Claude

[Borje Hagsten]

The problem with AI "programming", as I see it, is that while it can generate code quickly, you still have to read and test what it produces carefully and make sure the code will do exactly what you want in all possible cases. I can compare that to taking over a project from another programmer, where - if it's a big code - it can take days and weeks to follow and understand each step in code, then add a lot of time rewriting bad stuff and enhance code with better, faster and more secure ways to do this and that.

I once did this with PB Edit and PB Forms, when I started working for Bob back in 2004 and I can tell you - to dive into to MB's of code written and updated by several different programmers over the years took some time and effort. I can see similar or even worse problems, picking up on complicated code generated by an AI engine, where one knows there will be bugs and bad code hidden all over the place.

I guess it all boils done to - "I'm too old for this shi..."

[Stanley Durham]

That’s the best use case for AI.
I wrote an array for FB. I asked it to test all procedures with edge test, noting how many passed and failed. This code would be a lot of work if I had to write.
115 tests for 23 procedures.
You can tell it to do all kinds of tests because you don’t have to write the code.

---

I used AI to provide F256 floating point for FB. Just for kicks, I had it do an extreme test just now.
This is the prompt I gave it:
“test all procedures with all possible edge cases”
“at the bottom, note how many passed and failed”
208 tested, 208 passed.
Initial, it did find some rounding errors, “Note on the precision boundary: 78 significant digits (F256_NDIG) slightly exceeds what a 256-bit mantissa reliably holds (~77.06 digits), so values with non-terminating binary expansions (e.g. 1e-3, 1e-5) do not round-trip exactly as strings.”
Fixed it and tested it again.
3m, 45s; done deal.

[Dale Yarker]

Not likely. Just trying to propound the direction of our programing world.....

If you say so. But your tone looks more like advertising to me. Defending when someone pokes a hole in your premise.

[Walt Decker]

I used AI to provide F256 floating point for FB. Just for kicks, I had it do an extreme test just now.
This is the prompt I gave it:
“test all procedures with all possible edge cases”
“at the bottom, note how many passed and failed”
208 tested, 208 passed.

The premise that 208 was tested and 208 passed is faulty. There is no indication of what tests were used. Were the test them selves faulty? That is unknown. "AI" is known to produce faulse-hoods.

Scientifically speaking, one can claim that something is correct only if it can be varifed via duplication of results by more than one party. That was not done. Therefore, any results are automatically suspicious.

[Stanley Durham]

For F256, I can look at the test code. Since you don’t have to write the code, you can build extreme test apps.
I used DeepSeek to write it. I double checked the output with Google AI.
> Therefore, any results are automatically suspicious.
That’s the most important thing to understand. But the same can be said for your code.
So, you use it to test your and its code. For me, test code is extremely boring and tedious. A good test app may require more code than the source.
If you already understand all the implications, you don’t need AI.
Let’s say you want to do something you’re not that familiar with. I don’t have a clue on 3D graphics. With one prompt I had DeepSeek write a 3D test app in C. Moving forward with that would require a lot of work and testing, but that’s too powerful to ignore.

I was curious as to how one could write a BASIC compiler that translates to C, that supports dynamic strings, wide strings and arrays; and automatically frees them when they go out of scope. As it turns out, with GCC or Clang, not MS, you can do that with “`__attribute__((cleanup))”.
DeepSeek asked if I wanted a sample, so why not.
I now have a BASIC compiler that supports dynamic strings and arrays in a UDT, that gets freed when it goes out of scope. And it actually compiles code. I’m not going to try and write a full compiler, too many details.

It’s worth spending a little time playing around with. You don’t need it for what you already know, you need it for what you don’t know how to do or is boring and tedious. Too powerful to ignore.

[Graham McPhee]

As any code in a public repository on GitHub is used to train AI systems, and I've been pushing PowerBasic code to such a repository for years now, it's likely I'm partly to blame for Claude getting better, at least with PowerBasic.

It's not perfect by any means but it is improving. The trust that can be put in any piece of AI generated code should be similar to that placed with any human junior developer.
You have to test and review their code. It's also true that if you build  not just an application but a system, you'll spend longer supporting and maintaining that system, that you ever did writing it in the first place.

User requirements evolve over time and you will need to amend and upgrade systems as the years pass. AI is not currently good at this activity. Users are also not good at specifying requirements either. Perhaps we should use AI to replace the users before we replace the programmers