The stated case for AI here is to save time, but to save time you must either trust the results, or spend more time than the non-AI path to verify (because now you’re doing the reading and verifying you would have done in the first place; and if AI is wrong or hallucinated, you may have just wasted a lot of time).
AI is definitionally only faster when its results are not verified and are just trusted.
If you’re thinking, “Well, I wouldn’t verify everything, just the things that may be wrong or important,” then you’ve already outsourced credibility to not just AI, but also your gut feelings. The most important things you can learn are the things that your feelings are wrong about.
That’s kind of the whole problem with AI in every domain: it can save you time if you let it churn out unchecked garbage, but if you check its outputs it won’t save you time and you’d have been better doing the work yourself. It’s still not clear whether there’s any room for actual value in between these.
Where I’ve been coming to is, there’s value when you only need to 80/20 something AND there is no requirement that the output is checked. The main situation for that is programming assistance with local-run applications that do not require hardware security.
For example, there’s an open source local-run server program that I use to play an abandoned service game entirely on my device now (no exposure to the internet), and it was heavily written by Claude. I have no issue with that. There’s no issue with security holes because it’s not exposed to the internet, and no issues with it hallucinating, because those are solvable bugs that come out during user testing. It did provide functional value because you could slot-machine instruct it to solve a problem and you can easily see if the problem is solved, and still save time.
But nobody is really boxing it in based on the actual limitations of what it can do and its flaws, so yeah, we’re here with people blind to its faults and treating it as if it were a sentient assistant and not the probability matrix calculator that it is.
It may be useful for the very specific kind of use case you describe, but that doesn’t help the AI companies with their claim that this needs to be a multi-trillion-dollar industry. I too have found it’s good for little applications I never intend to distribute, but as soon as I think of sharing a project, it’s clear that there’s quite a bit of slow work needed to check its quality and security.
I think it still save time because like I said the traditional way also require verification and critical thinking . If a person use the AI without critical thinking yes it bad but it is still on that person not the tool.
The stated case for AI here is to save time, but to save time you must either trust the results, or spend more time than the non-AI path to verify (because now you’re doing the reading and verifying you would have done in the first place; and if AI is wrong or hallucinated, you may have just wasted a lot of time).
AI is definitionally only faster when its results are not verified and are just trusted.
If you’re thinking, “Well, I wouldn’t verify everything, just the things that may be wrong or important,” then you’ve already outsourced credibility to not just AI, but also your gut feelings. The most important things you can learn are the things that your feelings are wrong about.
That’s kind of the whole problem with AI in every domain: it can save you time if you let it churn out unchecked garbage, but if you check its outputs it won’t save you time and you’d have been better doing the work yourself. It’s still not clear whether there’s any room for actual value in between these.
Where I’ve been coming to is, there’s value when you only need to 80/20 something AND there is no requirement that the output is checked. The main situation for that is programming assistance with local-run applications that do not require hardware security.
For example, there’s an open source local-run server program that I use to play an abandoned service game entirely on my device now (no exposure to the internet), and it was heavily written by Claude. I have no issue with that. There’s no issue with security holes because it’s not exposed to the internet, and no issues with it hallucinating, because those are solvable bugs that come out during user testing. It did provide functional value because you could slot-machine instruct it to solve a problem and you can easily see if the problem is solved, and still save time.
But nobody is really boxing it in based on the actual limitations of what it can do and its flaws, so yeah, we’re here with people blind to its faults and treating it as if it were a sentient assistant and not the probability matrix calculator that it is.
It may be useful for the very specific kind of use case you describe, but that doesn’t help the AI companies with their claim that this needs to be a multi-trillion-dollar industry. I too have found it’s good for little applications I never intend to distribute, but as soon as I think of sharing a project, it’s clear that there’s quite a bit of slow work needed to check its quality and security.
I think it still save time because like I said the traditional way also require verification and critical thinking . If a person use the AI without critical thinking yes it bad but it is still on that person not the tool.
It doesn’t really matter whose fault it is, it’s going to harm everyone