iAI is obviously the hot new thing and prognostications for the future range from the unicorns-and-rainbows life of idleness for us all to the post-apocalyptic killer robots hunting down the cowering survivors of mankind a la Terminator 2. I confidently predict that immediate reality will be somewhere in between 😊
As we’ve become more permanent here in Camp Savage, I’ve been working on our network infrastructure. It’s pretty sophisticated and pretty complex. It’s all Unifi gear, with multiple VLANs, Wi-Fi networks, VPNs, NAS, long radio links, runs of fiber, and myriad APs, switches, sensors, and IoT clients. There’s a lot to configure and keep track of and I’ve learned, relearned, and remembered a ton in the last few months.
Although Unifi gear is really top-notch when it comes to interoperability of all the hardware in their range, there are lots of gotchas with configuration, physical connections, and other networking esoteria. Whenever I’ve become stuck, I’ve turned, of course, to Google to search for help. Its AI Overview has been mostly invaluable for cutting through the tedium of watching YouTube videos and scouring blogs for answers. But occasionally it’s taken me on wild goose chases because I didn’t know any better.
Check It Out!
The easiest example to explain is when I had random client disconnects happening in one building only. It took me down the path of isolating the most twitchy client onto its own network, and changing all kinds of parameters for hours to try to stabilize the connection. After several days of trying lots of things and seeing if overnight performance worked, someone on a forum suggested a totally different path. They asked what’s the connection to this building? Fiber. Any fiber errors in the Unifi logs? Yes! Upshot: I had multi-mode fiber SFP adapters with single-mode fiber cable. Swapping out the adapters fixed the problem instantly.
AI wasted my time because it didn’t have any way to think that far outside the box, because it hadn’t encountered that explanation or diagnostic step. Maybe a more powerful AI would have, but it took a human to actually point me down the right path.
The most frustrating examples were the many times I asked for steps to config X or Y and it confidently gave me steps and screenshots based on UIs or hardware from several versions ago, even when prompted with the exact version being used.
And over to SQL Server…
Jonathan has encountered several cases of long-term clients introducing AI developer tools that have instructed devs to make changes that have crippled their systems in production because the tools just did not get the complexity of the environment and made assumptions. Wrong assumptions, costing the client time and consulting dollars to identify and put right.
These devs blindly followed the advice of their expensive AI tools without understanding what they were doing and the possible ramifications and got bitten hard. If you’re trying to understand a performance problem and just blindly trust the AI tool to diagnose and propose a fix, it might get it right. It might get it wrong. Will you understand what it did and why?
And now more specifically to my specialty and my upcoming precon at PASS in a few weeks: internals, how stuff works, and wait statistics.
Mangling the well-known proverb: you can’t keep just giving fish to someone otherwise they get further and further from being able to fish themselves, and eventually when the fish-provider isn’t there, or if the fish are no longer trusted to be edible, how does the person eat?
Don’t stop learning. Don’t rely on tools that try to do your job. Absolutely use AI, but only as a tool to help guide you. Be sure to do further research or ask for references / links / sources, and verify them! Don’t give up your ability to understand SQL Server; it’s what will separate you from the rest and keep your servers productive.
Thoughts?
3 thoughts on “AI Stories and Thoughts”
I’ve been using Unifi for my home network for nearly 10 years and it’s been amazing. I recently upgraded some of the kit and used Claude to help me optimise it. As well as screenshots which it did extremely well with I got it to write some JavaScript (which I ran in the browser tools console when I was logged in) to download the configuration JSON which I then upload to Claude for analysis of misconfigurations and potential optimisations. Worked a treat – Claude found lots of issues that needed addressing! Since then someone has created this although I’ve not tried it: https://github.com/jhammant/unifi-cli.
These days I paraphrase the old IBM quote a lot: “A computer can never be held accountable, therefore a computer must never make a management decision.”
I think the latest frontier models are incredible tools and the local LLM scene is also very interesting.
But at the end of the day, you have to own the output and assume responsibility. Blind trust in AI is pretty similar to blind trust in Stackoverflow. By all mean use the tools but take the time to read, think it through and test.