Hacker News Reader

Comment by arisAlexis | Hacker News Reader

"The reason this distinction is critical is because these models are, in fact, nothing but mathematical functions. They take as input a bunch of numbers, and calculate and return a different bunch of numbers. They don’t do anything themselves — they can only calculate numbers."

This is entirely false, sounds like Andersen that seems incapable of understanding that agency and goals will be a thing.

View on HN

I'm going to go against the grain here and say that the quoted statement does not seem entirely false at all. It is reductionist, but it highlights the fact that the models are indeed just generating numbers confined to a GPU, and don't have inherent agency. Some human has to wire up the model to the internet and enable it to interact with the world by whatever mechanism, and even then it's limited to the digital world, unless there's an API to stab someone or derail a train.

We should focus more on the making infrastructure safe from abuse and not how to cripple AI models for them to not realize how to do harm. More like 'How to ensure humans don't drop atomic bombs on each other' rather than 'How to make atomic bombs safer to drop on people'.

>We should focus more on the making infrastructure safe from abuse

Humans are too fucking lazy to actually accomplish this. We hook more and more internet connected systems to things with a digitally controlled motor every day with things like "#disable in production, set for testing, disabling 2FA to..." that then get pushed to the customer.

> and even then it's limited to the digital world, unless there's an API to stab someone or derail a train.

Fly some airplanes into some skyscrapers, etc.

Litigation has been enough to put actual standards on aviation software. There just need to be one or two bad industrial accidents to make air gapping a standard practice elsewhere.

Would these measures have been adequate to prevent the 911 attacks though?

Yeah, the "it's just math" argument isn't really convincing anymore. Previous to this, math never had opinions about Taylor Swift songs.

> Previous to this, math never had opinions about Taylor Swift songs.

Nor do they now, of course.

Actually they do. You need to make a philosophical argument as how this different than us humans manipulating bits and neurons to come up with opinions. Good luck with that.

If you phrase the question differently enough, it might not have an opinion anymore. What sort of an opinion is that?

The operative word here, 'will', is in the future tense. Facile and dangerous AI agency may come soon, but agency alone does not eternally doom humanity to extinction. Again, agency within the information sphere is far different and much much more easily obtained than dominating agency in the physical world. Instead of focussing on illusory end of the world scenarios, the focus should shift to how AI is already being utilized to disrupt digital infrastructure. Much of the remediation needed is already well understood due to human black hat activities and threats.

"agency alone does not eternally doom humanity to extinction"

Can you prove this?

This is almost a textbook example of “reductionism ad absurdum”

Yeah. He seems to assume int and float are the only possible data types. But computers themselves have nothing to do with numbers in particular, they manipulate bits, and those can be used for far more than representing numbers.

People aren't grasping that uncontrolled compute already exists in the form of 20 year old computer viruses. They aren't benefiting anyone anymore, but they're still following their initial protocol, infecting other machines, causing harm for no reason, and we can't just "unplug them."

Reply on news.ycombinator.com