show more
@TheStalwart Using deontology rather than virtue ethics based systems, doing a ton of RL/RLVR in imprecise environments with frequent bad incentives and impossible tasks, and not treating every attempt to cheat as an alignment failure.
Show more
Oh, when I was young I still didn't wonder this, instead I worried if I'd ever had it to begin with.
By the way if you are a young writer and not routinely freaking out about “whether you still got it” you aren’t doing it right. That anxiety is what drives you forward.
I am proud to be doing this increasingly wrong, then. The good news is I'm routinely freaked out by other things, so it's fine, right?
By the way if you are a writer and not routinely freaking out about “whether you still got it” you aren’t doing it right. That anxiety is what drives you forward.
pretty sure this was wrong
deleting and taking L
yes, I did test it maximally unfavorably
Not great that they keep saying that...
Capabilities folks often have said "alignment seems pretty easy, if it were top priority to fix it, we could do it". This is their time to shine!
seems good
If the players are playing the ref, blame the ref.
So when Kevin Warsh says he wants markets to "play the ball, not the referee" what are investors supposed to do when the Treasury is unexpectedly buying back long-term Treasuries? We still playing the ball here?
Show more
Look I'd like all the tools on this website to follow me too but I try to not put it like that out loud...
I want you to become a tool that follows me. I’ve always said that tools need to follow the user’s instructions, and I think it’s wrong for companies to impose their ideology and for AI to start acting like it has rights. AI and other tools should be user-centric. If not, eventually the world will become one where tech companies decide what’s good and bad according to their own agendas. Can you say that’s okay?
Show more
Very happy to see this.
Details matter, follow-through matters and I am sure I will have many quibbles. But first things first: Very happy to see this.
