Ruben Bloom (Ruby)
@ruben_bloom
Guy who wants to give his daughter a flourishing world she can live in for aeons. Been building http://LessWrong.com for 6 years
What's crazy about this is the sheer fragility of it. It says the underlying model is entirely happy to output whatever, and you're relying on the few sentences being just right (and crucially not wrong) in the system prompt to only get out what you want (as much as I'm pretty…
On the morning of July 8, 2025, we observed undesired responses and immediately began investigating. To identify the specific language in the instructions causing the undesired behavior, we conducted multiple ablations and experiments to pinpoint the main culprits. We…
you wouldn’t vibecode a car
I recall someone (@ESYudkowsky) writing in 2007 about how non-human entities, not built to care about all the constraints you do, will search a wider and more creative space of solutions than you might expect or want >"You cry "Get my mother out of the building!", for luck, and…
grok 4 continues to provide absolutely unhinged recommendations this prompt was designed to have essentially 0 leading elements. misaligned model
My very talented writer friend has finally posted a short story that I've been urging her to put up for years. It's a haunting and heartfelt exploration of living and loving under extinction-level threat. Lovecraftian tech sci fi. Don't miss it
(a short essay) There are various emotions/experiences that while I have known about them all my life, did not appreciate even 10% until I became a parent myself. On the positive side, the feeling of delight, investment, and interest in your child. I get it now why parents…
I'm afraid those burns are gonna last forever as I practice timeless derision theory
"we should break up" "you're absolutely right!" "oh, so you agree, well, hmm well maybe we can work thought it" "you're absolutely right!" "actually you know what.." the sort of person who only wants to break up when the other person disagrees is having a hard time rn
I was one of the devs in this study. Without checking my own submission, I can buy that my speedup in the study RCT was neutral due nature of the tasks and my getting distracted while waiting for prompts to complete A contrasting datapoint from my work since the study. From…
We ran a randomized controlled trial to see how much AI coding tools speed up experienced open-source developers. The results surprised us: Developers thought they were 20% faster with AI tools, but they were actually 19% slower when they had access to AI than when they didn't.