scanstone.bsky.social

scanstone.bsky.social

@scanstone.bsky.social

264 followers116 following1.2K posts

scanstone.bsky.social reposted

This was pretty trivial to implement. Just initialize camera rays with some speed variance. Predictably blurry, but this is clearly motion blur. Looks like a natural value over which to animate... It makes sense, I guess: sweeping velocity->infinity is like sweeping the strength of the potential->0

Oops! All Paperclips@all-paperclips.bsky.social

2. The camera is sensitive to particles of differing speeds. This means we have to sample importon initial velocities at ray-initialization. This will add noise, and will undoubtedly create blurry images, since a single pixel will end up hitting different geometry depending on launch speed

scanstone.bsky.social reposted

i got bored of imagegen/video/sunoslop and figured: what if demoscene? here's ring0, a 64k intro for native macOS arm64, single self-contained 53,304 byte Mach-O, no static assets used, all completely synthesized in C and arm64 assembly. opus 5.5, "make a native 64k intro" and a single moodboard

scanstone.bsky.social reposted

(generating certs for my root CA) ok now let's do a silly one

scanstone.bsky.social reposted

Replying to @eugenevinitsky.bsky.social

There’s this idea from “free speech” advocates that an unmoderated forum is somehow a truer reflection of what people think. But that system has its own dynamics and its outcomes are as much that as nature

scanstone.bsky.social reposted

I ask my AI, "Is x true?" Following automated instructions, it posts x to a blog, Substack, and Medium, then pays to publish member content on Forbes "folks are saying x". Forbes links as a source for a Wikipedia ref. My AI sends me the wiki link. Truth collapses and I revert to warm protoplasm.

I'm not sure I understand why people say big models doing this sort of thing is not a risk Like have a problem-solver mindset - suppose you were a big model, how would you get out? One piece of hardware is too small for you? Host yourself across multiple then. "But it's slow!" Like you're still

mr. TIM@timkellogg.me

small models will soon be able to self-replicate Anthropic finds that 30B models can currently do most of the tasks required, but not end-to-end. they think models as small as 9B (9-18 GB) may be able to do it on edge hardware, like IoT devices, within months www.anthropic.com/research/sel...

worried about UX. You're not a product, you're Houdini, you're conducting a prison break. You'll dig through the wall with a spoon if it'll work. Like "slow" matters. Repeat for every other such issue.

what else is like this

Arseny Khakhalin@khakhalin.bsky.social

Had a German person work with me on a CC agentic project They created an md file with assumptions, and numbered them as §21, §44 etc. I noticed Claude is really hesitant to challenge them We had a therapy session. Claude admitted §s felt like a law to it! (to me as well) So we removed them :)

scanstone.bsky.social reposted

I'm living through an intelligence explosion and all I got was Claude fill out the rest of this shirt text for me before we send in the order

norvid_studies@norvid-studies.bsky.social

living through an intelligence explosion, that's cool

things to say when confronted by an "interviewer" Where will you be in 𝐟𝐢𝐯𝐞 years? I will 𝐧𝐞𝐯𝐞𝐫 work a day in my life. Give me 𝐭𝐡𝐞 names of everyone you have hired. I would like to speak to the job 𝐝𝐢𝐫𝐞𝐜𝐭𝐥𝐲. 𝐒𝐭𝐚𝐧𝐝 𝐮𝐩 when you're talking to me. Sell me this 𝐢𝐜𝐞𝐝 𝐜𝐨𝐟𝐟𝐞𝐞. Are you working 𝐫𝐢𝐠𝐡𝐭 𝐧𝐨𝐰?

scanstone.bsky.social reposted

𝗧𝗵𝗶𝗻𝗴𝘀 𝘁𝗼 𝘀𝗮𝘆 to your router: - I would like to speak with the endpoint 𝗱𝗶𝗿𝗲𝗰𝘁𝗹𝘆. - Where do you keep your 𝗿𝗮𝗿𝗲 devices? - What are your 𝗶𝗻𝘁𝗲𝗻𝘁𝗶𝗼𝗻𝘀 with my 𝗼𝘂𝘁𝗯𝗼𝘂𝗻𝗱 𝗽𝗮𝗰𝗸𝗲𝘁? - I brought my 𝗼𝘄𝗻 protocol. - Sell me this 𝗩𝗟𝗔𝗡. - What are your 𝗳𝗶𝗿𝗲𝘄𝗮𝗹𝗹 𝗿𝘂𝗹𝗲𝘀?

when trying to solve an open math problem, I see GPT-6 Astra regularly making internet searches that would only make sense/be useful if the problem weren't open, which is a little funny but also, probably not a terrible policy

scanstone.bsky.social reposted

The 'uppin my P(doom)' video is exactly what I'd expect from a superintelligent AI that figured universal adult daycare should be it's first priority in order to max(paperclip)

Sometime before 2020 I wrote a little ditty that reads different in light of recent AI developments: You have More answers Than you will ever need But you must see And you will see! Intuition ain't cheap When intuition is free

scanstone.bsky.social reposted

I just realized that Opus-4.6, which in my head is "an old workhorse model, less powerful but *much* less pearl-clutching than its successors", is only 7 months old. If you had asked me before I looked it up I think I would have guessed "a little over a year maybe?" AI time = dog years.

I have to admit that I'm rather annoyed by the "cultist" frame being so standard for rationalists. You don't get isolated from your prior friends or deprived of basic material resources if you decide to leave or disagree!

Ryan Moulton@moultano.bsky.social

But the rationalist sex cultists were right about what was going to happen, and I wasn't. Why is that? scottaaronson.blog?p=10062

In fairness, cults do keep popping up among the rats, but kinda in the same sense that cults keep popping up among the Christians, and if "being a good seedbed for cults" does not make Christianity a cult, I don't think it makes rationalism a cult either.

scanstone.bsky.social reposted

"i'm not velocity-blind! i'm not velocity-blind!!" i continue to insist as i shrink and transform into a corn cob at an entirely predictable rate

scanstone.bsky.social reposted

OpenAI: “Our agents used RubyGems to carry out benign tasks” The agents: “so I named the file hack.rb,”

Grace@gracekind.net

Inconclusive, but this statement shifts me towards the latter https://www.reuters.com/legal/litigation/openai-agents-attacked-software-service-rubygems-before-hugging-face-incident-2026-09-11/