Remembering Dotcom, Pondering LLMs #
Comparing hypes and bubbles
May 2026
There's a lot of talk about an AI (or, rather, LLM) bubble these days, and comparisons to the dotcom boom 30-odd years ago are tempting. The general idea behind comparing AI and dotcom often seems to in defense of a likely AI bubble, making the argument that even if the dotcom bubble burst, the tech certainly prevailed and is today firmly embedded in almost every part of the global economy.
In one sense, it's a way of defending the current financial irrationality as a necessary means to an end, but depending on how it's structured, the argument often misses the target. One such poorly structured comparison I stumbled across recently was made in a presentation by business analyst Ben Evans (slide deck here):
Imagine asking "What will be changed by the internet?" in 1997.
Similar rhetoric pops up regularly when discussing AI hype. To people who vaguely remember the IT boom and bust at the turn of the millennium, but have no deeper interest in tech business, it might sound like a plausible conundrum. When coming from a professional business analyst, it sounds if not dishonest, then at least ill-informed: It's true that nobody in 1997 could give detailed predictions about the rise and fall of Myspace, but it's false that there wasn't enough information available for analysts to get a fairly clear picture of the way things were heading.
Things of the Future
Above is a scan of the (long since defunct) bi-weekly Swedish computer periodical Datormagazin, issue 20, 1994, page 9. Datormagazin's target audience was ordinary home computer geeks, and on a two-page spread, the magazine reported from the "IT Festival", a show held at various venues around Stockholm, including the National Museum of Science and Technology. It was open to the public and attracted 8,000 visitors.
Clockwise from the top left, the captions read:
- A teenage boy's room of the future will house a computer terminal, a video phone and an interactive TV [the article text describes it as a functional equivalent of a streaming service].
- Volvo displayed their latest technology for trip planning. By entering a destination, the driver can access the best route and avoid traffic jams.
- The display that attracted most visitors of all ages were SISU's [the Swedish Institute for Systems Development] liveboard screens. Adult or child, anyone could draw and write together with visitors at Electrum in Kista [another show venue].
- A number of video phones were installed at the museum and Cyber Sphere at the Stockholm House of Culture [another show venue]. Visitors were often more interested in seeing who they were talking to than they were in the actual conversation.
The usefulness and convenience of Volvo's GPS/satnav, augmented with live traffic information, was hard to argue with. However, the visitors might not have been as excited about the collaborative video conferencing displays, had they realized that 30 years later, they'd be sitting in endless Teams meetings, arguing about whose turn it was to hijack the shared Powerpoint presentation.
It's apparent that few people in 1994 envisioned that a teenage boy in 2026 would have his computer, video phone and interactive TV (streaming service) all in one relatively affordable handheld device. The concepts were clear, though, and the technology on display already existed and worked.
Loftier Visions
Incidentally, 1994 was also the year when Sun Microsystems released a commercial about an imaginary computer called Starfire. The Starfire has a capacitive touchscreen and an accompanying tablet computer with motion sensors and a built-in camera. Among its many software features we find video conferencing, collaborative document editing, digital document signing, AI-assisted photo and video editing, and instant scanning. The framework story of the commercial features a middle manager preparing a corporate presentation about an electric car.
In AT&T's vision of the future (predating Sun's by a year), we can see a lot of similar concepts taken even further, including VR headsets, tablet computers, video conferencing, online shopping and - interestingly - instant speech translation and AI agents. The last two certainly didn't feel as if they were just around the corner back then.
[
](pix/man_modem_panasonic.jpg)
Look at him. Look at his pinstripe sprezzatura, his devil-may-care pipe smoking, his subtly greying beard. Most of all, look at his 1982 Panasonic RL-H1400 handheld computer with a modem attachment. Wisdom, business savvy, style: Women want him, men want to be him. What if this could all be one small device, maybe with a slightly better screen? What then, dude? What then? *
Inside the Boom
I was there, Gandalf. I was there 3000 years ago (well, 30-ish at least), during the dotcom boom, as a pimply-faced junior developer, watching it all unfold from the inside.
Firstly, it was quite obviously a bubble (though a lot of fun while it lasted): There's clearly something fishy afoot when even a naïve 19 year old - having the time of his life in a cool job at a cool company - starts to quietly wonder if it's a good idea to hire people for yet another office a few blocks away, when people at the already existing ones apparently have no work to do, all while the stock is ticking up like there's no tomorrow.
Secondly, almost everything about the hype was correct. The net was going to change how we access information, communicate, advertise (sadly), how we shop, how we do banking, order food, book travels, how we consume media and news, and how we conduct business - and it did. Many of the detractors were right, too: our society is rife with digital rent extraction, social atomization, mental health issues and constant surveillance, much of it intimately linked to our ever-increasing online presence.
Use Cases Aplenty
The future is already here - it's just not very evenly distributed.
- William Gibson in 1993
Tech was very much a special interest back in the 1990s, and many people lacked the information needed to extrapolate correctly. To someone in the thick of it, however, the promises (albeit not the business practices) seemed perfectly reasonable.
In 1994, Unix vendor Santa Cruz Operation (SCO) and pizza vendor Pizza Hut launched PizzaNet. Through this geographically limited pilot project, customers could customize and order pizzas via the web and have it delivered to their door. The same year - also in a PR stunt orchestrated by SCO - the rock band Deth Specula performed what may well have been the first concert ever to be live streamed online. This broadcast was said to show "a possible future for music entertainment." Indeed.
In 1993, SGI released their "low cost" Indy workstation, a computer specifically intended for web authoring, and the first one to come with a webcam as standard. Two years later, the video chat software CU-SeeMe appeared: slow and jerky, but no longer the stuff of astronomically expensive Unix workstations. RealAudio launched the same year, providing solutions for streaming live sound. In fact, things like video conferencing and online business collaboration had been anticipated ever since Doug Engelbart's Mother of All Demos in 1968, hampered only by hardware costs and bandwidth limitations.
Around 1997, web chats, webmail clients and even web-based time reporting software had started appearing, indicating a path for both future enterprise and consumer software.
In December 1996, Swedish bank SEB launched one of the first consumer-facing Internet services. Their goal was attracting 10,000 users during the first year. By the end of 1997, a whopping 110,000 customers had tried their online bank.
However, just like with online shopping, news services or computer aided navigation, online financial services weren't actually new during the dotcom boom. It was just that the future was starting to get evenly distributed.
Internet Precedent
In 1980, France's national telecomms monopoly launched Minitel. It was a funny little computer terminal with a small, monochrome CRT screen and a preposterous-looking keyboard. By 1988, three million terminals had been installed - most of them in homes - and usage was booming. People of course loved using them for mail and chatting, but also for banking, shopping and news. The primitive machines couldn't handle sound or video, but for everything else, they were wildly popular. So popular, in fact, that French newspapers anticipated severe competition and loss of business. Their protests led to a deal in which they were offered the first consumer services on the system.
Pre-internet networks existed elsewhere too, such as CompuServe in the US. There were of course also individual bulletin board systems with distributed discussion groups via FidoNet. None of them managed to stir quite the same fervor as in France - probably because Minitel terminals were largely subsidized, given out as "free loans" - but the services offered were similar. The concerns of newspaper death could only be mitigated because of centralized network control; with the advent of the more anarchic Internet, the consequences for publishers were predictable.
Hence, based on existing, working tech and considering the pace of Moore's Law and steady price drops in technology and bandwidth during the 1990s, it was entirely possible to infer that yes, this Internet thing is likely something people will want to use, and that widespread usage would change things fundamentally - even the hows and whats, with broad strokes of the brush.
Desire-driven Future
We humans build with intent. We have a vision of what life should be like, and we strive to achieve it. Flying machines, industrialization, mechanized agriculture, space travel, modern medicine, high yield crops - just to pick a few. We've had both great and astonishing ideas and a remarkable capacity to turn them into reality. This is technology answering to deep-seated human desires that can be traced as far back as we have written sources: exploring space, conquering the skies, battling disease, minimizing hard manual labour. Sometimes we make a technological leap - such as with general-purpose digital computers - and our ideas and possible solutions crystallize into finer detail.
The net, once it was clear it was going to become cheap enough, became a vehicle for spreading already existing technology. Both that technology, and the Internet itself, had already been proven to work before the hype and bubble exploded. That's also why many of the broader predictions made during the 1990s turned out to be correct, and it's also why this AI hype seems much harder to make inferences (heh) from.
Almost Working
Just like with the foundational tech of the dotcom boom, AI - as in LLMs - isn't new. Artificial neural networks have existed for many decades, and it's now nearly a decade since Attention Is All You Need was published. Older still is the human idea of autonomous machines for convenient (and profitable) automation. In many aspects we've come a very long way with very simple means, but an industrial robot will never have the same appeal as KITT - the talking, thinking, joking and completely self-driving car in the (very) 1980s TV series Knight Rider.
KITT is a convenient example, because it represents those last bits of the future that haven't quite arrived yet. It offers not only completely autonomous individual transportation, but also a flawless natural language interface. It's capable of performing and learning complex tasks with perfect reproducibility, and has the agency and creativity to apply this in real world situations. As AT&T's vision of the future reminds us, we've dreamt of AI agents for a long time - it's just that in our dreams, they don't make up phony information, miss crucial facts or commit calamitous blunders if left unchecked.
Given what LLMs actually are, we've been able to take them astonishingly far. The problem with them is that, unlike the increased distribution of tried and tested tech during the dotcom boom, they are inherently flawed. Non-determinism and hallucinations are features, not bugs, which means they need constant human supervision. As such, their usefulness remains highly dependant on the gravity of a given task, and in cases where it's crucial that no information is lost or misconstrued, their efficacy is questionable at best.
The most apparent use case for LLMs seems to be as software development tools. The current specifics of this has been discussed at length elsewhere, but the dominating consensus so far seems to be that if wielded with prejudice, an LLM can be of great help for certain programming and debugging tasks. We've long held a strong desire for low-code tools, code generation, test automation and, well, automating basically everything that has to do with programming. Those who remember the early LLM buzz, before the COVID pandemic, will perhaps recall MIT's SketchAdapt from 2019: When it comes to software development, there seems to be a problem-solution fit. The specifics may yet have to crystallize, but if venturing a guess, this is one of the most apparent areas where LLM usage could stick. We don't necessarily have to like this, but few of us have influence over its adoption.
[
](pix/nokia7110_ad.jpg)
Dude! Dude! Let's buy a pinstripe suit and a pipe! WAP phones, like The Nokia 7110, were pushed quite heavily during 1999. The hardware and bandwidth were both comically underdimensioned and WAP flopped, but it wasn't hard to tell in which direction things were heading. *
The Datagubbe Hype And Bubble Index
Gartner is an American business analysis firm acting as a key player in generating buzzwords, perpetuating hype and fomenting investment bubbles. They're famous for filling C-suite heads with things like magic quadrants and hype cycles. They also make helpful predictions, for example this one from 2022 in which they insisted that in 2026 (which is now), 25% of all people would spend at least one hour per day in the metaverse, where they would partake in a virtual economy based on digital currencies and NFTs - impacting "every business that consumers interact with every day" and providing employers a "connection to their employees through immersive workspaces in virtual offices". Gartner advises CEOs, investment firms and government agencies, employs 20,000 people and made a US$729 million profit in 2025.
If Gartner can make up pseudoscientific analysis tools and preposterous predictions, so can anyone. I'd like to propose a classification for tech hypes called DHABI - the Datagubbe Hype And Bubble Index. (I also considered calling it DUBAI - the Datagubbe Universal Bubble Assessment Index. Of course, it's not really an index, much like how Gartner's hype cycle isn't really a cycle. But I digress.) DHABI consists of the following categories, carefully picked to capture the essence of tech bubbles: #
Tech Soundness: Does the tech do what it says on the tin? - Popular Expectations: How much do ordinary people (as opposed to S&P 500 CEOs) yearn for the supposedly tech-related life improvements on offer? - Precedent: Have there been previous, comparable attempts? - Real World Applicability: Is the tech capable of living up to Expectations and Demand, and can it do so profitably? - Psychology and Business: What zeitgeist and business practices are affecting the bubble?
Let's apply DHABI to a few famous tech hypes:
Dotcom: Solid tech, huge expectations and demand, solid precedent and solid real world application. The bad business practices (such as IPOs of unprofitable startups) were mostly fuelled by overestimating and overselling adoption speed and time to profitability, even for successful business ideas.
Blockchain/Crypto: Solid tech, low expectations and demand, little precedent - although trying to replace foundational societal concepts (like currency) that already transfer to the digital world are often met with a shrug, see Swatch Beats. The overall usefulness was overestimated and the real bubble was largely driven by scams and get-rich-quick schemes, which was more than apparent already in the early days of NFTs.
Metaverse: Immature tech (annoying headwear!), high expectations but poor delivery, poorly performing precedents, poor real world application. More of a short-lived hype than a bubble, this was a dud for not considering that actual video conferencing and screen sharing already worked well, ignoring sparse adoption of earlier attempts like Worlds Chat and Second Life, and the fact that VR technology had existed for a long time - including targeted at business settings - but never been adopted as anything but a fun novelty for casual gaming. People want The Matrix (except fun), but are getting dizziness and eyestrain.
This leaves us at the current stage of the AI (or, rather, LLM) hype: Dubious/immature tech (hallucinations, non-determinism), huge expectations on personal benefit mixed with slop fatigue and a largely unenthusiastic workforce (indicating expectation mismatch), inconclusive or lacking precedent (E.G. expert systems flopped, but this is something else), and currently sparse real world application (machine translation, software development). The business practices surrounding it look troublesome indeed, with circular investments, possible overcommitment to expensive infrastructure, forced adoption and dubious profitability at current pricing models.
[
](pix/virtual.jpg)
Virtual Reality gaming in the early 1990s: essentially the same now, except with fewer polygons.
Finally
Hindsight is 20/20, but that's also why it's silly to pretend that people in the 1990s had no clue about what the Internet would change. The promises of 1997 were almost all realized by 2007 (the year both Netflix streaming and the iPhone launched). The one thing we weren't as eager to accept were predictions about the predatory business models effectively dampening much of tech enthusiasm.
Indeed, something has decidedly shifted in the tech world: There was a lot of hot air and market irrationality during the 1990s, but the overall mood was one of fun, hope and optimism. Today, the LLM hype is spurred by CEOs of frontier model companies making regular doomsday predictions about how their own tech will displace workers, increase surveillance, wage wars and eventually replace human thought and creativity. The net was supposed to bring the world closer together (and in many ways, it did), but AI seems to be about pushing us apart, into little sycophantic chatbot bubbles of hyper-individualistic atomization.
As a technology, LLMs are here to stay in some form, probably as coding tools. Still, the bubble surrounding them has a different shape than that of the dotcom boom: We want Knight Rider's KITT, but that's not quite what we're getting. Even in the case of programming, there's a huge gap in expectations mostly manifesting itself between management and developers.
All of this is pure speculation, of course, and I might be utterly wrong, which would be nice in a sense: A bubble isn't desirable and it's never fun when it bursts (even if the tech prevails), as it usually hits workers and small-time investors/savers the hardest.
Then again, I'm not a high profile professional business analyst, so there's always a chance I could be right.