Tag: AI

One of the things I’ve been experimenting with in AI is taking things other people do and seeing how well the AI works. In this case, I took a Prompt from Louis and tried to see how well it worked. This post describes the results from a few models.

This is part of a series of experiments with AI systems.

The Challenge Using Claude

Louis was using an AI to generate some test data to play with the new Regex functionality in SQL Server 2025. He put this in ChatGPT: “create a set of random string data with 1000 rows for a sql server table named RegExTest. It should insert data into a column named Value”.

He had mixed results with ChatGPT, which I tend not to use. Instead, I tried this in Claude. I like Claude as I get an explanation separate from the code. Here’s a view of the page:

Here’s the code, which is nice because I could steal certain patterns out of here if I had the need. The explanation was also good, and I’ve reproduced that here:

<start claude explanation>

This SQL script creates 1000 rows of diverse random string data for your RegExTest table. The data includes various patterns that are useful for testing regular expressions:

• Email addresses (user@domain.com)
• Phone numbers ((123) 456-7890)
• Dates (MM/DD/YYYY)
• Mixed alphanumeric with special characters (ABC-123_XYZ)
• Numeric strings (123456)
• URL patterns (https://www.example.com/path)
• Code patterns (AB12C345)
• Words with spaces (Hello World)
• Mixed case strings (AbCdEf123)
• Random alphanumeric strings

The script uses a recursive CTE to generate 1000 rows and applies different string generation patterns based on the row number modulo operations. This ensures you get a good variety of data types to test your regular expressions against.

The script also includes table creation (if it doesn’t exist) and verification queries to confirm the data was inserted successfully.

<end claude explanation>

There was an issue in the else part of the CASE. The last CONCAT call only had the first parameter and CONCAT needs two parameters, so I copy/pasted the 2nd parameter from the last THEN statement, which worked fine. In this case, that parameter was an acceptable choice to create a random alpha string.

Perplexity

I got a very weak result from Perplexity. Essentially, write insert statements, though you could use a script.

I didn’t see a script, so I asked. I got this.

Here’s the SQL Code:

That’s weak. It works, but it’s weak and I think I could have written that in not much more time than the AI took.

It did end with this sentence: Let me know if you want the script in another format or with different string lengths!

I did and then got a series of different scripts for different formats, but each was producing a separate 1000 rows of only that format.

DeepSeek

I download the DeepSeek coder model and decided to try that. The basic prompt was disappointing. The model basically wanted to use Python and only imported modules without other code.

When I said that and asked for code, I got more details, but with the script separated out into sections. The top seemed to repeat a bit.

Then I got each part of the script.

I didn’t run this, but it’s a reasonable way to do things for developers. For data people, this seems like overhead.

I asked for SQL, and got a script for Oracle, but more interesting, the code is for 1000 rows, but the comments say ten thousand. Can’t the GenAI count?

I asked to change this to SQL Server and got this code.

I don’t know what to say except that I’m disappointed in the local deepseek model, which is not only slow, but hasn’t produced a good answer.

Summary

Claude clearly wins this experiment.

I wrote an editorial on the view of GenAI tech from execs and someone commented that they wanted references for results. Most of us might take a result from a co-worker and use it without asking for any proof of where it comes from. If we trust our co-workers, this can work well. If we don’t, then we might want to know where they learned this and how they know it works.

For GenAI, we might want some references to help us learn more or check something. Or, more likely, learn how to modify something that isn’t perfect.

In this sense, GenAI can be better as this gives us the references we need to learn how the answer was generated.

This post looks at how I got references in a few cases using Claude, Perplexity, and Copilot.

This is part of a series of experiments with AI systems.

Asking a Question

I decided to ask a question that I was asked by someone and I used Copilot to find the answer. Here is my prompt:

I need to schedule a Powershell script to access a SQL Server and want to use a managed service account to run the script. How can I do this in powershell

Let’s look at how the various tools I’ve used respond.

Claude

Claude.ai is from Anthropic and it’s a tool I’ve enjoyed using. When I pasted in the prompt, I saw this response:

I got to the bottom and no references, but when I asked for them (see my prompt in the image), I got a few to look at.

That’s one of the tricks with AI. You can ask it to explain itself.

Perplexity

My second test was perplexity, which I started to use a bit after Grant mentioned it.

Right at the top, I get some references that I can click. When I get to the bottom, I have a list of possible future prompts, which aren’t references, but they get me thinking about things I might need to consider.

That worked well and I liked having the references, though I’d prefer them at the end.

Copilot

I decided to try Copilot inside of VS Code, which is where I’d likely use it. I entered the prompt as a comment in a blank .ps1 file. After my first line, it suggested a few more comments to give context. I accepted a few and then added the last #include references line.

I didn’t get references and the generation stopped at the #End of script line.

I then added the comment below asking for URL references.

Each of the items above appeared as a line in italics, like other suggestions. I hit Tab to accept and I’d get another reference. When I saw the last one repeat, I stopped the responses.

Summary

I’m not evaluating the code here, just trying to see if I could get some idea of where the code came from. In this case, I don’t deal with MSAs a lot, so it’s good that I can get some URLs to check docs or get more options other than the code I found.

If I asked this question on Stack Overflow, I might get the same code (or similar), but getting URLs or references isn’t always easy there. Some people provide them, some don’t, and some might just close my question and say I should have searched better.

I think the GenAI does a good job here of giving me a starting point, and a place to go, which is helpful.

I read Brent’s first look at SSMS and Copilot in there. He didn’t have a great opinion of the tool, especially comparing it to Gemini or Claude. I haven’t tested this yet, mostly because I’ve got other stuff going on, but I decided to take his prompt and the code and drop it into the dedicated Copilot App on my Windows 11 laptop.

This is part of a series of experiments with AI systems.

I have a button that starts Copilot, so I pressed it and put in this prompt:

This was copied from Brent’s post, and then his code copied and dropped in.

I was curious to see if this worked.

My Results

I thought my results were better. The first part of the response is shown below. Copilot sees the scalar function as a problem and rewrites things. I don’t love that this is just the query and not the CREATE OR ALTER PROC part, but still.

Below this, it notes that I should test indexes and gives me the ones it would recommend. Then it goes into the fact that order by desc can be problematic. It also has a TOP 1 recommendation for the first query.

The rest of the response is OK, and gets into some general things that might help me go past the recommendations.

The prompting matters, as if I click on the “what are more perf opt techniques” I get a lot of generic stuff.

Is AI going to make you a better DBA or developer? Maybe. It will give you things to think about, but you need to test them. You might even ask a GenAI tool for help on that. I did and got this:

And below, some more steps.

Not bad, and likely what a lot of people do, though perhaps not with #5, in which HammerDB is suggested to run this proc.

Learn to use GenAI tools, learn where they help (And save time) and where they can’t be trusted. That’s likely an ongoing battle for most of us, but these tools can help us be more productive.