Just got back home after a conference and wanted to jot down a few thoughts to mark the moment in time. I wanna be able to look back at this in a couple years and revisit where we were then, kinda like how I made notes about what life was like during COVID.
William Gibson supposedly said, “The future is already here – it’s just not very evenly distributed.”
Self-driving cars are like that. A close friend of mine has a brand new Tesla Model X, one of the last ones, equipped with the Hardware 4 gear and the Full-Self-Driving 14 software. It opens his garage door for him, meets him in the driveway, opens the car door as he approaches, closes the car door after he’s inside, closes the garage door as it drives him out of the subdivision, navigates through the suburbs and onto the highway, and takes him right to his destination. Granted, we live in Vegas, a nearly-always-sunny location with clearly marked roads, but for all practical purposes, it feels like the future we were promised.
I have other friends with Teslas that have earlier versions of the hardware, not capable of getting FSD 14, and they have very different levels of experiences. And then of course there’s me, and none of my cars even stay in their own lanes, let alone do any kind of self-driving. It’s not like I’m poor, either: it’s just that the Venn diagram of cars I wanna own, and cars that have good self-driving, have nearly no overlap.
So if you ask six people to describe the abilities of self-driving cars today, you can get six different answers ranging from “it’s amazing” to “it will kill you”, much like the parable of the blind men describing an elephant. All of those answers are simultaneously correct, and yet also incorrect.
AI is like that today, too.
The exact conference, presentation, and presenter doesn’t really matter here, so I’m not going to name them. The point is that at this recent conference, I sat through sessions where I heard a few presenters say things from the podium like:
- “AI doesn’t have runtime access to your database, so it can’t see how its recommended query changes will perform”
- “AI’s knowledge is limited to only what it was trained on, and it can’t infer new things about new systems”
- “AI can’t evaluate query optimizer cost calculations because it doesn’t understand your data distribution”
- “AI can’t identify the root cause of performance issues”
- “AI can’t reason about CPU, memory, and I/O tradeoffs while tuning a query”
Uh… maybe your AI can’t.
But there are most definitely AIs out there today that do all of that stuff, and I should know, because I use them on a daily basis.
I’m a Microsoft SQL Server consultant and trainer, and I run Claude Code Desktop on my MacBook Pro. I use it for writing demo queries, First Responder Kit changes, and client work. Claude Code can still comprehend exactly what’s going on, despite being on a Mac, and it can test its work against Azure SQL DB, Amazon RDS SQL Server, local SQL Server VMs, and some in Amazon EC2. It’s fast, works well with Github, and accomplishes things in minutes while I’m off doing other stuff.
When something in my environment goes wrong, I can ask Claude Code to troubleshoot it. I don’t maintain production Availability Groups or failover clusters, but if I did, I’d be quite comfortable using Claude Code to troubleshoot those as well, plus make plans for upcoming deployments or environmental changes.
Claude Code can connect to your database, read your statistics, try different versions of a query, measure its overhead in terms of CPU/memory/IO, help you understand the tradeoffs, and much more. Just yesterday, while sitting in an airport, it was helping me use undocumented tricks against undocumented DMV columns – stuff that most definitely wasn’t a part of its training data, but it was experimenting and trying new stuff so we could both learn together.
The only thing that would have been cooler is if all this was happening in a self-driving car.
Careful when you listen to AI sessions at conferences.
The state of the art is advancing really quickly, and the people standing up on stages talking about the state of the art are like six blind folks describing an elephant.
- Some vendor staff will over-sell what their own products can do
- Some vendor staff will be honest about what their own products can’t do
- Many vendor staff will just be plain ignorant about what their competitors can do
- End users in the field will have the most relevant experience to share, but they need to be clear about which AI platforms they’re discussing, and make it clear that they’re only describing the part of the elephant they’ve been able to put their hands on
- Some (most?) of this stuff will be outdated within 2-4 months of the conference talk
- Some of it was already outdated when the slides were due, but the busy practitioner didn’t have time to update their slides, so listen really carefully to what they say aloud, and ask plenty of questions
I don’t think any of these people are being purposely misleading or malicious. They’re just blind folks describing an elephant. We’ll get a better picture of the elephant over time.
7 Comments. Leave new
Lol, self driving, for me, is overrated. While I am amazed at my MYLR with FSD, I cancelled the subscription since I didn’t use it that much and a few things it did were really, really annoying. That being said, in about 10 years, I’m likely only going to want a car that self-drives, since I’ll be much, much older.
The AI stuff is very distributed unevenly, and even the experiences. For me, it’s mostly really helpful, sometimes stunningly amazing, and a little too frequently can be horrifyingly bad.
That being said, it’s getting better all the time.
I get why my friend loves FSD. He doesn’t really enjoy cars, but he has to visit a bunch of Vegas businesses each week as part of his work, and FSD lets him be productive during that time.
Because I don’t have to commute for work, I’m still not quite sold on it. Usually when I get into the car, it’s to go somewhere for fun, so I don’t mind driving. However, for road trips, I can see how it would be spectacular.
If I could read while it drove me, I might reconsider, but for me, I don’t need to get a bunch of texts or other things that could work with voice done. So it’s less helpful.
It is way more relaxing to cruise along and I do like Autopilot. If it did lane changes on demand, it would be amazing. I hope that’s the bar for driver assistance moving forward
FSD 14 does indeed do lane changes on demand or automatically. You can put it in “chill” or “Mad Max” modes, and in mad max, it jumps around from lane to lane as it passes people.
Online reports about it are all over the place, but I rode with my friend in his X, and watched it zip around from lane to lane, dodging traffic, speeding past folks, all without any input from him.
I should add – this is why I keep saying HW4 and FSD14. Evidently the hardware version 3 (HW3) cars can’t do FSD14.
HW4 came out in Model S built from March 2023 on, X from March 2023, Model 3 from Highland (2024), and Model Y around 2024. Before that, the cars may never be able to do full FSD 14, and there’s a big controversy around that.
My HW3 car worked really well. With FSD it lane changes, but the base Autopilot doesn’t, which makes it worse than SuperCruise and BlueCruise. That’s the “base” I think should be met.
My last subscription with Aggressive mode ran slower than expected. Maybe Mad Max isn’t with HW3, but I was surprised how often it went slower than I wanted. Quite the change from late 2024 when I last subscribed.
I work for one of those vendors you mention. Spot on. This stuff is moving so quickly that if a vendor is selling you an “AI product” you should probably consider NOT buying it. Things are moving too quickly. The lifecycle of the “product” from inception to GA is still measured in months, and that’s too long. What we really need are good, inviolate patterns. Start with “open source as much as possible”, deploy things as consumptive services, make sure everything is plug/play. Said differently, instead of buying an AI tool baked into the query interface/tightly coupled to the db…instead ensure the vendor as a good MCP for the db, understands the patterns to access it from vscode/claude code…and interops nicely with them.
I’d also add that the paradigm is changing radically quickly towards “solve the business problem first, deal with the infra/code/architecture later.” So, if I need to deploy Microsoft Fabric….instead of doing that via CLIs, etc, you should be able to describe the business problem and the LLM should generate the code.