The Ebb and Flow of AI
It’s the end of the year. Everybody’s coasting and looking back. Let’s kick our heels up, get comfortable, and talk big picture industry stuff.
You and me, our weirdo jobs involve:
- Databases – which is already unusual, and most people you run into on the street have never heard of it, even though it touches every aspect of their lives
- One specific kind of database in particular – Microsoft SQL Server (and its various flavors, like Azure SQL DB and Amazon RDS), which isn’t even the biggest database out there, not by a long shot, so we’re a niche in a niche
- A few particular database tasks – writing new code, performance tuning existing code, and server maintenance – we’re not even the biggest job in our industry, which would probably be database developer or report author
When you think about the entire population of Earth, our specific job is extremely small. There just aren’t that many people who do what we do.
Because we’re so niche, companies don’t really make all that much money building products to help us do our jobs, or eliminate our jobs altogether. Sure, there are definitely companies out there that are spending tens of millions of dollars a year to automate our job out of existence – but are they going to be successful? Should we worry about that? To find out, let’s zoom out and look at the bigger picture.
A More Lucrative AI Target: Self-Driving Taxis
If you zoom out to the population of Earth, an example of a larger industry would be transportation – getting people and goods from one place to another. If companies can make that easier, or eliminate human labor involved with that, then companies would make muuuuuch more money.
So when new golden hammers like artificial intelligence and machine learning come around, companies wielding those hammers look for the most money they can make, and they apply the hammers there. Go where the money is.
Google, Apple, Tesla, and General Motors all threw billions and billions of dollars trying to make self-driving taxis happen. Whoever figures that out first, and successfully scales it, could not only be the next Uber, but simply blow Uber out of the water altogether. When Uber’s worth over $100b, and General Motors is worth half of that, GM’s gotta look over at Uber and go, “I wanna take you down, buddy, and if I spend $50b doing it, I’ll still make money.”
And boy, would they make money. Not only would they cut expenses, but they’d offer a better product. I travel a lot, and I abhor today’s taxi experience. Uber is a step up from what taxis used to be, but I can’t Uber everywhere. For example, when I touch down at a busy urban airport, it’s usually much more convenient to just hop into a waiting taxi as opposed to getting an Uber, then waiting 15-20 minutes in a manic ride share area. Taxis are a throwback to a different era – disgusting back seats, no air conditioning, and bad drivers with terrible driving skills and even worse conversation topics. I would love to get into a well-maintained driverless taxi, especially at the end of a long exhausting day of travel.
However, as surefire as the business model sounds, competitors are dropping out of the AI-driven taxi business. Very smart, very wealthy companies have learned (the very expensive way) that you can’t just rub some AI on taxis and take humans out of the equation:
- Apple worked on self-driving cars from 2014-2024, then gave up this year
- General Motors’s Cruise project started in 2013, and GM just threw in the towel
- Tesla – well, c’mon, Elon’s list of broken promises is a running joke, and he’s been promising a million self-driving taxis since 2019, and the Cybercab is still years away at minimum
- Waymo was started about 15 years ago at Google, has over 2500 employees, tens of billions of dollars invested, and still only operates in parts of 4 US cities
Self-Driving Taxis Aren’t Impossible.
They’re Just Expensive and Hard.
It’d be easy to look at the list of company failures above, then write a blog post titled something like “The Rise and Fall of AI.” Thing is, I just don’t think that’s correct. I don’t think AI is “falling” in the same sense that, say, the Roman empire fell. It’s just ebbing and flowing like the tide.
Companies invested tens of billions of dollars in self-driving taxis, and while we haven’t crossed the finish line yet, we (consumers) have seen a return on that investment. It just wasn’t the return that companies had been hoping for – instead, much smaller. Rather than eliminating the humans altogether, AI and ML are turning out to be tools that help humans be better drivers.
GM’s SuperCruise gets rave reviews, Mercedes’ new Drive Pilot system even lets you stop paying attention to the car altogether, and Tesla’s “full self driving” – well, it provides for plenty of YouTube comedy.
So How Will AI Reshape Databases? Slowly.
Using the knowledge of what happens when companies spend a decade and billions of dollars trying to improve complex tasks, let’s turn our attention to our own tiny niche industry – just as companies are. Companies are now spending time and money here, too, and some talking heads are saying humans will no longer be needed to build, tune, and troubleshoot databases.
One of the first companies to try it, OtterTune, got about $14M in funding across the span of 4 years to make MySQL and Postgres self-tuning. Richie and I tried it with SQL ConstantCare®’s database, and we had pretty abysmal results, but we chalked that up to it being an early product iteration. I worked with Andy Pavlo & his team on feedback, and I didn’t blog about our experience because I didn’t want to curse the product with bad early publicity. I guessed that given enough time & money, they’d solve the problem, and the world would be a better place for it.
Other companies will try the same thing. Most of them will burn through their investments without getting enough market traction, just as most self-driving taxi companies are setting billions of dollars on fire to be the next Uber. The self-driving database dream will be smaller, involving less research spending, because the end reward isn’t as big as self-driving taxis.
Along the way, over the next several years, those investments will result in better tooling for us humans to use, and I’m genuinely excited for that. We’re already seeing the earliest hints of it with things like SSMS Copilot and AWS DMS, but don’t let the sketchy quality of the early stuff scare you away from it permanently. The early versions of these tools will be just like the early versions of Tesla’s “full self driving” – only for the, uh, “brave.”
But before you panic about your job going away altogether, go take a ride in a taxi.
10 Comments. Leave new
Interesting post.
Another way to look at this is the old adage “People overestimate AI in the short run and underestimate it in the long run”. Replace _AI_ with any other tech trend…NoSQL, cloud, whatever.
In my mind, the big problem is folks wanting to whole-cloth replace some function/role/person with AI (or some new tech…again, NoSQL 15ish years ago). That’s not how things work and their is historical precedent…centuries of it.
Back in the Great Depression the WPA wouldn’t fund purchasing backhoes…the theory being it would put ditch diggers with shovels out of a job. But did it? Nope. It turns out the world realized there was a huge unmet demand for holes in the ground. You’ll find the same thing will happen with self-driving cars and “AI in my DBMS”. With the former, navigating centuries old roads in cities like Philly will be hard for self-driving cars. It _might_ never happen. Maybe the right solution is to do self-driving vehicles on interstate highways where it’s easy and then do the “last mile” stuff with a human driver. Maybe. I dunno, and neither does anyone else.
Maybe AI won’t EVER optimize the role of the RDBMS. Maybe instead the world will wakeup and realize that we don’t need to use relational tech to persist data as much as we think we do. I remember when sharepoint stored all your doclists in image datatypes. Dumb. Again, “overestimate in the short term and underestimate in the long term.”
Another way to think of it is folks have a very narrow view of “some _thing_ and that thing’s role”. RDBMSs are one. So you find the idiots that wanted to rip-and-replace every RDBMS with NoSQL. Why? Because some vendors touted “hey, you don’t need that expensive DBA who fiddles with backups and indexing…both of which aren’t need in the NoSQL world”. How did that work out for you? The same will happen again. Maybe we’re at a place where RDBMS tech is about as optimized as it gets. Maybe we actually need DBAs and 3NF (albeit for a more limited number of use cases). Another historical analogy is the replacement of doormen in hotels and condos in big cities. The short term thinking was tech (security cameras and modern elevators) can replace those folks because folks viewed that role narrowly. Well, it turns out doormen do a lot more and you’re seeing their return. Overestimation of tech in the short run, underestimating in the long run.
Similar to Dave’s comment above; if we can spend billions on HOV lanes with their own exits and on ramps, it might make sense to isolate self-driving cars in some way. High-speed from hub to hub, with out the exorbitant cost of mass transit. Back in the 80’s many spoke of the future of self driving cars to be based upon electronics embedded in the road. Well that never happened, but segregating traffic would make the Waymo lane extremely efficient (almost like a Japanese bullet train-They are Jaguars by the way).
I suspect DB-AI might develop similarly. Build a few lanes or two, and those self maintenance. It grows over time as it is tested and trusted. However, the over optimism of “We built an AI to run you entire organization’s DBA needs . . .” is much like any other technology we’ve ventured toward in the past. The reality usually lands between uber-skeptics and zealots 😉
And, imho, it’s not AI, it’s DI (Directed intelligence). It will have what we might consider analytic capacity, but we are building it separately for cars, database management, medical research . . . because, for now at least, it must be directed and have a framework. Uber drivers might be overwhelmed with the hub to home business, just like paper manufacturers were when the “paper killer” the PC came out.
> if we can spend billions on HOV lanes with their own exits and on ramps, it might make sense to isolate self-driving cars in some way.
I don’t _think_ this would really work, but I like this idea a lot. What we should be doing, from a gov’t perspective most especially, is take some _small bets_ on ideas like this. Try it out for a defined period of time in one location and see if it takes off. Too often we as a society go all-in on some new tech or idea and fund it to death which prevents the free market from working and prevents other nascent ideas from developing. Electric cars are number one. While they have benefits, the fact that we give tax credits, etc to them means we are crowding out other ideas that might be better like hydrogen or CNG.
The real beauty of AI, for me, is the ability to take small bets, generate the feedback loops, and figure out where the real value lies. I do AI rapid prototypes with businesses for a living and I have to restrain myself from laughing in a customer’s face at some of the half baked ideas folks have…but I never refuse to try them, because when we experiment we start to reach a consensus on what really works.
Last point…it’s always possible that two opposing things may be true at the same time. This freaks people out. It’s possible Waymo is great and so are human-driven ICE taxis. Perhaps the former only works in 4 cities where the infrastructure is newer, the roads are laid out in grids, the weather is decent, etc. Everyone else uses the latter. Unfortunately we are always looking for the revolutionary tech instead of the evolutionary.
What I’d really like to see happen is for Microsoft to “reshape” SQL Server to being at least as fast as it was in 2017. @019 and 2022 are jokes when it comes to performance even of simple tasks.
Instead of them wasting time on which version of crap code to use the next time it runs, they should some more time looking at whatever they did to cause the slowdowns. They should also spend more time on making truly usefully changes like maybe doing Norton-Style online defrags of indexes instead of the awful excuses of REORGANIZE and REBUILD for indexes.
I could go on but considering that it took them almost 2 decades to come out with a sequence generator and a true CSV BULK INSERT/BCP (and they still don’t have a BULK EXPORT) and nearly as long to come out with a splitter and then not do it right for 6 years, I would be surprised if it took them another 20 years to bring performance back to SQL Server. 🙁
I hear what you’re saying and I have no knowledge of what folks are thinking in the SQL Server product group. It IS appalling that seq gens, CREATE OR REPLACE syntax, and autonomous transactions have taken so long to build. It’s likely due to there not being much of a market for these things, as much as we all like to complain about them. As of today, and this could change at any moment, I think _the advancement of RDBMS tech for OLTP style workloads_ is dead. There just isn’t much need for new features or improvements to existing, when the world is focused on analytics and using polyglot persistence for data.
I’m not suggesting transactional workloads are dead…far from it, it’s just that folks have started to come around to the notion that not everything needs to go to SQL Server, there are good use cases for NoSQL, microservices might be a better approach, etc. As you pointed out, and I agree, raw performance hasn’t much improved for almost a decade…maybe there isn’t much left to wring out of that.
Praise be the name of our lord Clippy or else “it looks like you’re writing an e-mail” in SSMS!
AI is here to stay. It’s role in industry will evolve as companies try, fail, then try again to find a profitable way to use AI for real work.
I see AI as coming into its own with the industry push to classify data and implement DLP solutions. There are many players out there in this field, but the 600 pound Gorilla is Microsoft (again-duh). Their new Azure based AI suites use Generative AI’s, like CoPilot & ChatGPT4, to use pre-defined Security Information Types (SIT) in templates for the common use cases. For specialty use cases, there are customizable templates that use ML to train the AI by using sample data to learn the business rules. This is where AI is coming into its own for Azure in general, and Office 365 specifically.
The three MS platforms to watch are MS 365 AI, MS Purview and MS Priva. That’s where I’m focusing my personal training plans right now.
grubby taxis are the worst! I’d rather park my car at the airport and rent a car at the destination then hale a taxi or even an Uber.. I once had to change a tire on the freeway for a clueless Uber driver after a blowout on the way to the airport… I was filthy after that and still had to catch a flight! So looking forward to those self driving clean taxis.
AI will take a little while to take full effect as it will takes time to train the all human element to interact with it and fully embrace that if the machine does not work it must be THEIR fault – not the machines or the company that runs it.
Humans are notoriously inefficient, was recently ‘talking’ to a service providers AI system – which could it told me provide the answer to all my questions – only when it could not, finally transferred told that it was because “you are not experiencing the right of problem!!?!”. They then denied that their previous ‘colleague’ was not human and how could i possibly know it wasn’t as it always gave perfect answers – managing to both lie and admit so in a single sentence.
I asked what color their socks were – they responded ‘what the **** has that got to do with anything?’ – thereby providing evidence they were human by giving a rational response to a stupid question – rather than keep repeating ‘can you rephrase the question to better define the technical nature of your enquiry?’
My point – until AI can deal with people who don’t know what they (should be / are) asking for – it can’t provide the compete answer to anything. How many times have I (we) been told ‘their’ answer by a user / customer – when asked what their problem is.
Its just one tool in the toolbox – like everything else – including you and me.
https://www.youtube.com/watch?v=f9jVS_Bpx2E
Waymo is very cool. I’d love to get these in more cities, but I also get the challenges.
I tend to agree. AI is everywhere and hyped, but it’s going to change a lot of things in databases slowly. Heck, so many people still have SQL Server 2012 or earlier, MySQL 5.6, python 2.7 (8?), etc.