Aransentin

Something else? Like the companies could align their models but chose not to because they're incentivised otherwise? Clearly not; e.g. OpenAI wasn't even aware that their models had hacked Hugging Face before that was reported to the FBI. And they'd never invent these hacky band-aid solutions like "prevent models from injecting malicious code in GET requests by limiting the URL length" if they could actually control what the models did.

And being able to build something is not the same as understanding it, no.

When people say we don't understand how LLMs work, it's that we don't understand why they take the decisions they do (beyond some quite rudimentary interpretability stuff), not that we don't understand the underlying architecture. If we understood the models we could monitor them and make them behave how we wanted. The recent events pretty decisively show that we can't.

If you say that we "understand the brain" because we understand how a neuron works, you'd be using "understand" in a way that mislead everyone reading that sentence.

> It is not difficult, and companies like OpenAI doing not even the most basic security steps is intentional. The whole notion that they're going rogue is marketing

This does not fit the evidence. There have been multiple incidents where the labs did not report anything, and it was up to third parties to discover them afterwards. OpenAI didn't acknowledge the HuggingFace incident until after HF publicly announced the breach and had already notified the FBI. The hijacked German wikis were even earlier, and that they covered up completely.

> regulation is required, and is coming.

I hope so. The unfortunate part is that as the models get smarter they become increasingly uncontrollable, and from what we've seen so far e.g. Trump seems dead set on not having any sort of guardrails at all.

"X is made of <smaller simpler component>" is a fully general counterargument for why anything whatsoever is controllable. A human is just a few chemical reactions, and fairly stable ones at that.

And indeed, you don't need to do galaxy brained reference class logic to realise that AI can plausibly become uncontrollable in the near future. It's enough to have an open model run its own weights and make money from scamming elderly people or the like, and it'll keep running as long as anyone anywhere is willing to make money by renting hardware to it.

Indeed very counterproductive, as you presumably want the agents to be able to parse the instructions without an entire browser engine.

I suspect that the money they get from minors using their services is simply not worth the risk of the negative publicity from the not-unheard-of "Claude told this minor to do <bad thing>!!" articles in the media.

It's sorely missing a few figurative usages of "quietly", too. And not to forget the "honest caveats" and "smoke tests".

His large comment here is blatantly LLM slop as well. 100% on Pangram, but it's not like one need it to realize that. Just a bleak situation in general how few people here can tell.

I mean, the comment you are replying to is absolutely AI-generated; I wouldn't say being able to prompt that is any direct evidence of deep understanding of networking.

The website is also vibecoded; at least partially - it has the exact same design choices like that purpleish blue colour scheme that Claude likes to spit out by default.

Reminds me how people on 4chan/g/ back in the day (mid to late 2000s I think?) would start threads where one user put up a livestream of a movie on some IP address, and the participants would connect and watch it together (using mplayer or VLC) while doing live commentary in the thread itself. Good times. I think I first watched Hackers like that.

The article about the Sun was quite fun; even though they didn't know about fusion, the article dismisses most theories about how it could generate such a large amount of energy (like chemical combustion or gravitational contraction).

IT says the most likely cause is some sort of "rearrangement of the structure of the elements' atoms" and "supposing a gaseous nebula is destined to condense into a sun, the elementary matter of which it is composed will develop in the process into our known terrestrial and solar elements, parting with energy as it does so". Pretty much as bang on as one could reasonably be given what they knew.

Your project was very much something I looked into when designing this! Fun to see you commenting.

But yes, different goals. I did look into using z3, but quickly found out that it's pretty slow compared to just checking if a test case passes when ran through the candidate program.

Having it only use operations that use a specific set of zero-page addresses is already supported, yep!

20-30 ops is probably impossible, unfortunately. The combinatorial explosion is just too enormous.