24 ms·

Here's my falsifiable predictions: 1. We won't be able to evolve these systems such that they become 100% accurate. 2. Despite this, because they are so conve

Here's my falsifiable predictions: 1. We won't be able to evolve these systems such that they become 100% accurate. 2. Despite this, because they are so convenient, we will lower our standards to accept some falsehoods as acceptable in areas where we previously did not. 3. Real human expertise will become a 'premium product' across multiple industries.

lionkor 3y ago
4. You will not be able to know that an answer you get to a question you pose, however complex, was not word-for-word in the training set
omginternets 3y ago
Agreed. I also think point 4 has an analogy in domains like art/marketing. As humans become better at recognizing the idiosyncrasies of AI-generated content, it will become ghettoized. I'm expecting something like a revival of organic, human-produced content (with a premium cost, of course).
revertmean 3y ago
1. There is no such thing as 100% accurate. Not only is it not physically possible (there can always be hardware errors or bit flips) but it's not even theoretically possible (you'd require a checker that was 100% accurate to tell, which is equivalent to solving the halting problem). 2. We already have, since even these early days models are in current use. 3. The assumption here is that human expertise will always be more accurate than model expertise, which seems unlikely. I wouldn't be surprised if someone - even just for fun - tries to set up a software company with a traditional management/developer structure, but where AI plays all the roles. It sounds like an interesting experiment.
- Tainnor 3y ago
  > 1. There is no such thing as 100% accurate. Not only is it not physically possible (there can always be hardware errors or bit flips) but it's not even theoretically possible (you'd require a checker that was 100% accurate to tell, which is equivalent to solving the halting problem). You don't have to solve the halting problem to prove a mathematical theorem (which includes proving things about a computer program), either manually or via an automated theorem prover. One consequence of the halting problem (or more precisely, Rice's theorem) is that there is no algorithm that can determine a non-trivial property of an arbitrary program. It doesn't imply that you can't prove things about a specific program. I suppose you can always be philosophical about it and say "how do I know the axioms are true" (whatever that means), or "how do I know there's no mistake in this proof" - but then you'd have to extend that same level of scrutiny to the theorem that the halting problem can't be solved, I guess.

lionkor 3y ago

4. You will not be able to know that an answer you get to a question you pose, however complex, was not word-for-word in the training set

omginternets 3y ago

Agreed. I also think point 4 has an analogy in domains like art/marketing. As humans become better at recognizing the idiosyncrasies of AI-generated content, it will become ghettoized. I'm expecting something like a revival of organic, human-produced content (with a premium cost, of course).

revertmean 3y ago

1. There is no such thing as 100% accurate. Not only is it not physically possible (there can always be hardware errors or bit flips) but it's not even theoretically possible (you'd require a checker that was 100% accurate to tell, which is equivalent to solving the halting problem). 2. We already have, since even these early days models are in current use. 3. The assumption here is that human expertise will always be more accurate than model expertise, which seems unlikely. I wouldn't be surprised if someone - even just for fun - tries to set up a software company with a traditional management/developer structure, but where AI plays all the roles. It sounds like an interesting experiment.

Tainnor 3y ago
> 1. There is no such thing as 100% accurate. Not only is it not physically possible (there can always be hardware errors or bit flips) but it's not even theoretically possible (you'd require a checker that was 100% accurate to tell, which is equivalent to solving the halting problem). You don't have to solve the halting problem to prove a mathematical theorem (which includes proving things about a computer program), either manually or via an automated theorem prover. One consequence of the halting problem (or more precisely, Rice's theorem) is that there is no algorithm that can determine a non-trivial property of an arbitrary program. It doesn't imply that you can't prove things about a specific program. I suppose you can always be philosophical about it and say "how do I know the axioms are true" (whatever that means), or "how do I know there's no mistake in this proof" - but then you'd have to extend that same level of scrutiny to the theorem that the halting problem can't be solved, I guess.