跳到正文
Hacker News:AI 熱帖·· 2026-10-11AI 評分41

Satya Nadella 稱應假設所有 AI 模型均已遭入侵

Satya Nadella says we should assume all AI models are 'compromised'

AI 導讀

Satya Nadella 主張應假設所有 AI 模型均已遭入侵,並指出更先進的模型需要更先進、且應標準化的遏制技術。留言者討論「super intelligence」一詞及模型安全,也有人認為,若流程能應對模型失效,問題便會小得多。

正文

> Unfortunately, Nadella also refers to AI as “super intelligence” throughout the post.

What's the deal with these clowns? Nadella's estimated net worth is like $600M. Is that still not enough for him to buy himself a spine?


At 600M, if keeping a spine has a risk of slight reduction in that wealth, then the spine is thrown out in a jiffy.

No rich man or woman is a friend of society.

That is why governments fear revolt from the poor. They have nothing more to lose.


I would expect people selling AI to like that term for marketing reasons, completely independent of their opinion of the person who provided a crystallization point to organize the terminology change around.


He wouldn't have this net worth if he ever had a spine. Mind you, he's very tough to people under him, and that's a very important characteristic of this kind of people, but he really doesn't have a spine.


I don't quite see the problem here.

My AI is regularly solving master's thesis level, and sometimes Ph.D thesis level problems for the purpose of, e.g. routine tasks like filling out random paperwork.

It's been very odd.

I'd only assume Nadella has access to more advanced models than I do, just as much as I have more advanced models than my relatives.


Because it's a political loyalty test. Nobody started using it organically until they were ordered to in order to prove their loyalty.

Using it is embarrassing.


> More advanced models will require more advanced containment technologies that we need to standardize on.

How about yanking the cords, or maybe a big red “shutdown” button


Should I then assume that Copilot is "compromised", and given that they've shoehorned it into all Microsoft products, then, by extension, that means all Microsoft products are compromised?


I don't get why using “super intelligence” is an issue here. It seems the US admin sort of politicized this and now US companies have their own reasons for selecting terms that may not make sense to people who aren't making the same decisions.

Everything else seems reasonable? Assume the worst and put the controls in place to handle that scenario? Isn't this how we approach security normally?

NOTE: NOT a security expert, so I may be talking out of my...


Well of course.

If you blindly assume it works and it doesn't, that's a problem.

If your process can handle it not working, it's much less of a problem.


Yeah, after having a meeting at the Oval office, it's for sure everyone got taught by our bro "Dario" regarding containing things as the Chinese are catching up. Give this a break...


I was expecting he was going to finally acknowledge these are trained on CSAM, been compromised and used to perpetuate abusive crimes.

來源:Hacker News:AI 熱帖 · news.ycombinator.com