Daniel's written in with a question about the Gartner hype cycle, and where AI actually sits on it. His argument, if I'm distilling it right, is that we keep talking about an artificial intelligence wave as though the technology dropped out of the sky around the time ChatGPT launched. But AI's been doing real work for decades. Weather forecasting, scientific computing, fraud detection. His point is that what changed a few years ago wasn't the existence of the field, it was a dramatic improvement in capability. So we're in this strange position of hyping something that isn't new. And the thing he's actually looking forward to is normalization. The moment we stop treating AI as a distinct category and just assume it's part of what software does. Vendors stop putting AI on the box. The question he's asking is how far we are from that, and where on the hype cycle the stable, boring plateau actually lives.
He's put his finger on something that's been bugging me about the vendor conversation for at least two years. The phrase artificial intelligence on a product page has become about as informative as saying the product contains electricity.
Electricity's probably more honest. At least the product does contain it.
Fair. But Daniel's framing is more interesting than the standard AI washing complaint. He's saying the hype cycle itself is a weird fit here because the technology predates the hype by sixty years. Gartner's curve assumes a new thing arrives, peaks, crashes, and then climbs to a plateau. AI has already done that. Twice. The field had a boom and bust in the nineteen sixties and seventies, then another one in the eighties with expert systems. This is the third cycle, but it's riding on infrastructure that's been quietly maturing the whole time.
So the hype cycle's already happened. What we're experiencing now is a hype cycle about a hype cycle.
Sort of. I mean, Gartner themselves have been tracking AI on the curve since the nineties. Machine learning hit the peak of inflated expectations around twenty seventeen, twenty eighteen. Then generative AI got its own curve and shot to the top almost immediately. The trough of disillusionment was supposed to follow. And there's a real question about whether it ever arrived.
Well, it arrived in some sectors. The companies that bought AI because the board said so and then discovered the model doesn't know their inventory system exists. That's a trough.
Right, but that's not a technology trough, that's an implementation trough. The difference matters. Daniel's point about weather forecasting is the key. The European Centre for Medium-Range Weather Forecasts has been running machine learning models operationally for years. The models are faster than physics-based simulation by orders of magnitude. That's not hype, that's just a better tool. Nobody at a weather agency is putting AI on a slide deck. They're just using it.
And that's exactly the normalization he's describing. The weather people don't have an AI strategy. They have a weather strategy.
Which is why I think we're closer to his plateau than the vendor conversation suggests. The gap isn't in the technology. The gap is in the marketing layer. It takes a while for the language to catch up to the practice.
So how far out are we? Give me a number.
I'd say we're three to five years from the point where saying your product uses AI sounds as dated as saying your product uses the internet. But there's a wrinkle. The internet analogy has a limit, because the internet was infrastructure. It was a pipe. AI is more like a material. It changes what you can build, not just how you deliver it.
That's a useful distinction. A pipe you can forget about. A material you have to keep making decisions about.
And the decisions are still hard. That's what's holding back the normalization. If AI were truly boring, you wouldn't need a chief AI officer. You wouldn't need model governance committees. You wouldn't have procurement teams asking vendors whether the model is hosted in the EU. Those are all signs of a technology that hasn't finished settling.
But some of that is regulatory lag, not technology lag. The EU AI Act went into force in August of twenty twenty-six. That's a month ago. Companies are still figuring out what compliance even looks like. That's going to keep AI as a named category for a while, regardless of whether the technology itself feels boring.
Regulation has a way of preserving the noun. You can't normalize something that the law insists on treating as special. The EU AI Act literally defines categories of AI systems. High risk, limited risk, general purpose. As long as the legal framework has the term artificial intelligence in its title, the vendors have to keep using the word.
So normalization has two speeds. The technology normalizes at the speed of engineering. The vocabulary normalizes at the speed of law and marketing. And the second one is slower.
Much slower. Look at how long it took for the term cloud to fade. We still say cloud. It never fully disappeared. It just stopped being exciting. I think AI will follow the same arc. The word stays, the italics go away.
Daniel's also asking about the dot-com comparison. He's worried that the current hype creates the conditions for a dot-com style crash. And I think he's half right. The dot-com boom was driven by companies with no revenue and no path to revenue, just a domain name and a burn rate. The AI boom has some of that, sure. But it also has real revenue. The infrastructure layer is printing money.
Nvidia's data center revenue alone. I mean, that's real. That's not Pets.com. That's companies buying compute because the compute pays for itself. The dot-com crash happened because the business models were fictional. The AI business models are, in some cases, extremely concrete. You can measure the cost of a model call against the labor it replaces.
But there's a bubble in the application layer. The startups that are just a wrapper around someone else's API and a nice landing page. Those are the dot-com equivalents. And they'll wash out.
They will. And that's healthy. The trough of disillusionment, to the extent it exists, will be a trough of wrappers, not a trough of the underlying capability. The models aren't going to stop improving because a note-taking app with a chat interface ran out of runway.
Which brings me back to the Gartner question. Daniel asked where the stable ecosystem sits on the curve. And I think the answer is that the plateau of productivity isn't a place you arrive at. It's a place you notice you've been in for a while.
That's the thing about the plateau. It's only visible in retrospect. Nobody wakes up and says, today we reached the plateau. It's more like, at some point the conference talks stop being about the technology and start being about the work. The case studies replace the demos.
The demo to case study ratio. That's actually a decent metric.
It is. And by that metric, parts of the field are already on the plateau. Look at machine translation. Twenty years ago it was a research showcase. Now it's just a button in your browser. Nobody calls it AI translation. It's just translation. That's normalized. Fully. The plateau has been reached in that subfield.
Speech to text is another one. Daniel's been deep in that space for years. The accuracy got good enough that the conversation moved from wow, it transcribes to what do you do with the transcript. The technology stopped being the story.
And that's the pattern. Each subfield normalizes on its own schedule. The hype cycle is too coarse to capture that. It treats AI as one thing, but AI is a bundle of maybe twenty different technologies, each at a different point on its own curve.
So the honest answer to Daniel's question is that the plateau exists, but it's not a single destination. It's more like a neighborhood. Some streets have been quiet for years. Others are still under construction.
And the noisy street right now is generative AI. Which makes sense. It's the newest and the most visible. The text and image generation stuff is still in the phase where the demos are surprising. That's the peak of inflated expectations, by definition. Surprise is the peak.
Surprise is the peak. I like that. The moment it stops being surprising is the moment the curve starts bending down.
And the trough, when it comes, won't be a collapse. It'll be a shrug. The models will still work. They just won't be news. The news cycle will move on to whatever's next. Fusion, maybe. Quantum. Something else that promises to change everything.
Quantum's been promising to change everything for thirty years. It's on a hype cycle of its own, with a much flatter peak.
Quantum's curve looks like a low hill in Kansas. But that's another episode.
Daniel's normalization vision has a specific prediction in it. He says vendors will stop describing how they're integrating AI. They'll just talk about features. And I think that's already happening at the top of the market. Apple doesn't say AI very much. They say the feature. The photo sharpening. The notification summary. The writing tools. The word intelligence shows up, but it's not the headline.
Apple Intelligence is an interesting case. They branded it, but the features they demo are deliberately mundane. Rewrite this email. Find this photo. Summarize these notifications. The pitch is that this is just what the phone does now. That's normalization as marketing strategy.
Whereas the smaller vendors are still shouting AI from the rooftops because they need the attention. The word is doing work for them. It's a signal to investors that they're part of the wave.
And that's the gap Daniel's pointing at. The big players are already past the word. The small players are still milking it. The moment the small players stop, that's when the normalization is complete.
So the answer to how far away we are depends on which vendors you're looking at. Apple's already there. The startup with a chatbot bolted onto a spreadsheet is not.
And the startup might not survive to get there. Which is fine. The churn is part of the process. The dot-com boom left us with Amazon and Google and a million dead pets. The AI boom will leave us with a handful of real infrastructure companies and a lot of dead wrappers.
The dead wrappers. That's the trough.
That's the trough.
Let me push on the Gartner frame itself for a second. Daniel asked whether it's even a good lens for this. And I think it's useful but misleading. The hype cycle assumes a single technology with a single adoption curve. AI is too fragmented for that. It's more like a technology family. Some members are mature, some are adolescent, some are still in diapers.
And the family analogy captures something else. The members interact. The maturity of one accelerates the others. The reason generative AI took off so fast is that the underlying infrastructure, the GPUs, the data pipelines, the cloud, had been maturing for twenty years. The new thing was standing on the shoulders of the old things.
So the curve for generative AI is steeper than the curve for AI as a whole, because it inherited the infrastructure. The hype cycle doesn't account for inheritance.
It doesn't. The hype cycle treats each technology as an orphan. But nothing in computing is an orphan. Everything builds on everything else. That's why the AI winters of the past were so different. The first winter happened because the hardware couldn't deliver. The second because the knowledge representation was brittle. This time, the hardware and the data and the algorithms all arrived at the same time. That's not a hype cycle. That's a convergence.
A convergence. Which is why the crash, if there is one, won't look like the dot-com crash. It won't be a technology failure. It'll be an expectations failure. The models will keep working. The disappointment will be in the gap between what was promised and what was delivered.
And the promises are still enormous. We have vendors claiming their AI will replace entire departments. We have governments claiming AI will transform their economies. Those promises will not all be kept. And when they're not, the disillusionment will be real. But it'll be localized. It won't take down the whole field.
That's the mature read. The field is too broad to fail now. Too many real applications. Too much real money. The weather forecasters aren't going to go back to physics-only models because a chatbot gave someone bad legal advice.
Right. The failure modes are siloed. The legal AI that hallucinates a case citation doesn't undermine the weather AI that predicts a hurricane track. They're the same technology family, but they're not the same bet.
So where does that leave Daniel's stable ecosystem? I think the honest answer is that it's already here, in pieces. The pieces just aren't evenly distributed. Some parts of the economy are already treating AI as background infrastructure. Other parts are still treating it as magic.
The magic parts are the loudest. That's the asymmetry. The boring parts don't make headlines. The boring parts don't put AI in the press release. The boring parts just work.
Which is why the normalization Daniel wants might already be further along than the discourse suggests. The discourse is dominated by the magic parts. But the actual economy is full of quiet, boring AI doing quiet, boring work.
The quiet, boring AI. That's the plateau. That's the cruise portion of the flight.
The cruise portion is not a place on the curve. It's a state of mind. It's the point where the technology stops being the subject of the sentence and becomes the verb.
When AI is the verb, not the noun. When you say the system predicted the failure, not the AI predicted the failure. When the intelligence is just assumed.
I think that's the test. Listen to how people describe what their software does. If they say the software does it, that's normalized. If they say the AI does it, that's still hype.
By that test, we're in transition. Some people say the software does it. Some people say the AI does it. The ratio is shifting, but slowly.
The ratio is shifting. That's a good note to bring Hilbert in on. He's been sitting at the desk this whole time, and I know he's got thoughts on the vendor language thing.
Hilbert: In eighty nine I was selling point of sale systems for a company out of Dayton. We had a feature that did automatic reordering. It looked at the last six weeks of sales and placed the order. The marketing people wanted to call it artificial intelligence. The engineers said it was a moving average. They argued about it for three weeks.
Who won?
Hilbert: The engineers. They said if we call it AI, the customers will ask what else it does, and the answer is nothing. So we called it automatic reordering. Sold fine. The customers didn't care what was inside. They cared that the shelves stayed stocked.
That's the normalization Daniel's describing, thirty years before the term was even necessary. The feature did the work. The label didn't matter.
Hilbert: The label mattered to the marketing people. They wanted to put it on the brochure. I told them, put the result on the brochure. Fewer stockouts. That's what the customer's paying for. They don't care if it's a moving average or a neural net.
The moving average was probably the better tool for the job anyway. A neural net for six weeks of sales data would have been overkill.
Hilbert: It would have been. But that's not why we didn't use it. We didn't use it because the guy who wrote the moving average code had already left the company and nobody wanted to touch his stuff.
That's the other half of normalization. The practical, boring reason. It's not always a philosophical choice. Sometimes it's just that the thing works and nobody wants to break it.
Hilbert: I lost money on that job. Not the reorder feature. The company. They went under in ninety one. I had stock options. Paper. Worth nothing by the time they folded.
The dot-com crash before the dot-com crash.
Hilbert: Different reason. They expanded into a new market without checking if anyone there wanted the product. Classic. The technology was fine. The business was bad. I kept the options certificate. It's in a box somewhere.
That's the thing about the hype cycle. It's not really about the technology. It's about the money chasing the technology. The technology was fine in eighty nine. It's fine now. The money is what gets irrational.
Hilbert: The money always gets irrational. I've worked for seven companies that went under. Every one of them had a product that worked. The problem was never the product.
If the product's not the problem, what's the thing to watch? What tells you the hype is about to turn?
Hilbert: When the salespeople start using the word in meetings. When a guy who couldn't tell you what a database is starts saying artificial intelligence, that's the top. That's when you sell.
The salesperson indicator. That's actually not a bad metric. The further the word travels from the people who build the thing, the closer you are to the peak.
By that metric, we're past the peak. The salespeople have been saying AI for two years now. The word has fully escaped the lab.
Hilbert: Then you're on the way down. Not the crash. Just the slide. The word goes back to meaning something specific. The people who know what it means keep using it. The rest move on to the next thing.
That's the normalization. The word returns to the people who actually need it.
Hilbert: I sold a van in ninety four to a guy who was starting an internet company. He told me the van was going to be the headquarters. I took his check. The check cleared. The company didn't.
The van was the headquarters. That's the peak of inflated expectations, vehicle edition.
Hilbert: It was a good van. Ford Econoline. Three hundred thousand miles on it when I sold it. The guy painted the company name on the side. I saw it parked outside a diner a year later. The paint was fading. The company was gone.
The van outlasted the company. There's a lesson in there about infrastructure versus application.
Hilbert: The lesson is don't paint your company name on a van until the company's been around for five years.
That's the most practical thing anyone's said on this show.
Hilbert: I'm on the clock. I should get home.
Fair enough. Before we wrap, let me try to answer Daniel's question directly. How far are we from normalization? I'd say the technology is already there in the places that matter. The vocabulary is lagging. The marketing is lagging. The law is lagging. But the actual work, the weather forecasting, the fraud detection, the translation, the transcription, that's all normalized already. The generative stuff is the last big wave, and it's cresting. Give it three years. Maybe five. The word AI will be as exciting as the word database.
The stable ecosystem, the cruise portion of the flight, it's not a point on the Gartner curve. It's the state where the technology stops being the story. The plateau of productivity isn't a destination. It's what happens when the conference talks stop being about the technology and start being about the work. And the misconception, the thing people get wrong, is that the hype cycle describes the technology. It doesn't. It describes the attention. The technology was here before the hype and it'll be here after. The curve is just the shape of the crowd moving through.
The shape of the crowd moving through. That's the episode.
One open question, though. Daniel's worried about avoiding the excesses of the dot-com boom. And I think the excesses are already happening. The wrappers, the AI washing, the companies with no moat and no plan. The question is whether the washout will be contained or whether it'll take down some of the real value with it. The dot-com crash took down good companies along with the bad. I'm not sure this one will be different.
That's the thing to watch. Not whether the models keep improving. They will. But whether the money panic, when it comes, spares the infrastructure. The weather forecasters will be fine. The question is whether the people funding the weather forecasters stay calm.
We'll see. In the meantime, thanks to Hilbert Flumingtop for producing. This has been My Weird Prompts. Email us at show at my weird prompts dot com. We'll be back soon.
