4 ms·

The assumption of the article is that future AI will be more intelligent or at least equivalent to humans. Although I could perhaps see a possibility of equival

The assumption of the article is that future AI will be more intelligent or at least equivalent to humans. Although I could perhaps see a possibility of equivalence I seriously doubt we can make AI that is more intelligent than us. Why? Well, what is it that we use to train AI? Human generated content. What is the most revolutionary AI technique that came out recently? Direct human feedback during the training process. Do you see a pattern here? If we attempt to feed AI generated content to train other AIs all performance metrics drop... Its a chicken and egg situation if a human mind is the best content creation device we have, how anything trained on that content can surpass it quality wise? Of course AI having very fast access memory compared to our meatware can recall more and faster, but this is _just_ a quantitative improvement. Not a qualititative one.

cubefox 3y ago
Future AI doesn't have to be trained on text, it doesn't need to be a language model. It can be trained on video data to predict its sensory inputs. It would predict the future of reality, not the next text tokens. Gains from training on text are bound by the human intelligence which produced this text. But training on sensory data is only bound by reality itself.
MacsHeadroom 3y ago
This is demonstrably incorrect on every level. Reinforcement learning from human feedback (RLHF) makes models dumber and less capable but more compliant. Meanwhile, high quality synthetic training data (generated by AI) makes models smarter and more capable. RLHF and human sourced training data is not the ceiling of training intelligent systems, it's the baseline-- the floor.
tim333 3y ago
Chess, Go etc are counter examples.
js8 3y ago
That's just a repackaging of the famous "evolution cannot create complexity" argument against evolution. There are good rebuttals, one is intuition from Dan Dennett that we can build buildings that are taller than the highest crane.
bottlepalm 3y ago
If I was chatting with you in one window and ChatGPT in another, please tell me how you would prove that you are the ‘more intelligent’ one.
- Mordisquitos 3y ago
  Intelligence is hard enough to define already, but consider that human intelligence can only be evaluated inasmuch as it allows its holder to successfully satisfy his or her own interests and purposes in the current environment and circumstances. That implies of course having interests and purposeful objectives. Are you+ChatGPT more intelligent than I am by myself? Almost certainly. You have extended your abilities to fulfill your purposes, and can demonstrably do so better than I can. However, is ChatGPT itself more intelligent than I am? Well, what are ChatGPT's purposes and objectives? What is its environment? What does ChatGPT want, so we can measure its ability to achieve it?
  - bottlepalm 3y ago
    Three paragraphs desperately trying to rationalize the fact that you wouldn’t be able to prove that you are the more intelligent one in the scenario I presented. Time to face the truth.
    - Mordisquitos 3y ago
      You seem to be under the misapprehension that I have any motivation to enter a dick-measuring competition with ChatGPT, let alone one that would not measure intelligence but instead would be biased towards the outcome you want to prove. How do you understand intelligence? How do you judge it? And why should it matter to me?
cma 3y ago
> Well, what is it that we use to train AI? Human generated content. Systems like alpha zero learn through self play.
victroflux 3y ago
[dead]
hinkley 3y ago
For an AI to wreak havoc it doesn't have to be smart. It just has to be capable. Industrious but stupid is the theme of half of speculative fiction on intelligent machines. Grey goo and paperclips are examples. Tangentially, I love that Murderbot thinks that killing all humans would be stupid, because who would make TV shows for it to watch if the humans are dead?
starburst 3y ago
We are already making humans more intelligent each generations, surely we can make machine more intelligent than us especially since they are not limited the same way we are.
hollowpython 3y ago
I agree that the feedback loop wouldn't make the models smarter than us, but training data would. I can see how with more computing power we will be able to feed it the whole of YouTube, Spotify and perhaps also millions of physics experiments. With this data, perhaps a model will be able to understand the hidden patterns of the world (e.g. physics equations governing reality) better than us, don't you think?

cubefox 3y ago

Future AI doesn't have to be trained on text, it doesn't need to be a language model. It can be trained on video data to predict its sensory inputs. It would predict the future of reality, not the next text tokens. Gains from training on text are bound by the human intelligence which produced this text. But training on sensory data is only bound by reality itself.

MacsHeadroom 3y ago

This is demonstrably incorrect on every level. Reinforcement learning from human feedback (RLHF) makes models dumber and less capable but more compliant. Meanwhile, high quality synthetic training data (generated by AI) makes models smarter and more capable. RLHF and human sourced training data is not the ceiling of training intelligent systems, it's the baseline-- the floor.

tim333 3y ago

Chess, Go etc are counter examples.

js8 3y ago

That's just a repackaging of the famous "evolution cannot create complexity" argument against evolution. There are good rebuttals, one is intuition from Dan Dennett that we can build buildings that are taller than the highest crane.

bottlepalm 3y ago

If I was chatting with you in one window and ChatGPT in another, please tell me how you would prove that you are the ‘more intelligent’ one.

Mordisquitos 3y ago
Intelligence is hard enough to define already, but consider that human intelligence can only be evaluated inasmuch as it allows its holder to successfully satisfy his or her own interests and purposes in the current environment and circumstances. That implies of course having interests and purposeful objectives. Are you+ChatGPT more intelligent than I am by myself? Almost certainly. You have extended your abilities to fulfill your purposes, and can demonstrably do so better than I can. However, is ChatGPT itself more intelligent than I am? Well, what are ChatGPT's purposes and objectives? What is its environment? What does ChatGPT want, so we can measure its ability to achieve it?
- bottlepalm 3y ago
  Three paragraphs desperately trying to rationalize the fact that you wouldn’t be able to prove that you are the more intelligent one in the scenario I presented. Time to face the truth.
  - Mordisquitos 3y ago
    You seem to be under the misapprehension that I have any motivation to enter a dick-measuring competition with ChatGPT, let alone one that would not measure intelligence but instead would be biased towards the outcome you want to prove. How do you understand intelligence? How do you judge it? And why should it matter to me?

cma 3y ago

> Well, what is it that we use to train AI? Human generated content. Systems like alpha zero learn through self play.

victroflux 3y ago

[dead]

hinkley 3y ago

For an AI to wreak havoc it doesn't have to be smart. It just has to be capable. Industrious but stupid is the theme of half of speculative fiction on intelligent machines. Grey goo and paperclips are examples. Tangentially, I love that Murderbot thinks that killing all humans would be stupid, because who would make TV shows for it to watch if the humans are dead?

starburst 3y ago

We are already making humans more intelligent each generations, surely we can make machine more intelligent than us especially since they are not limited the same way we are.

hollowpython 3y ago

I agree that the feedback loop wouldn't make the models smarter than us, but training data would. I can see how with more computing power we will be able to feed it the whole of YouTube, Spotify and perhaps also millions of physics experiments. With this data, perhaps a model will be able to understand the hidden patterns of the world (e.g. physics equations governing reality) better than us, don't you think?