Why Isn't Everything Different Yet?

A patient explanation of adoption lag for impatient people, from someone who is famously impatient.

“Stochastic parrot” has entered my personal lexicon, but not the way the people who coined it intended. The people who coined the term meant that LLMs can’t perform anything truly novel because they are just ramified Markov processes, repeating random bits of their training corpus. To me, “stochastic parrot” now means: an argument so embarrassingly out-of-touch with reality that you wonder just how many lead paint chips the person making it ate as a toddler.

At this point, the catalogue of bad AI arguments iswell-establis hed: stochastic parrots, “AI can never truly understand,” water use estimates off by a factor of 1000, the occasional “it’s just autocomplete” said with the breezy confidence of someone who hasn’t had to code up an autocomplete system four times in their career. There’s a thriving cottage industry of goalpost relocation around AI, as the lag between confident predictions of “AI will never do X” and headlines reporting that “AI does X” gets progressively shorter. The goalposts are now somewhere in the parking lot, possibly in another ZIP code.

My audience is certainly too smart (and surprisingly handsome) to make those sorts of arguments but there are anti-AI arguments that I take seriously. One I’ve been hearing a lot recently is of the form:

“Okay, but if this thing is actually going to change everything — why hasn’t it changed everything yet? I’ve been waiting since the ChatGPT press cycle and the economy hasn’t been completely re-engineered yet!”

This is actually a fair question. It comes from the grand American tradition of borderline-psychotic impatience, a tradition I respect deeply and adhere to religiously. The question deserves a real answer rather than the usual hand-wavy “large transformations take time” non-answer that accomplishes nothing except making the speaker feel like some sort of Zen-ass master.

The technology arriving is not the same as the technology being used

When electricity became commercially available, you know what most factories did? They replaced their steam engine with an electric motor. One motor. In the same spot the steam engine was. Driving the same central driveshaft that spun all the same belts and pulleys to all the same machines.

This seems stupid in retrospect, but it was also completely rational. The factory was built around a central driveshaft. The workers knew how to operate the machines. The safety codes (such as they were) were written for the machines. The insurance was priced for the machines. You don’t rip out your entire capital plant because a new thing arrived.

It took about thirty years before factories were redesigned from the ground up around the idea that every machine could have its own motor. That’s when the productivity revolution happened. Not when the generator was invented. Not when the first electric motor was sold. Thirty years.

By my reckoning, we’re at forty months.

Software is not a breakfast food

Here’s something worth noticing: the first place AI is visibly hitting hard is software development. There are plenty of good reasons for it that I’ve covered previously. It also explains why the broader transformation feels slower than the hype would suggest.

Nobody wakes up in the morning craving a heaping bowl of software. Software is not a primary consumption product. It is infrastructure — the thing that makes other things better. You use it to run your hospital, manage your supply chain, file your taxes, book your flight, or decide which ad to show which human at which millisecond. Software is electricity, not the toaster, and certainly not the toast.

Which means that for AI to transform anything by this point, one of two things has to happen: either someone figures out how to embed AI directly into that thing (your hospital, your supply chain, your tax software), or AI makes software development itself faster/cheaper/better — which then propagates improvements into everything software touches, which is more-or-less everything.

Both are happening. But for each of the millions of domains where software currently operates, someone (an actual real, nameable person who has other stuff to do) has to do the work of figuring out specifically what AI changes and how. That’s not one problem. That’s millions of problems, being worked in parallel by people who also have day jobs and PTA meetings to go to and occasionally just get stuff wrong. It’s basically what about half of the entire population of software developers on the planet has been painfully retasked to accomplished over the last forty months, while our industry is being broken down and rebuilt from the ground up. We’re working as fast as we can. You’re welcome.

The checklist nobody talks about

Here’s roughly what has to happen before a transformative technology actually transforms anything at scale. I’m going to use “AI” as the example, but this list would have applied equally to electricity, automobiles, the internet, or indoor plumbing.

• The tech has to exist. Check. Moving on.

• The tech has to be reliable enough to trust. This takes longer than people expect. When technologies are in their early stages, we aim to make the “happy path” work before hardening it. Early automobiles required a dedicated mechanic. Early internet connections required a dedicated IT person to restart them. Early electrical connections probably required Thomas Edison himself to occasionally fix things. AI is currently in the “sometimes it happily and confidently lies to your face” phase. Most industries consider this a known-issue-to-be-managed rather than a dealbreaker, but it still slows deployment and occasionally makes us reconsider our life choices.

• Infrastructure has to be built. Not metaphorically. Literally. Data centers, power plants, chip fabs, fiber. This takes years and billions of dollars and permitting approval from people who also think stochastic parrots are real, and a disturbingly important amount of it is happening within the missile range of the entire Chinese navy. You can’t imagine how happy I am that all of this is not my problem.

• Workflows have to be redesigned. Not “add AI button to existing web page.” Redesigned. Tore up from the floor up. This requires figuring out what many workflows were actually supposed to accomplish in the first place, which turns out to be whimsically unclear in many organizations. Adding an AI chatbot button is basically just software developers tapping the microphone and saying “Is this thing on?”. It’s nowhere near the real payoff.

• People have to be trained. Not just to use a tool, but to develop judgment about when not to trust the tool. This is hard. People are bad at calibrating trust. This is the entire reason we have a word for “overconfidence.”

• Legal and regulatory frameworks have to be written. Who owns AI-generated output? What’s the liability when an AI-assisted decision hurts someone? What disclosures are required? Nobody actually knows yet. Jurisdictions are figuring this out at different speeds in different directions. Many of them are trying out directions that I think are deeply stupid and dangerous, but honestly that’s to be expected. Props to them for at least starting the process, fewer props for the “stupid and dangerous” bit.

• Business models have to be invented. Some things that become possible with AI aren’t just better versions of existing things — they’re things that didn’t exist before and have no pricing, no sales motion, no established market. Building those from scratch takes time even when the underlying technology works perfectly. A lot of these come under the rubric of “marketing”, work that is both slow, annoyingly probabilistic, and sadly utterly necessary. “Demand generation” sounds like an oxymoron until you find yourself working in a startup that doesn’t have enough of it.

• Competitors have to force each other to adopt. This is how most corporate technology adoption actually happens. You don’t upgrade because it’s smart. You upgrade because your competitor upgraded and suddenly your customers are asking uncomfortable questions. Adlai Stevenson famously said about politics: “When I feel the heat, I see the light”. Same energy. For many, change doesn’t come before discomfort, and discomfort only comes when your opponents change.

Each item on this list is a multi-year project. We as a species are running most of them in parallel, which is impressive and a sign that things are moving fast by historical standards. But they all have to mostly complete before you see the kind of macro-level numbers that would make a journalist write “AI has transformed the economy.”

Fast by historical standards is still slow by Tuesday standards

The internet became “commercially available” around 1991. Most people consider it to have genuinely transformed commerce sometime around 1999–2001. That’s a decade, and the internet didn’t require nearly as much retraining for every knowledge worker (and there were a lot fewer of them back then), didn’t require building much physical infrastructure beyond laying some fiber, and didn’t require resolving novel questions about who is responsible when the internet makes a mistake about your medical situation.

For reference: it took about 40 years after the invention of the steam engine for Britain to see meaningful GDP effects from industrialization. We are operating at about 5-10x the historic speed of technology-driven economic transformation. This does not feel fast. This is because we are impatient. Both things are true.

So: where are we? The technology exists and is impressive. The infrastructure buildout is underway and massive. Workflows are being redesigned in early-adopter organizations, often via guesswork. We’ve got one (1) product area (software development agents) where we’re past “early adopter” and moving onto mass-market. Legal frameworks are being written badly by people who have never used the technology, which is traditional. Business models are being discovered by trial and error, also traditional. Fortunes are being made and lost, another time-honored tradition.

The critics who say nothing has changed are measuring at the wrong resolution. The critics who say change should have been instantaneous have a broken model of how change works. The honest answer is: this is going extremely fast, it will often feel slow until suddenly it doesn’t, and the people who have built understanding now will not be scrambling in three years.

---

This post was constructed as always with the able assistance of Claude Sonnet 4.5. Eventually Claude and his kind will change everything, but this morning he got my name wrong, and wow was he snippy when he read this afterword.