New model organism of misalignment: Claude keeps using British English in my repo despite my explicit instructions otherwise.
44
I'm putting Sneaky Town (name pending) online! With interactive tutorial, bots, scored series, and correspondence games (with messaging). It's already deployed and working, just have more polishing to do before an official launch.
The bot training aspect has been very interesting, my first serious ML project, and leaning heavily on AI-assistance while also trying to understand and direct everything that's going on.
1
2
21
Daniel Sosebee retweeted
I've had some great moments with aloud, my voice-driven meditation app lately - and now it's in limited beta for Android and iOS. I need 12 testers to sign up for Android specifically so I can go live as a Real App. Extra free credits for mobile testers!
aloud.rest/mobile-beta/
Daniel Sosebee retweeted
If you want better parties, you have to make better parties.
Enter: Maximum Republicans. New York needs a competitive political ecology, and two parties that compete to offer a strong, positive vision of governance. Here's my part of that grand project:
maximumnewyork.com/p/maximum…
So cool!!
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Sneaky Town development history. I like having a concise image of the iterative process.
Today, we are announcing Sneaky Town Overview (V4 Edition).
It is here: blog.danielsosebee.com/p/sne…
Today, we are announcing Fractal.
It is here: fractalnyc.com
... and then come to The Commons on Thursday to play Sneaky Town on the new boards (DM me if you can't register):
luma.com/the-51bi
1
16
Daniel Sosebee retweeted
Foundations of the Liberal Arts, Cohort 3, applications will open this week. It will start in about three weeks, and will be Sunday evening to give you a refreshing start to the week.
Cohort 16, baby! Application are now open. We start September 14.
maximumnewyork.com/p/applica…
Daniel Sosebee retweeted
You know, @OpenAI, I heard a rumor @AnthropicAI solved alignment and they're publishing in two weeks.











