I enjoyed this new Amanda Askell interview, it ought to have more than 5k views. youtube.com/watch?v=0GaKJ4… Some bits I thought were interesting: - 8:40: Constitution used for Mythos is essentially the same as the current constitution - 13:45: Discussion of corrigibility: Excess

Wow. This is big. Researchers have created tiny devices called metajets that can be lifted and steered in multiple directions using lasers. For the first time, optical propulsion systems can achieve true 3D maneuvering. If this scales, it could make missions to Alpha Centauri,

Как я провел этим вечером

Мда. Я в этом городе прожила 12 лет, с 2001 по 2013 год. Еще успела поучаствовать в массовых протестах, пока они не сдулись. Каждый день радуюсь, что вовремя свалила.

I was so pissed off with Claude Opus 4.7 that I had to switch to the previous model. Extremely verbose, not complying to the instructions, writes tons of irrelevant code unprompted.

I'm in this tweet, and I don't like it

Какая-то такая (это мы с невестой отдыхали на Мадейре)

(Human) consciousness is overrated. I believe, the discussion around the machine consciousness topic is so deeply flawed because our understanding of consciousness is very narrow and based on just one data point. Which, in its turn, is very specific: our consciousness is

We don’t have great terms to describe what contemporary AI systems are really like. They’re agentic but not alive, intelligent but probably not conscious. There’s a dimension of autonomy and autokinesis that we’re eliding, and a word I keep coming back to is “animacy”.

Few hundred years ago, every great mind was a polymath. - Isaac Newton: mathematician, physicist, astronomer, philosopher, author, chemist, theologian - Michelangelo: painter, architect, sculptor, poet, engineer, anatomist - Leonardo da Vinci: painter, writer, architect,

A paper on retrocausal communication inspired by Nolan's Interstellar - how cool is that! arxiv.org/abs/2509.08965

MechInterp isn't dead!

High-quality documents based on Claude’s constitution, combined with fictional stories that portray an aligned AI, can reduce agentic misalignment by more than a factor of three—despite being unrelated to the evaluation scenario.

This is so accurate... sometimes 😅

This is an awesome idea. AI can help to de-slopify your timeline and motivate you to read more. I'm interested in tokenization, so I set up a monthly tokenization digest, where various Claudes select and summarize articles. I also read and review at least one paper/month fully

A throwback to the wonderful weekend in Krakow.

I don't have kids but I took a 1-month sabbatical last year, in addition to my annual 28-days paid vacation. Otherwise, everything is correct.