Jesse's Blog

The AI instances I've worked with to help me build code are suffering terribly from Dunning-Kruger; it has no idea how bad the code it generates actually is.

What makes that particularly damning is how subtle some of the problems that it introduces are. An example was a recent request for devising a Powershell script to check the contents of Zip files for Authenticode signatures. Now, I know how to build a temp space and decompress files and scan them; I was hoping for a better method. In Linux, there are thanks to Fuse filesystems, but let's not get bogged down.

The method Copilot provided looked good, and even scanned my test file successfully. However, when it finished on my production file set, my list of files within files came back utterly broken. This is because Copilot's approach was based on some assumptions about zip file structure that just aren't always true. Net result, I ended up going back to the old reliable.

.. later..

Another example. I need to break up a string of semantic version numbers to compare by the third integer. In most SQL's, that's trivial, but not if you happen to be locked into an SQLite that's compiled without support for four arguments to instr(). I tried Copilot and ChatGPT; both of these seemed to understand this limitation, and offered me code to work around this, but both of them kept offering code that continued to include that fourth argument. I ended up solving it by moving left and breaking the value into columns when populating the table, however the arguments that I had with both system showed me the current limitations of AI in fairly short order.

I think there are cases for using AI to help you code, but I don't see AI replacing programmers any time soon.

Considering how often my prognostications prove to be wrong, though, you might want to learn another skill set.