Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Licensing is neither feasible nor effective for addressing AI risks

aisnakeoil.substack.com

Post reply on HN
Print view View original

71–80 of 161 posts

Re: Licensing is neither feasible nor effective for addressing AI risks

#71

Post by circuit10 »

adsfgiodsnrio wrote: post #60
circuit10 wrote: post #36

Earlier quoted context omitted.

It’s almost impossible to supervise something more intelligent than you because you can’t tell why it’s doing things. For now it’s easy to supervise them because AIs are way less intelligent than humans (though even now it’s hard to tell exactly why they’re doing things), but in the future it probably won’t be

"Supervised" does not mean the models need babysitting; it refers to the fundamental way the systems learn. Our most successful machine learning models all require some answers to be provided to them in order to infer the rules. Without being given explicit feedback they can't learn anything at all. Humans also do best with supervised learning. This is why we have schools. But humans are capable of unsupervised learn…

Oh, sorry, I’m not that familiar with the terminology (I still feel like my argument is valid despite me not being an expert though because I heard all this from people who know a lot more than me about it). One problem with that kind of feedback is that it incentives the AI to make us think it solved the problem when it didn’t, for example by hallucinating convincing information. That means it specifically learns how to lie to us so it doesn’t really help

Also I guess giving feedback is sort of like babysitting, but I did interpret it the wrong way

Re: Licensing is neither feasible nor effective for addressing AI risks

#72

Post by Al0neStar »

circuit10 wrote: post #35
civilitty wrote: post #26

Earlier quoted context omitted.

Yes, really. Scifi fantasies don’t count as evidence and we’ve learned since the Renaissance and scientific revolution that all this Platonic theorizing is just an intellectual circle jerk.

The fact that something has been covered in sci-fi doesn’t mean that it can’t happen. https://en.m.wikipedia.org/wiki/Appeal_to_the_stone “Speaker A: Infectious diseases are caused by tiny organisms that are not visible to unaided eyesight. Speaker B: Your statement is false. Speaker A: Why do you think that it is false? Speaker B: It sounds like nonsense. Speaker B denies Speaker A's claim without providing evidence…

The burden of proof is on the person making the initial claim and i highly doubt that the reasoning behind AI ruin is as extensive and compelling as the germ theory of disease.

We assume human-level general AI is possible because we exist in nature but a super-human self-optimizing AI god is nowhere to be found.

Re: Licensing is neither feasible nor effective for addressing AI risks

#73

Post by wizzwizz4 »

circuit10 wrote: post #71
adsfgiodsnrio wrote: post #60

Earlier quoted context omitted.

"Supervised" does not mean the models need babysitting; it refers to the fundamental way the systems learn. Our most successful machine learning models all require some answers to be provided to them in order to infer the rules. Without being given explicit feedback they can't learn anything at all. Humans also do best with supervised learning. This is why we have schools. But humans are capable of unsupervised learn…

Oh, sorry, I’m not that familiar with the terminology (I still feel like my argument is valid despite me not being an expert though because I heard all this from people who know a lot more than me about it). One problem with that kind of feedback is that it incentives the AI to make us think it solved the problem when it didn’t, for example by hallucinating convincing information. That means it specifically learns ho…

> One problem with that kind of feedback is that it incentives the AI to make us think it solved the problem when it didn’t,

Supervised learning is: "here's the task" … "here's the expected solution" *adjusts model parameters to bring it closer to the expected solution*.

What you're describing is specification hacking, which only occurs in a different kind of AI system: https://vkrakovna.wordpress.com/2018/04/02/specification-gam... In theory, it could occur with feedback-based fine-tuning, but I doubt it'd result in anything impressive happening.

Re: Licensing is neither feasible nor effective for addressing AI risks

#74

Post by wizzwizz4 »

circuit10 wrote: post #67
AuthorizedCust wrote: post #62

Earlier quoted context omitted.

> We can’t even explain how current LLMs work internally. You sure can. They are just not simple explanations yet. But that’s the common course of inventions, which in foresight are mind-bogglingly complex, in hindsight pretty straightforward.

You can explain the high level concepts but it’s really difficult to say “this group of neurons does this specific thing and that’s why this output was produced”, though OpenAI did make some progress in getting GPT-4 to explain what each neuron in GPT-2 is correlated to but we can also find what human brain regions are correlated to but that doesn’t necessarily explain the system as a whole and how everything interac…

> but it’s really difficult to say “this group of neurons does this specific thing and that’s why this output was produced”,

That's because that's not how brains work.

> though OpenAI did make some progress in getting GPT-4 to explain what each neuron in GPT-2 is correlated to

The work contained novel-to-me, somewhat impressive accomplishments, but this presentation of it was pure hype. They could have done the same thing without GPT-4 involved at all (and, in fact, they basically did… then they plugged it into GPT-4 to get a less-accurate-but-Englishy output instead).

Re: Licensing is neither feasible nor effective for addressing AI risks

#75

Post by davidzweig »

dwallin wrote: post #20

Any laws should be legislating the downstream effects of AI, not the models themselves. Otherwise we will quickly get to a place where we have a handful of government-sanctioned and white-washed “safe” models responsible for deleterious effects on society, with plausible deniability for the companies abusing them. Legislating around the model is missing the point. There is no evidence that a runaway artificial intell…

>> There is no evidence that a runaway artificial intelligence is even possible.

In the space of a century or so, the humans have managed to take rocks and sand and turn them into something that you can talk to with your voice, and it understands and responds fairly convincingly as it were a well-read human (glue together chatGPT with TTS/ASR).

Doesn't seem like a big stretch to imagine that superhuman AI is just a few good ideas away, a decade or two perhaps.

Re: Licensing is neither feasible nor effective for addressing AI risks

#76

Post by circuit10 »

wizzwizz4 wrote: post #73
circuit10 wrote: post #71

Earlier quoted context omitted.

Oh, sorry, I’m not that familiar with the terminology (I still feel like my argument is valid despite me not being an expert though because I heard all this from people who know a lot more than me about it). One problem with that kind of feedback is that it incentives the AI to make us think it solved the problem when it didn’t, for example by hallucinating convincing information. That means it specifically learns ho…

> One problem with that kind of feedback is that it incentives the AI to make us think it solved the problem when it didn’t, Supervised learning is: "here's the task" … "here's the expected solution" *adjusts model parameters to bring it closer to the expected solution*. What you're describing is specification hacking , which only occurs in a different kind of AI system: https://vkrakovna.wordpress.com/2018/04/02/spe…

Oh, that seems less problematic (though not completely free of problems), but also less powerful because it can’t really exceed human performance

Re: Licensing is neither feasible nor effective for addressing AI risks

#77

Post by circuit10 »

wizzwizz4 wrote: post #74
circuit10 wrote: post #67

Earlier quoted context omitted.

You can explain the high level concepts but it’s really difficult to say “this group of neurons does this specific thing and that’s why this output was produced”, though OpenAI did make some progress in getting GPT-4 to explain what each neuron in GPT-2 is correlated to but we can also find what human brain regions are correlated to but that doesn’t necessarily explain the system as a whole and how everything interac…

> but it’s really difficult to say “this group of neurons does this specific thing and that’s why this output was produced”, That's because that's not how brains work. > though OpenAI did make some progress in getting GPT-4 to explain what each neuron in GPT-2 is correlated to The work contained novel-to-me, somewhat impressive accomplishments, but this presentation of it was pure hype. They could have done the same…

When I said about a group of neurons I was talking about LLMs, but some of the same ideas probably apply. Yes, it’s probably not as simple as that, and that’s why we can’t understand them.

I think they just used GPT-4 to help automate it on a large scale, which could be important to help understand the whole system especially for larger models

Re: Licensing is neither feasible nor effective for addressing AI risks

#78

Post by milsorgen »

jahewson wrote: post #46
circuit10 wrote: post #36

Earlier quoted context omitted.

It’s almost impossible to supervise something more intelligent than you because you can’t tell why it’s doing things. For now it’s easy to supervise them because AIs are way less intelligent than humans (though even now it’s hard to tell exactly why they’re doing things), but in the future it probably won’t be

The government seems to manage the task just fine every day.

Does it?

Re: Licensing is neither feasible nor effective for addressing AI risks

#79

Post by nradov »

arisAlexis wrote: post #70
dwallin wrote: post #47

Earlier quoted context omitted.

If that’s your concern than lets direct these government resources into research to improve our shared knowledge about them. If humans only ever did things we fully understood, we would have never left the caves. Complete understanding is impossible so the idea of establishing that as the litmus test is a fallacy. We can debate what the current evidence shows, and even disagree about it, but to act as if only one par…

You know, when nuclear bombs were made and Einstein and Oppenheimer knew about the dangers etc, there were common people like you that dismissed it all. This has been going on for centuries. Inventors and experts and scientists and geniuses say A and common people say nah, B. Well, Bengio, Hinton, Ilya and 350 others from the top AI labs disagree with you. Does it ever make you wonder if you should be so cock sure or…

[dead]

Re: Licensing is neither feasible nor effective for addressing AI risks

#80

Post by killjoywashere »

m4rtink wrote: post #69
killjoywashere wrote: post #63

The only way to address generative AI is to strongly authenticate human content. Camera manufacturers, audio encoders, etc, should hold subordinate CAs and issue signing certs to every device. Every person should have keys, issued by the current CA system, not the government (or at least not necessarily). You should have a ability, as part of the native UX, to cross-sign the device certificate. Every file is then sig…

Seems like something that could be very easily missused for censorship, catching whistleblowers and similar.

That's actually why it's urgent to set this up outside of government, like the browser CA system, and develop methods to issue verification of "human" while preserving other aspects of anonymity.
Post reply on HN