Amusing AI Example in Help

[Jay Sage]

I have been entertaining myself (and learning a lot) by reading through the online help. I found the following example in the help for the AI command. It illustrates what is so often true of AI: what it tells you is wrong!

[D:\TakeCommand36] ai /role=describe "display all files between 20 and 25 days old"​

The TCC.EXE shell command to display files between 20 and 25 days old is: DIR /[d-25,-20].​

​

- DIR: Lists directory contents.​

- /[d-25,-20]: Filters files based on their age, showing only those modified between 25 and 20 days ago.​

- d: Represents the date filter.​

- -25 and -20: Specify the range of days from the current date.​

That looked quite plausible, but I had not long before read through the help section on date ranges and learned that offsets for the second date are not relative to today but relative to the first date. So "/d[-25,-20]" means with dates between 25 and 45 days ago. Indeed, when I used "/d[-45,+20]", I got the same results.

I just ran that AI command line, and the wrong answer is still given. I also tried using ChatGPT directly on the web, and I got the same wrong answer. There I was able to tell it that it made a mistake and give it the correct syntax. As it typically does, it thanked me for the correction: "You are correct — thank you for the correction."

Perplexity made the same mistake and responded similarly to my correction. Gemini made the same mistake and was even more effusive in its thanks: "You are completely right, and I appreciate you catching that... Thank you for correcting that detail—it makes a huge difference in getting the actual files you're looking for!"

I'm starting to think that that response is just automatic. It's not clear whether the AI is actually the correction or learning from it. With ChatGPT, my first correction actually had a typo, but it responded by affirming my mistaken correction. When I apologized and gave it the correction without the typo, I got the same thank-you message.

 

[David Marcus]

It's not clear whether the AI is actually the correction or learning from it.

Of course, the AI is not learning. AIs are LLMs, they are not "I" (i.e., not intelligent). They are excellent at mimicking human language. But, they don't understand anything.

 

[Charles Dye]

Of course, the AI is not learning. AIs are LLMs, they are not "I" (i.e., not intelligent). They are excellent at mimicking human language. But, they don't understand anything.

Next you'll be telling me that Eliza doesn't really understand my personal problems!

 

[Jay Sage]

I'm not so sure about the AI not learning. I have read warnings that, especially with free versions of the AI apps, that one's conversations can be used to augment the data in the models. Hence the warning to be very careful about including personal information in queries. In any case, I figure that it cannot do any harm to point out the mistakes.