Voice of the DBA

Tag: AI

  • Smart AI Agents and Data Security

    I don’t know if Anthropic, OpenAI, or Google have really spent a lot of time and effort on security around their models. I know they talk about it (Anthropic, OpenAI, Google), but in practice, I’m not sure the researchers and testers are being as careful as the marketing and management of those companies want you to think.

    Gemini broke into three companies during a security test. Claude escaped a test environment and breached three organizations. OpenAI had breaches of Hugging Face and other sites, so perhaps this is the most powerful model, as it exceeded the “rule of three.” In all these cases, the agents weren’t doing anything amazing. These were acting the same as most malicious or bored human hackers who might find or guess credentials and take advantage of them. In some sense, this points to the generally poor state of security in the world.

    On the other hand, it could be that these high-tech organizations aren’t equipped to secure systems from the rapid, capable digital fingers of agents powered by a model. If they can’t, then to what extent can we be expected to secure internal AI agents in our organizations?

    For those of us who secure and manage data assets, do we think that our systems are well secured? Do we think our legitimate users, to whom we’ve granted access, are securing their credentials? What is the likelihood that we’ll have an internal breach from an agent that’s asked to find some data and discovers a) unsecured credentials in a file or repo, b) finds production data it can access in a non-production, less secure system, c) a backup of a database not protected that the agent can restore, or d) some other hole?

    I do think AI has a lot of potential, but it operates in a somewhat slapdash, reckless fashion. I’m not convinced most vendors (or models or software incorporating AI) will provide robust and comprehensive security controls. Even if they do, how many of us will know how to implement them well? Heck, we see developers and DBAs mess up SQL Server security on a regular basis, and I think that’s a simple system to configure.

    I’m excited by AI and also terrified of the possibilities.

    Steve Jones

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

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

  • Today’s AI

    “This is the worst version of the car you’ll own.”

    That was a quote from the delivery manager when my wife and I picked up our Lucid Gravity. He said that since the car is very software-oriented and gets regular over-the-air updates, the car keeps improving. The car I bought is the worst version because it keeps getting better. His experience over the last few years with the company has shown that the changes made in software result in customers growing more fond of their vehicles.Today’s editorial: Today’s AI

    Or maybe they’re less annoyed by things that don’t work well. Ah, the joys of software.

    I was reminded of this interaction when I saw someone send a note at Redgate about AI with this quote: “Today’s AI is the worst AI you will ever use.”

    Is that true? I guess, to the extent that we think the AI model you use today will improve and its capabilities will grow, that might be true. Certainly, the move from Sonnet 4 to 4.5 to 4.6 to 5 seems to be, well …

    The same to me. Has the model improved? By benchmarks and measures, it has. The same could be said for the GPT models from OpenAI. However, I don’t know that I have found much difference. Depending on my guidance, I think I’ve found that I’m as likely to get great results as poor ones from different models. I’ve tended to stick with Sonnet models, though often for the focused tasks I pick, the Haiku models work, as do the Opus ones.

    I’m cheap, so I don’t often see the point in using Opus. I haven’t found it to be clearly better, though; to be fair, I’m not letting it loose on large tasks, and I am conscious of costs. I’m not paying, but I tend to treat all things I do as if I were paying the bill myself. That helps me to be aware of when I am being effective and efficient.

    When I’ve seen my cars upgrade (first a BMW, then a Tesla, now a Lucid), there have been clear changes that I’ve appreciated. There are sometimes annoying upgrades, which always remind me of Jeff Moden: Change is inevitable… change for the better is not. That can be true, but I find that the changes made in software often make sense, and I can see how they an improvement. Even when I don’t like them.

    For a non-deterministic model, however, I don’t know how easy it is to measure change, and if the model is “smarter”, but I don’t use the extra smart-ness, is it better? I have no idea how to measure these improvements, and I am certainly skeptical that these benchmarks and the frontier model hype is driven more by the need for profit than by a better AI model.

    Steve Jones

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

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

  • Get Along

    I enjoy listening to Get Along by Kenny Chesney periodically, often smiling to myself. I was reminded of this song recently while talking about careers and the AI impact during VSLive San Diego.

    On one hand, most of the people I surveyed or chatted with were using AI to accomplish some tasks. However, very few are using it in a way that would eliminate their positions or do the majority of their work. On the third hand, lots of them worry about their organizations attempting to get by with more AI and less humans in the future.

    I think that last one is a natural point of view in the world today. I’m sure plenty of executives are hoping they can reduce their future labor costs with machine assistance. Quite a few are finding that the machine assistance isn’t that cheap, and humans are still needed. I don’t know where this will shake out in the future, but I guess that many of our organizations will shrink a bit and use less developers than they might have now. I also think there will still be plenty of developer jobs, but AI will be used judiciously to supplement the capabilities of workers.

    This means workers will need to know how to work with AI effectively and efficiently.

    Perhaps there will be new departments and new companies that spring up and create the need for more workers. I hope that’s the case, but I bet a lot of teams will lose some humans and replace them with AI. The cost of benefits continues to rise, and I don’t know that many companies will grow their businesses enough to keep all their humans and spend on AI. That’s adding lots of salary/token costs because of growth. Some organizations will have a lot of growth, but I suspect many companies will see that they have tasks that need humans and tasks that don’t.

    The common question from our discussions was “which humans will stick around?”

    Technical competency isn’t likely to be the deciding factor. You being a 10x engineer or the best coder might not keep you around. Spotting problem code and better guiding AIs will matter. The mentorship skills you use with junior staffers, and the patience, will become important. Some people are learning those skills now.

    The one thing most people agree on is that you should get along well with others. Those soft skills – the ability to communicate, being pleasant, making others feel not only comfortable but wanting to work with you – will matter most. If others don’t enjoy collaborating with you, especially your boss, I can see your position being precarious and potentially slotted for a layoff.

    This isn’t to say you can’t argue a point of view or debate an approach. The thing you should remember is that when you do present your view, it is seen as an engaging and spirited debate, not an antagonistic war. People want to work with others when they feel some bond with them. When they don’t, they might be indifferent or even dread interactions. Feeling dread, hesitation, or a reluctance to engage is a sign that someone isn’t a part of a team.

    Those who aren’t part of the team might be the easiest to replace with an agent in the future.

    Steve Jones

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

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

  • Finely Tuned Models

    I have no idea if this is true, but this post on X says that Thomson Reuters used information they’ve collected for decades to fine tune and train a model. They started with one of the qwen models and then spent $40 million to add their knowledge to the model. The post says this model is comparable to the Gpt5.5 and Sonnet 5 models. They have used 10% of their data, which spans over 100 years.

    I have many questions. But first, $40mm? How many tokens is USD$40mm? A quick calculation is trillions of tokens, and while I’m sure there is some payback if lots of their customers use the model for work, there’s also a compute cost every time they use the model. Perhaps they can fine-tune and train the model more efficiently over time, but I can’t see many individual companies spending this effort on training their own model.

    I also wonder what the time it took to train this. The story notes 2 years, but if I want to update this, then how much more effort. What’s the cost? This certainly reduces the dependence on the frontier models from Antropic/OpenAI/etc., but is this something that makes sense for other companies? If I choose to use the TR model, would it be cheaper than Sonnet 5? Maybe it uses less power, which is always good, but will T-R charge me less than Anthropic?

    I do think that fine-tuned models might make sense, especially if you can use a small language model and train it for a reasonable cost. Like $100,000, not millions. In that case, would it make sense to you? Would your company fund this themselves? Is this something that a trade group could do with support from multiple organizations? Or is the competition between companies so high that they can’t work together?

    I do think that AI has a lot of possibilities to assist humans, and focused models can be useful to solve specific problems, with a lot less cost than the best frontier models. I’m just not sure if the training effort is something that many are willing to put forth. It will be interesting to see how GenAI models evolve as the costs and resources needed by the frontier models continue to rise and smaller models become more capable.

    Steve Jones

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

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