hey @OpenAI I'm gonna burn a @tszzl mention token to kindly request \parr to be added to the latex formatter (web & desktop, cmll package) my eyes are burning 😂

Going to bed these days

Dear @Battlefield, the moment the game ends for you, matchmaking penalty ends. Instantly. Not when the animation ends. Pls fix. this sucks

did apple tv just leak silo ep 10?

hey @tszzl, you know what would be an actual good solution? the big labs should start targeting _every single company out there_, in a systematic way you release reports, share 0 days in an appropiated manner, etc this will lead to a safer world, and faster timelines.

I hear this argument all the time that people in academia are not there "for the money" and therefore, so the implication, you can trust them. It's true in the sense that academics didn't chose their profession to become rich-rich. But it's untrue in the sense that they want to keep their job because most of them haven't learned anything else. They have very strong economic incentives to want to keep their job, even though it won't make them super rich, they probably have a cozy position. And that is why they exaggerate the relevance of their work and ignore shortcomings, because the show has to go on, they need the next grant, the next job, they need to publish the next paper. No, you cannot trust them to be honest about the relevance of their work, and for the most part you can't trust them to be honest about their colleagues work either, because their income depends on what their colleagues think about them.

“we don’t know how to get to ASI so let’s pretend”

I wrote about why we believe the future is for everyone. More coming about a positive vision for a world with superintelligence soon.

I believe the highly technical term for this situation is "tough shit." Achieving technical breakthroughs does not entitle you to economic rent in perpetuity. The period under which you can extract rent depends entirely on: - your customers' price sensitivity - switching costs - Who your competitors are and what they're willing to do - your ability for your competitors to copy your product - legal recourse They just happen to be in an industry where it costs an exceptional amount of money to be the leader, but there are some fairly significant problems when it comes to long term rent generation.

you know someone is too deep on their own kool-aid when they can’t look at the mirror and laugh about themselves, even for just a tiny bit what would absolve anthropic is just admitting that at least once they might have overreached they just have to meet in the middle

100%. I already think this aggrandizing labelling has been culturally disastrous, but we must prevent further damage. AI corps is a good shorthand I like.

I think it's important to stop using the word "labs" when describing the AI companies "Labs" infuses them with a nobility they don't deserve They are for-profit companies like any other for-profit companies and deserve to live or die as a result of the value they provide

dario pulled the plug out of the GPUs because everyone's bullying him into signing the open weight manifest

thank you @sama

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

Orthologic Type Systems arxiv.org/pdf/2507.10482 Another great paper. Give the type system union, intersection, negation, variance, and explicit subtype assumption, but drop distributivity. And that is the RIGHT THING to do since Ax(B+C) should not be the same type as (AxB)+(AxC). Isomorphic, sure, but distributivity confuses extensional equivalence with representational identity.

I guess CLL is ambiguous with something /shrug

I was laughing and now experiencing the european mindset first hand the universe has a preferred direction: the one that pumps heat in the other is inherently evil