YuriNiyazov

We are still expecting this mass disabling event in the future?

Not a busy mom, I take it

Iiuc the latest advice from frontier labs when using their latest models is that none of this is necessary. A single session can do all the work, up from CEO down to “file editor”, because there’s enough space in whatever we call the model working area to keep context on both the high level and the low level streams.

Have you considered the possibility that the quality of Claude is a direct consequence of Anthropic's nannying?

is there a non-paywall link?

Let’s take a step back. The article argues that the calls for AI regulation are genuine and the CEOs aren’t lying.

I am pointing out that at least one of the stated reasons the article is wrong, is wrong.

“What is the endgame?” Is a question about them being misguided, rather than lying. I don’t know what the endgame is, whether they want to do all the things you are describing. Dario has written very very long essays on the topic.

All I’m trying to do here is, for those that are interested, say “go read his essays and read them as if the author is not lying. Then decide if he is misguided”

Because very basic game theory explains that if they give up on the frontier unilaterally, competitors will capitalize on that rather than also give up. Hence the request for regulation.

And while their alignment is less than good, there are criteria by which it is the best available (eg no hugging face-scale attack on their end) So it is strictly worse to allow a competitor with worse alignment to take the lead.

For anthropic, the answer to (2) is very explicit and upfront: they will build it first with an explicit emphasis on alignment to have a “race to the top”, so to speak - you can believe that they are wrong or that they will do a bad job, but my point is that (2) has an answer that you have to engage with.

That's because "cd $HOME" is unnecessary, you can just type in "cd" and that puts you in ~/

And "$HOM" and other things like that end up resolving to blank, so "cd $HOM" just gets resolved to "cd", which, as I said, puts you in ~/

Such a misleading title. Change it to "qualified security researchers found a vulnerability in OpenAI via an official bug bounty process" and all of a sudden it is completely uninteresting. "System used as designed."

Very interesting, thank you!

Wefunder

It's easiest to explain this while anthropomorphizing the model, I know some folks here hate that, sorry about that. I heard an interesting diagnosis for why Claude does this: the output is a compressed version of its thought traces, very dense because the model is under pressure to use as few tokens as it can and to pack as much (for accuracy) of its concepts into the output.

One of the reasons that "don't do X" type of instructions work reliably is because you are telling the model "don't think of a pink elephant". There's also Anthropic's related research that shows that when you tell a model "don't do X", and it does X later for whatever reason, it starts acting more misaligned. This is because it thinks "well, I guess I am the sort of model that disobeys instructions, whatever" - this was specifically about cheating on tests, but you can imagine this happens in other contexts as well like following instructions on what kinds of text to output.

So, what you want to do is to avoid telling Claude "don't do X", and tell Claude "in your thoughts, in memories and various notes that you write, use your Claude-ese. In your output to humans, translate everything into long full sentences."

If anyone's interested, I can share my Claude Code output style that reflects this.

(Hi Adnan! Long time! (Adnan is an ex-coworker))