If your code doesn’t work because you rely on humans understanding it, you don’t have a business you can run. We already are there where humans have no idea why the computer does this or that decision because it’s so complex especially with all the machine learning and complex training data, etc. let’s not pretend it will get less complex with time.
So your argument is that people will rely on AI entirely without making any redundancies, unlike now where they have more than one human so they can check for these issues because humans make coding errors?
My argument is that already today no human is able to and checks it when it comes to decision making models like for example if the car should go left or right around a obstacle. And over time we will have less straight forward classical programming doing decisions and more and more models doing decisions with hundreds or thousands of sensor inputs.
So you’re saying no code would be worth error checking by a human at all? There is no level of simpler code that an AI could get wrong and would need someone to fix it?
My assumption is that it would even be the other way around, AI checking all human code, especially writing all the tests. So I guess AI would also write tests for it’s own code too.
When it comes to humans, I think they would probably be changing the prompt they give to AI instead of changing the code at the end. You can already see how it’s done with generation of pictures, while theoretically you could take a not so perfect AI picture and use Photoshop to fix it, but most of the time people change the prompt instead and regenerate the picture.
I’m not sure why you think it’s not a good example. The picture itself is code (PNG, JPEG, etc.) which some AI wrote and no human at the company has checked but it gets delivered to the customer who pays for it. It’s not as good as if a human would do it, but it’s way way way cheaper so you can generate a couple of times until you get what you want. So in this domain humans already are losing jobs to AI even though AI is so bad that it is giving people extra legs, etc.
The same thing happens with hallucinations of ChatGTP, which despite that is still preferable by customers to a human assistant who would summarize articles, etc. with much better quality but very low speed and very expensive.
Code is not much different from summaries of longer texts.
Or let me rephrase it with the context of the original assumption that if people don’t check the code which AI wrote the company will lose customers because the quality is bad.
Right now there are tons of models out there which no human can understand why they decide this or that, still they bring so much more value that they get shipped even though they make some mistakes. If a company would try to only ship code checked by humans they would not be able to ship the products and would lose their customers to a company which does not check it but does ship.
Except we already have fields (like pharma manufacturing) that have to deal with hundreds or thousands of inputs and variables, are automated, and we still manage to fully understand the stack as well as fully check everything.
Hint: when someone tells you they “can’t” check or understand what their software is doing, it’s a scam.
Normally they should be told to go back and figure it out before being allowed to ship any product. If you tried this in any other industry it would be laughable. Even in software it’s outrageous, imagine getting accounting software or even a simple file backup tool that doesn’t work some of the time and nobody can tell you how it works. Yet these companies get a pass putting cars like this on the road.
I kinda agree with them. Currently coding already is an abstraction. The average developer has very little idea what machine code their compiler actually produces, and for the most part they don’t need to care about this. Feeding an AI a specification is just a higher level of abstraction.
For now, we’ll need people to check that AI produces code that does what we expect, but I believe at some point we’ll mostly take it for granted that they just do.
If your code doesn’t work because you rely on humans understanding it, you don’t have a business you can run. We already are there where humans have no idea why the computer does this or that decision because it’s so complex especially with all the machine learning and complex training data, etc. let’s not pretend it will get less complex with time.
So your argument is that people will rely on AI entirely without making any redundancies, unlike now where they have more than one human so they can check for these issues because humans make coding errors?
My argument is that already today no human is able to and checks it when it comes to decision making models like for example if the car should go left or right around a obstacle. And over time we will have less straight forward classical programming doing decisions and more and more models doing decisions with hundreds or thousands of sensor inputs.
And that means AI code shouldn’t be error-checked?
That means that it right now can not be error checked and it will be even more difficult in the future.
So you’re saying no code would be worth error checking by a human at all? There is no level of simpler code that an AI could get wrong and would need someone to fix it?
My assumption is that it would even be the other way around, AI checking all human code, especially writing all the tests. So I guess AI would also write tests for it’s own code too.
When it comes to humans, I think they would probably be changing the prompt they give to AI instead of changing the code at the end. You can already see how it’s done with generation of pictures, while theoretically you could take a not so perfect AI picture and use Photoshop to fix it, but most of the time people change the prompt instead and regenerate the picture.
The generation of pictures is full of fuckups like giving people extra legs, so I’m not sure that’s a very good example.
I’m not sure why you think it’s not a good example. The picture itself is code (PNG, JPEG, etc.) which some AI wrote and no human at the company has checked but it gets delivered to the customer who pays for it. It’s not as good as if a human would do it, but it’s way way way cheaper so you can generate a couple of times until you get what you want. So in this domain humans already are losing jobs to AI even though AI is so bad that it is giving people extra legs, etc.
The same thing happens with hallucinations of ChatGTP, which despite that is still preferable by customers to a human assistant who would summarize articles, etc. with much better quality but very low speed and very expensive.
Code is not much different from summaries of longer texts.
Until even the tests are bugged
Have you seen tests written by humans? 😅
Or let me rephrase it with the context of the original assumption that if people don’t check the code which AI wrote the company will lose customers because the quality is bad.
Right now there are tons of models out there which no human can understand why they decide this or that, still they bring so much more value that they get shipped even though they make some mistakes. If a company would try to only ship code checked by humans they would not be able to ship the products and would lose their customers to a company which does not check it but does ship.
Except we already have fields (like pharma manufacturing) that have to deal with hundreds or thousands of inputs and variables, are automated, and we still manage to fully understand the stack as well as fully check everything.
Hint: when someone tells you they “can’t” check or understand what their software is doing, it’s a scam.
Normally they should be told to go back and figure it out before being allowed to ship any product. If you tried this in any other industry it would be laughable. Even in software it’s outrageous, imagine getting accounting software or even a simple file backup tool that doesn’t work some of the time and nobody can tell you how it works. Yet these companies get a pass putting cars like this on the road.
I kinda agree with them. Currently coding already is an abstraction. The average developer has very little idea what machine code their compiler actually produces, and for the most part they don’t need to care about this. Feeding an AI a specification is just a higher level of abstraction.
For now, we’ll need people to check that AI produces code that does what we expect, but I believe at some point we’ll mostly take it for granted that they just do.