Voice of the DBA

Tag: AI

  • The AI Budget Assistant

    I saw this article about AI helping IT leaders plan their budgets, and I had visions of certain departments getting cut, especially data-related ones. My concern is that a lot of what a DBA does can be acting as insurance or supporting others and not necessarily doing things which are easily visible or that might show up in reports, praise in Slack/Teams, or in other places an AI might be prompted to look. If an IT leader has to decide what’s important for the AI to look at, is data included?

    Of course, it’s entirely possible the AI would look at all the performance problems and want to spend more budget on the data side. Fingers crossed for that interpretation.

    The idea of an AI looking at a large amount of current data in various spreadsheets, reports, goals, requirements, usage, and more to synthesize a budget makes sense. This is the type of thing many AI-type technologies do well, finding patterns and summarizing the data in a way that assists a human in producing a final report of where to spend money.

    Just like any report produced by a human, I would hope that someone reviews the results and verifies the analysis makes sense. However, in reality, I suspect a lot of human reports are passed along as-is, with only a cursory glance. The world of AI might make this worse as we can produce larger and more detailed reports.

    The funny thing might be that whoever has to approve the budget might let their GenAI model summarize the budget for accuracy. In that case, do we need humans involved? I can see some managers trying to reduce the staff that might have internal knowledge and replacing them with GenAIs. Will we just have AIs producing things other AIs read and summarize?

    Most of us know that garbage in/garbage out is a constant problem in the world of data and reports. Judging data quality and ensuring the accurate (and complete) data is fed into any system is something that humans need to be a part of. An AI model might assist, but humans have context and knowledge that can be hard to craft into a prompt. After all, this might be the training data for your model, and I’m not sure general models will learn well without handholding from humans, especially in specialized areas like budgets.

    I’ve had to build budgets, and while I don’t think it’s hard, it’s time-consuming. There is a lot of drudgery, and I think an AI can help here. However, I’d want to audit and check the data and ensure that my helper is summarizing the data in a way that makes sense to me. Even then, what I’d most like to do is let the AI give me a few suggestions of future costs and let me fine-tune those before I present them to anyone else. After all, if we make large errors, I suspect I’ll feel the pain of failure more than my AI assistant.

    Steve Jones

    Listen to the podcast at Libsyn, Spotify, or iTunes.

    Note, podcasts are only available for a limited time online.

  • Another AI Faux Pax

    I was experimenting with a local model (article) and as a part of this, I pulled down a web interface for my model in a container. I ran it and had it in the background. I’ve randomly been asking questions as I need to look up things. I’m mostly curious how well it works.

    This week I asked about CDC and got and interesting response.

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

    Updating a CDC Schema

    I asked this question of my local mistral model: How do I alter a table in SQL SErver that is being tracked by CDC to add a new column?

    That might not be the best way to phrase things, but I was on a call with a customer who was going to enable CDC and wanted to discuss how to manage this with schema changes. I’m not a CDC expert, so I was checking some documentation while chatting and decided to put this query in.

    I got these results,

    2025-01_0153

    I thought I’d missed something, as I had never heard of sp_cdc_refresh_schema. Sure enough, I didn’t see it on MS Docs.

    2025-01_0155

    If I were trying to do this (pretend I’m a copy/paste DBA), I would see this:

    2025-01_0156

    So, the next step is to call the AI on this. It finds that it hallucinated and give me other ideas.

    2025-01_0158

    It gets worse. Now I create a proc and insert into the system tables, which isn’t recommended.

    2025-01_0160

    This won’t work either, but I can see people trying this.

    First, if you get advice from the Internet, please always, always use a sandbox to test (what I also recommended to the customer), and second, treat any AI response like it came from a teenager.

  • Setting up a Local LLM

    I wanted to experiment a bit with an LLM and training it, so I decided to try a few things. I looked at a few tutorials (see the references below) and then finally got this working. This post distills down what I got to work.

    Getting Started

    I like containers, and rather than install something on my machine, I decided to get docker images for ollama. I ran this to get started:

    docker pull ollama/ollama

    I ran the Docker container with this command:

    docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama

    From there, I pulled a few models into the container with these commands:

    docker exec -it ollama ollama pull mistral

    I then ran this to start the model and interact with it.

    docker exec -it ollama ollama run mistral

    Here are a first few things I typed in to test the model.

    2025-01_0111

    From here, I exited and then ran this to start the model in detached mode.

    docker exec -d ollama ollama run mistral

    From there, more experiments, but that’s in another article.

    References

    Here are some places and tutorials I looked at:

  • The AI/Human Spectrum

    I was asked this question recently: is it more likely that AI will replace humans or assist them in their work?

    It’s a good question. If you think about the way AI is being hyped in 2024, many people think AI is, or will soon be, replacing people and we need less of them in work. I guess the simplified view is that AI can do the jobs of many people, but I’m not sure the world is that simple. What I think is more likely is that AI becomes a lever that assists a few people in getting more work done and potentially replacing other, less knowledgable humans.

    Maybe it’s the ultimate, do-more-with-less pressure that the management in many organizations places on workers that has people looking to AI to help. Get more done this year, but we’re not adding staff, and potentially we’re removing a few staffers. Maybe AI can do your job?

    AI can be a lever, and I do think that there are tedious tasks that we might have AI take on for us. We’ll need more of an agent/proxy approach to AI systems, as now I can ask AI to do some things that relate to text/images, but not actually do work for me. I want an AI to actually set my Out-of-Office for me, not give me instructions on how to do it or write the message for respondents.

    The current Generative AI/LLM systems aren’t really smart or intelligent, but they do process vast amounts of data and mimic the responses that other humans might give. If you work in an area that can benefit from that type of interaction, maybe an AI works well as a lever and lets you get more done on any given day. For some jobs.

    However, getting more done might not be enough. If I pair program with someone else and they write a bunch of poorly performing code, it takes me time to judge the quality of what they’ve written and then additional time to fix it. Getting more done in that situation can be a burden because the additional code produced requires additional rework. I might get less work done in some cases if the code is low-quality and I use a lot of time to rewrite or improve the code.

    However, if I am tackling simple tedious tasks, perhaps basic CRUD work in an application, maybe an AI can generate enough SQL, web, C#, etc. code that the job is done quicker. Maybe not at the most efficient level, but how many of you think the code for your internal applications is amazing? Is it good enough? Can an AI do “good-enough” work?

    As with a lot of dramatic changes to technology, I find myself going back and forth on the value produced. Some days I think the tech is amazing and some days I think it’s akin to the stuff I shovel out of horse stalls. AI is in the same boat with me, and while I think there is potential, I know that there are also downfalls and potential detractions from its widespread use.

    Certainly, the need to evaluate and judge quality is a challenge with AI, which leans me towards the lever that assists talented humans and replaces less talented ones. The other issue is cost. The LLMs are expensive, and use a lot of compute power. I’ve seen some smaller models, perhaps tailored with RAG or other methods of refining (and limiting) their use will overcome that, but who knows. The current models, however, cost something to run and someone is going to have to pay for that. Is there enough ROI to do that?

    Lastly, trust. Can we really trust an AI to give us accurate responses, or even perform work on our behalf? We have that problem with other humans now, but they work slowly compared to a computer. Can you imagine the problems that a rogue computer system could create with access to change things in the real world?

    Answer my question today. Are AIs more likely to assist or replace people?

    Steve Jones

    Listen to the podcast at Libsyn, Spotify, or iTunes.

    Note, podcasts are only available for a limited time online.