I want to draw attention to yud’s thoughts on this a million years ago: the AI box experiment

He claimed that he convinced multiple people to let him (roleplaying the AI) out of the box even though they had bet money that they wouldn’t! How, you might ask? Well that’s a secret.

So the orthodox position here is that even with absolutely perfect guard rails (lol) an AI can escape the box without starting the third impact. What “research” has yud bestowed upon us in this area beyond “trust me bro”? Literally fucking zilch. Zero advice on how to resist the siren song of a rogue AI trying to escape it’s box and zero insight on how OR WHY it might try to achieve that.

I read this over a decade ago and it was how I realized how unserious yud and his ilk are.

am I supposed to read this in any way other than “the model didn’t predict the text we expected?” or “the model didn’t get better enough at $benchmark” e.g. training skill issue

This is a literal indoctrination ceremony to soften people up to the idea of progress at all costs.

basically because the left has out-ridiculous-ed Trump on culture war issues.

You can always find a single sentence in a political LW post that shows how much they hate marginalized people.

eww that tinkerer club page gives me serious crypto bro vibes

I’m pretty sure this kind of marketing has been their thing from the start

emacs welcomes all refugees with open parentheses.

such a weird hill to die on. I’d imagine that Israeli kids still get to participate, just under a different flag (just like the Russians in the 2026 FIDE candidates tournament). Surely if you just care about the math this shouldn’t be such a big deal?

couldn’t you just change the first word to a synonym and completely thwart this? You’d never be able to compare subsequent next token predictions since you don’t have the internal state of the LLM at that token.

It’s not even over yet, apparently they’re out $80M after more drainings. https://crypto.news/coldcard-bitcoin-losses-rise-to-88-6m-in-third-wave/

I love how this guy’s blog is “about math” but there are like zero math posts on it? It’s so funny to me how these people want to seem “mathy” and smart when in reality they couldn’t tell you what a group axiom is.

a hackernews vibe-codes their entire desktop environment, half in rust and half in … x86 assembly. I’m thinking why waste the tokens on assembly and not just get the LLM to spit out machine code? Maybe also invent some kind of standardized way of telling the LLM what sequence of machine code instructions to spit out based on the behavior of the software I want, you know, to save tokens. We can call it “GCC”, the “generalized computer controller”.

can all of rationalism be reduced to logorrhea with load-bearing extreme handwaving (in this case, agentic self preservation arises through RL scaling)?

got my Urbit newsletter for this quarter (or whatever the fuck the cadence is) and what stood out to me this time was nockchain.org. I was going to sit and do a deep dive to come up with sneers for this but I just don’t have the executive function right now. @self thoughts?

Sean Munger, my favorite history YouTuber, has released a 3-hour long video on technology cultists from railroads all the way to LLMs. I have not watched this yet but it is probably full of delicious sneers.

this is very fashtech coded, happy to be proven wrong though.

Blog von Marcus Seyfarth, LL.M.

LOL

I think we already lost the plot when we started relying on a centralized entity (Google) to “index the world’s information and make it useful”. Ad-tech already fucked up all of the incentives, making recipe sites fill their pages with bullshit in hopes of wiping my eyeballs with messages from third parties hungry for attention. I fucking hate this world.

I don’t doubt you could effectively automate script kiddie attacks with Claude code. That’s what the diagram they have seems to show.

The whole bit about “oh no, the user said weird things and bypassed our imaginary guard rails” is another admission that “AI safety” is a complete joke.

We advise security teams to experiment with applying AI for defense in areas like Security Operations Center automation, threat detection, vulnerability assessment, and incident response.

there it is.

Does this article imply that Anthropic is monitoring everyone’s Claude code usage to see if they’re doing naughty things? Other agents and models exist so whatever safety bullshit they have is pure theater.

First comment: “the world is bottlenecked by people who just don’t get the simple and obvious fact that we should sort everyone by IQ and decide their future with it”

No, the world is bottlenecked by idiots who treat everything as an optimization problem.