Look, Huang is smarter than me and certainly spends more time thinking about AI but fundamentally, advanced AI seems to present employee-type problems, not machine-type problems, so you can't treat it like a simple engineering challenge. We have foolproof methods for building safe machines, but not for building safe employees.

I'm not sure I understand the argument you're making, then. Are you saying the models won't get proficient at hacking if they aren't rewarded for it by escaping the sandbox?

Like why would you think the models are worse at hacking before they get out of the sandbox than after?

HuggingFace employed pretty normal security practices and the models didn't have much trouble breaching their defenses. So I don't think "the models can hack around non-negligent defenses" is speculative.

Basic security practices may not be sufficient for current models, much less future ones! Making unexploitable software is pretty hard

For the CONTROL Resonant launch, we are deploying a hotfix to address enemy health pools and Dylan's combat strength, especially during the early hours of the game. The changes introduced in the hotfix will make the combat more dynamic, retain the challenge, but reduce the friction, making upgrading and respeccing more straightforward. Read more on our blog: controlgame.com/en/blog/cont… #CONTROLResonant

are u being fr now i have to preload 92gb again

Here's pairsort: github.com/ericflo/pairsort ericflo.github.io/pairsort/ (PyPI wouldn't allow jevsort) - created by Opus 5.5

jsort exists!

I'll be giving a talk next week at the @PlanetScale SF Database Meetup about how we made @pgrustdb and how we made it 300x faster than Postgres for analytics

Thanks! Here's a link to the slides: drive.google.com/file/d/1_wD… I believe PlanetScale was recording it, so there's probably a recording somewhere

Might be a good time to share that I’ll be joining @stripe next week to lead design for AI and experimentation. Cannot wait to work with @lil_dill and fam on the future!

On August 13, 2026 at 2:59pm, my little sister Chizoma passed away at the age of 32 after a 2.5 year battle with leukemia. We were firmly planted in our identities of being the youngest boy and youngest girl of the four of us, so I cannot describe the feeling of how lost I feel without her, my only younger sibling. Words cannot describe how much of a light Chizo was in every aspect of people's lives. If I had to describe her in one word, it would be generous. Generous with her love. Generous with her happiness. Generous with taking care of others, even sometimes at her own detriment. The amount of love and care she showed her friends, family, and loved ones was ever-present throughout her entire life, even throughout her battle with cancer. Through this leukemia journey, I learned that she had an O- blood type, the universal giver, which I would joke with her is the most fitting blood type for her personality. In the final weeks before she passed, as I saw in pain and hurting, I was crying while holding her hand and she started caressing my hand, comforting me. I smiled while crying and asked, "Why are you comforting me? I'm the one that's supposed to be comforting you." The day before her last, she commented to our mother on her weight loss and was like, "We need to take you out shopping and get you a new wardrobe, these clothes are too baggy on you." To show this love and care to others even in her final moments speaks to how compassionate of a person she was. I never thought I'd experience a day without the ability to hear her loud laugh. No longer will I be able to feel annoyed by her calling me "my bank account." Even now, I'm yearning to hear her voice and call me that one more time while hugging me, returning back to the old times. There is so much more to say than I can fit in a post. The world lost truly one of the brightest and kindest human beings, and I am honored to forever be your big brother #2 (and bank account). Rest in peace Chizo, I love and miss you so much already and I cannot wait until I see you again.

every time you see the word "intelligence" in the context of LLMs, try mentally replacing it with "reasoning" it changes how i interpret things and much closer to the truth IMO first we figured out how to scale compute, now we figured out how to scale reasoning

But the end product is still reasoning, right? The rest is just debating reasoning efficiency. Like, compute is still compute whether it's RISC or has a bunch of branch prediction and efficiency cores or whatever.