why you almost certainly don't need to be a part of the a.i. revolution

Not everyone needs a "pivot to AI." Far more often than not, off the shelf models are all your company needs, and even that might be overkill.

One of the most overhyped dishes on social media and the news is the 24 karat steak at Nusr-Et, the steakhouse chain owned by Nusret Gökçe, better known as Salt Bae. It is exactly what it sounds like; a tomahawk steak wrapped in gold leaf and served on a wooden cutting board. If you’re a celebrity, Gökçe himself will come out to assume his signature frown, which could be easily confused with either severe constipation or the aftermath of getting overzealous with an anal toy, and very dramatically slice and — of course — salt it himself. The price for all this? Just $2,000 before tax and tip.

Now, if you really want a great tomahawk steak but don’t want to spend monthly rent on one, you can buy amazing aged Wagyu cut for $90 and order some 24 karat gold leaf on Amazon for $9 to have pretty much the same exact experience. Just let it rest after grilling or a reverse sear, serve it on a wooden cutting board, and… You know, I’m not going to finish that joke, but the point is that about $1,900 of Nusr-Et’s steak is a combination of smoke, mirrors, and hype.

Believe it or not, the same concept applies to the vast majority of AI startups you see dominating your news feed, as was recently detailed with a great deal of, admittedly, very justified rage and expletives by data scientist Nikhil Suresh. And while I’m not a data scientist, I design enterprise software, tooling, and frameworks, which, yes, can include the occasional AI model, so I can very much second his argument, with just a bit less frustration.

You see, many of the AI startups in question are little more than automated wrappers around OpenAI’s ChatGPT, meaning that there is no algorithm or model running in the background, just some basic code which talks to ChatGPT to deliver 98.5% of their advertised functionality. When a major update allows ChatGPT to do something new, these startups will be dead in their tracks and have no reason to exist.

This isn’t exactly a real business model because the only real asset they have are the authentication tokens for ChatGPT API calls, and the only thing they do is give you a way to access said API through a different app, a kind of DoorDash for chatbots that doesn’t actually do anything but present their own versions of restaurant menus but you still had to call the restaurant to order and drive to pick up your food.

Other AI companies aren’t going the parasitic middleman route, which is nice. But the actual AI they use are existing models from open source libraries to run on a standard AI toolkit to do a part of their work, or add it so they can process input from a text box to do something that used to be very easily and straightforwardly done by a button, or a dropdown, or a drag-and-drop.

It’s revolutionary in the same way as having Clippy replace a simple wizard, or adding Stripe or Square SDK (development kit) to process payments would be breaking new ground. Kind of like the the Nusr-Et gilded tomahawk, what they’re selling to investors is 95% hype by volume, their core competency — which could be executed extremely competently and fully satisfy their customers — wrapped in cargo cult ML.

trying to revolutionize in all the wrong places

If you remember the days of crypto mania, you may recall that a small company called Long Island Iced Tea changed its name to “Long Blockchain Corp” and saw its stock almost quadruple in a move that eventually led to insider trading charges. Well, a lot of AI startups and companies “pivoting to AI” are doing very much the same thing.

As long as they can credibly claim they’re running a Tensorflow/Keras model to do a thing, investors will happily flood them with cash and analysts will rate them as buy and holds. But the dirty secret is that many of them don’t even really need AI models because they don’t have the necessary data to make them useful, don’t know what problem they’re actually solving with it, and probably already struggle with a simple CRUD, or create, read, update, delete app as it is.

“But come on Greg, companies have so much data these days! Can’t they, you know, use AI to mine it for insights like the commercials say?”

Short answer? Probably not. A lot of executives tend to think that their data lives in a digital museum where every exhibit is carefully inspected, investigated, labeled, and indexed in a catalog. In reality, any data that’s not crucial to the core mission lives in a sort of hoarder’s nest, with stacks of moldy, decaying newspaper and rats scurrying between stacks of trash bags. Or, very likely, deleted to save on storage costs.

You logged everything? Awesome. We dumped it in a file collection or a document database with everything we could. Good fucking luck making sense of 67 terabytes of every click, error, and crash report that cost you millions a year to maintain. Why? Because these are not the things the business actually cares about, like SEO stats, or web traffic analysis, or sales numbers, or referral data. None of that required AI. Just experts who knew what questions to ask so the right reports could be created.

You see, that’s the thing with AI. You can’t just feed a ton of data into is and say “okay, make me a trillion dollars” because it has no context, frame of reference, or know how to parse your data, much less what kind of meaningful outputs you expect. (Someone should probably tell that to Sam Altman because that’s his exact pitch for OpenAI.)

To get all those new insights, you need to understand your goals, and no, “make me more money” is not a real goal. Rating the probability of certain parts breaking down given specific parameters of wear and tear, and the intensity of its use to figure out if certain assembly line setups will be more cost effective than others is a real goal. But do you have reliable data for all of that? For many companies, probably not because it isn’t their core concern, but just a cost of doing business.

Does it even matter to be as precise as that if you already have a budget that covers a whole bunch of configurations and it would cost tens of thousands of dollars to come up with the data, train the model to an acceptable accuracy, and cough up a range an experiences plant manager already more or less knew? That money could’ve went to bonuses, repairs, new equipment, upgrades, places where it would make more sense. And yet, you need to add that AI so the market sees you as an “AI company” because every company now has to be that because investor reasons.

if you have to ask, the answer is probably no

Of course, none of this means that AI is useless, or there won’t be an AI revolution. It’s just that said revolution won’t happen in the lands of gray cubicles, TPS reports, and too many Microsoft Teams weekly status report meetings. Virtually all of those places will be served perfectly fine with simple linear regressions, naive classifiers, and a few tokens to use a cloud-based chatbot or a nudity detection model. It’s frankly, a waste of time, effort, electricity, and cash for them to do anything more.

This is not me trying to be mean or look down at anyone. It’s just that while tech CEOs and their fan club pitch it as the solution to all our problems now and forever — based on fanciful flights of fantasy presented by a handful of people who had no idea what they were talking about — and plan a grandiose future of a trillion humans across the solar system based on the idea that AI will certainly make in happen, the reality is that these models are highly specialized tools custom made to solve a specific problem.

Sure, that problem can be vast in its scope but so will the model. There’s a reason for ChatGPT requiring over half a million kilowatt-hours and 175 billion parameters to do what it does. But its design focuses on breaking down phrases, sentences, and what words are and how they relate to each other.

You can’t expect to just pour a random bucket of data into a random AI model and get anything meaningful to come out unless you have a great way to break that data down into a consistent structure, with accurate, consistent metadata, what formulas you will need for data processing, and have a very good idea of what you expect the output to look like and why so you can train it. Otherwise, the model will either break as soon as you try to run it, or barf out random numbers because it has no idea what the hell you want from it.

What problems does a widget wholesaler need to solve that require a super custom AI model? What do insurance companies need that hasn’t already been covered by one of the dozen linear regression models they already use? If you’re creating brand new medical imaging devices, materials, antibiotics, or analyzing scientific discoveries, a custom model will shave literal years of effort off your work and help point you to new breakthroughs that can change the world and billions of lives for the better. If this isn’t in your wheelhouse, you can probably just grab something off the shelf.

Lest you think I’m being condescending, all of this applies to my work too. In the past, I tried to implement a models to analyze service states and uptimes, and trigger self-recovery tasks if a crash was imminent. Sounds awesome and new, right? Nope. After a few attempts I quickly found that a Prometheus alert could’ve done the same job. All I had to do was configure already existing settings.

In short, if you find yourself scratching your head and wondering how you can use AI and are finding yourself either looking at downloading mature models or coming up empty, it’s okay to say that maybe you don’t really need to start your a data science department and pay in the mid-six figures to hire experts to swear under their breath as they rummage through your company’s data ware-dumpster, looking for anything remotely useful and interesting they can maybe use in a model. You’ll save yourself a lot of time, money, and nerves.