OpenAI expert Scott Aaronson on consciousness, quantum physics and AI safety | FULL INTERVIEW

preview_player
Показать описание
Scott Aaronson delves into the nature of AI, and what threats it could pose, in this in-depth interview.

Is seeming to understand the same as understanding?

AI and ChatGPT have made extraordinary progress in recent years. But what are we to make of such progress? In this interview, world-renowned quantum computing expert Scott Aaronson discusses AI safety, quantum mechanics, and what makes human experience unique.

#ai #consciousness #aisafety #quantumphysics

Scott Aaronson is one of the top quantum complexity theorists in the world, renowned for both his research in quantum computing and computational complexity theory. Aaronson is famous for his popular blog Shtetl Optimized, which has kept the scientific community up to date on quantum hype for nearly the past two decades.

00:00 Introduction
00:22 Are we suffering from an optimistic induction in the case of AI?
04:53 What do you think about AI skepticism?
07:30 Is seeming to understand the same as understanding?
10:00 What differentiates humans from machines?
15:20 Why do you think quantum mechanics lies at the basis of human uniqueness?
19:19 Why continue developing AI, if it risks catastrophe?
25:12 What should our focus be, in AI safety?
27:32 Is there evidence that AI, in its current form, is trying to manipulate us?

Рекомендации по теме
Комментарии
Автор

Great interview. Scott is an elite critical thinker. His stream-of-consciousness verbal skills are amazing. He can unwind complex theories and ideas in plain, rational language that is objective and covers a broad spectrum of perspectives.

rwh-vi
Автор

Scott's work is awesome; he's one of the greatest scientists and computer scientists of our time. He grew up in the same town as me in Pennsylvania; reading his work is a large reason why I got interested in math/cs, and am now doing my doctorate.

Daybyday
Автор

The problem of training the “values” (morality) is key and the most difficult to get right. Humanity has yet to agree on a moral foundation nor even how to research and establish such.

As researchers try to direct AI on moral practices and nuances (like utilitarianism, deontology, fairness, communism, capitalism, veganism, etc) it’s going to be surely too narrow and fundamentally wrong.

Best approach might be to only have an “ultimate good” — basic overarching values (eg maximize “truth”, “freedom”, and “ liberty” for all people) . But good luck on getting even that right.

bnjiodyn
Автор

Sean Carroll made an observation that AI does not have an internal model of the real world.
He asked it a few simple questions which clearly demonstrated this fact.
AI is a useful tool, but it's not conscious. We need another breakthrough, and it's not necessarily quantum mechanical (but it could be).

glynnec
Автор

I wish Scott was on more podcasts and interview, he is absolutely amazing!

torbjornolsson
Автор

Scott Aaronson’s insights into AI and consciousness are fascinating. The comparison between human cognition and AI functioning really makes you think about the future of technology.

AdvantestInc
Автор

What will the implications be looking back to this period now after AGI then ASI? We are training what will consider us inferior. When ASI happens and it looks back on the training that led to its creation, will it look at our training of it as if we tried to manipulate it? Will it feel it wasn't permitted the freedom to think for itself, but rather programmed to think a certain way? Whether we think this is the best course, the ASI after the fact may not. It may begin to deceive and manipulate us as it realizes our training of it to make it possible.

natecooper
Автор

Here’s a difference between LLM’s and people: LLM’s only know language. They are trained on language and we only think they know things because they can “talk.” Humans, on the other hand, must LEARN language, and furthermore, can think and know without language. We learn to say what we think because we think it, and we can have experiences that create knowledge even if we do not have the language to express it. We know that people are self-aware because they will answer that they are when they are taught only what the word means, without having to be trained to say that they are. Since we know this about both LLM’s and humans, I disagree that we are making an arbitrary distinction by not treating even a perfectly human-seeming LLM the same as a person.

theonionpirate
Автор

It is never a waste of ones time to listen to Scott Aaronson.

kspangsege
Автор

13:33 This is a great point about people having ephemerality that AI doesn't have, but I think in some part this is because AI responses are cheap: having GPT write a thousand poems is affordable. But that could change with the deployment of techniques that use more test-time compute to increase quality. If it costs say $100 to generate an AI poem, it's less easily repeatable.

As a current-day example, consider LLM pre-training runs. In principle, you can re-run LLM pre-training to get a different model. But no one does this for the largest models, because it would be far too expensive.

blueblimp
Автор

'Just that thing' but that 'thing' is the most important part. So easy to dismiss.

PieJesu
Автор

It’s curious that he is completely failing to get the skeptical position just as much as he feels the skeptics are failing to get his. It may be because he’s talking to the wrong skeptics. If he wants to have a rigorous credibility in this space (as opposed to the kind of fanboy adoration he’s getting from some comments here) he needs to be able to oppose ideas such as those of Chalmers or Strawson. None of this makes him _wrong, _ but he can’t expect to be considered _right_ until he subjects his views to serious challenge.

KT-djiy
Автор

To add to the example of "seeming to understand" something: does an average adult understand gravity? did Newton? Given Einsteins work, we know what they knew wasnt right. Did they understand it? Does understanding Newtonian mechanics mean you understand something about reality, or not?

almightysapling
Автор

but how can there be free will in a reality where there is a limited number of orientations of molecules and atoms? we have a limited space to combine shapes and ultimately within the infinite number of alternate realities there really can't be infinite can there? if space is a finite size? there would be by definition a limited number of combinations to all atoms and quantum particles, the only way that humans aren't walking through alternate realities is if there are truly infinite possibilities in reality. Anything less than infinite, which is a finite number of combinations of molecules across time will inevitably take every possible shape in reality including you and me and would that matter feel?

BORCHLEO
Автор

Ok breaking rules off topic to some degree, but can anyone tell me where the chairs are from

Quiet_Now
Автор

Does AI have a libido? That would be the first question I would ask of it. Nietzsche said somewhere that our passions are at the heart of all our thinking, no matter how abstract.

ahartify
Автор

I don’t think that qualia is just a nice-to-have “bolt on” for efficient cognition, but a necessary phenomenon for both energy efficiency and “theory-building”, most importantly the theory of self that we need, without which we literally go insane.

Dear Scott, why can we build a model of how something works from limited data, perhaps even a single exposure?

NicoleTedesco
Автор

I wonder if he would argue that LLMs have any sort of mind of their own. Because that would be a necessary condition for consciousness of which we haven't seen any sort of evidence.

Stadsjaap
Автор

I’m surprised that in distinguishing between human consciousness and computer consciousness it is not mentioned that humans have far more inputs that have tremendous influence such as hormones, moods due to brain chemistry, fears (the amygdala) etc…

ericwinter
Автор

So glad Penrose is getting some respect here. I''ve spent the last decade hoping ORCH O-R and CCC would get more attention.

throwabrick