Lila Soto: Hugo, hey — did you get through that prediction audit before today?
Hugo Vance: I did. Read it twice, actually. The number stopped me.
Lila Soto: Eighty-seven percent. Out of 214 specific, dated predictions — 2014 to 2024 — eighty-seven percent wrong. And I keep sitting with that because these weren't random Twitter takes. These were calibrated forecasters. So, mm, that's what we're getting into today — why tech prediction keeps failing at that scale, and whether the history of how technology actually spreads gives us anything better.
Hugo Vance: Yes. And the failure mode is always the same one. Not capability. Infrastructure. The assumption that society reorganizes at the pace capability advances — that is the recurring delusion.
Lila Soto: Okay but — and this is my actual question — if that pattern is that durable and that visible, why does it keep fooling smart people? Rodney Brooks, MIT robotics, iRobot co-founder, has been publicly scoring his own 2018 predictions every year since. His eighth annual scoring came out January 1st, 2026. Forecasts he explicitly framed as pessimistic — still too optimistic. So pattern literacy isn't the cure, or it would be working. What's underneath that?
Hugo Vance: Well. The pattern is legible in retrospect and invisible in the moment. That's — I'd say that's almost the definition of the problem. Electrification took fifty years to reach half of American factories. Half. And today's forecast says AI reshapes labor markets in three years. That's not a prediction about technology. It's a prediction about regulatory speed, organizational retraining speed — and those clocks haven't compressed.
Lila Soto: Right — but the part that doesn't fit is Brooks himself knowing the history and still overshooting. So what's the smaller question hiding inside yours? Is it that the pattern stays invisible, or that we're choosing not to see it because there's money in the hype?
Hugo Vance: Both, I think. And they reinforce each other in ways that are — you see, that's precisely what makes this durable.
Lila Soto: And that's — okay, that's what makes Everett Rogers so useful here, I think. Because what he actually did in 1962 was pull together 508 diffusion studies — anthropology, rural sociology, medicine, education — and say: here's the shape that keeps showing up. Slow start, fast middle, slow end. The S-curve. Every time.
Hugo Vance: Five hundred and eight studies. That's not a theory. That's a census.
Lila Soto: Right — and the plain-language version is just: every new thing starts as a rich person's toy. Then your neighbor has one. Then it's everywhere. Then your grandmother finally gets one. That arc — slow, fast, slow again — Rogers named five actual slices of it. Innovators, about two and a half percent. Early adopters, thirteen and a half. Then the early majority and late majority, thirty-four percent each. And then laggards, sixteen percent. It's basically a normal distribution laid on its side.
Hugo Vance: Yes. And Rogers built on George Beal and Joe Bohlen — rural sociologists, late 1950s, studying how farmers took up new agricultural practices. That lineage matters. The original data wasn't about gadgets. It was about whether a family planted a different seed variety.
Lila Soto: Which is — mm, that's a different kind of adoption cost than downloading an app.
Hugo Vance: Exactly the problem. Because the S-curve describes the shape beautifully. What it cannot tell you — and this is where I'd push back on anyone using it as a forecasting tool — is where you are on it. The lag data alone spans seven years for the internet to a hundred and twenty-one years for steam and motor ships. That range is so wide it almost invalidates the clock.
Lila Soto: Wait — a hundred and twenty-one years?
Hugo Vance: Steam and motor ships. A hundred and twenty-one years from invention to widespread productive deployment. So knowing the curve exists — knowing it will go slow, fast, slow — tells you nothing about whether you're living in a seven-year story or a century-long one.
Lila Soto: So knowing the shape of the curve doesn't tell you where you are on it — isn't that the whole problem? Like, the map is real but it has no 'you are here.'
Hugo Vance: The map is real and it has no 'you are here' — yes, but here's what breaks the clean version of that. Joel Mokyr, working through 1750 to 1914, found it wasn't just slowness. He found *specific* brakes. Skilled-labor shortages. Institutional misalignment. Infrastructure gaps. Not philosophical fog — actual named obstacles. Which means the lag isn't random. It's structural. And that structure hasn't dissolved.
Lila Soto: Okay but — and I want to press on this — AI tools already live inside existing digital infrastructure. People carry them in their pockets. So isn't the infrastructure prerequisite actually weaker this time? Like, the rails are already there.
Hugo Vance: That's the category error. The brakes Mokyr documented were never about whether the device existed.
Lila Soto: Wait — say that again.
Hugo Vance: The steam engine was operationally viable. The electricity grid got built. The brakes came from organizational redesign, labor retraining, regulatory alignment. Those clocks — I mean, picture it concretely. A manager in 2024, she's got AI running in her morning workflow, genuinely faster, real gain. But her team still runs on eighteen-month planning cycles. Her company still hires on credentials from five years ago. Her regulator still moves like molasses. The tool arrived. The world did not restructure.
Lila Soto: That's exactly — mm, that's Mokyr's brakes, but in a cubicle.
Lila Soto: Which means we might be measuring consumer trial adoption and calling it transformation. Those are different things running on different clocks.
Hugo Vance: General Purpose Technologies — steam, electricity, computing — they diffuse slowly precisely because each sector has to build its own complementary adaptations. The breadth is the reason for the slowness. That hasn't changed.
Lila Soto: And whether the compression from sixteen years to seven to five actually extends to societal restructuring — or only to surface tool usage — that's the question Carlota Perez's installation and deployment model sits right at the center of. That's the part we haven't gotten to yet.
Hugo Vance: And that's — yes, that's the precise place Perez cuts through. Personal computers, sixteen years to fifty percent penetration. Internet, seven. Smartphones, five. AI tools projected at three. The compression is real. I'm not disputing the data. But Perez's model says: that fast diffusion, that speculation, that's the Installation Phase. Financial capital floods in, the bubble builds, the crash comes. And the productive restructuring — the Deployment Phase — that follows the crash. Not the hype.
Lila Soto: So the clock is faster but the phases are the same.
Hugo Vance: The clock is faster. The sequence — installation, crash, deployment — Perez ties that to Kondratiev's long waves, the fifty-to-sixty year cycles. The trough between phases is often a major economic depression. That structure hasn't been abolished by faster diffusion.
Lila Soto: Okay, I'll — yeah, I'll give you that. But what's the smallest version of where I'm still not sure? It's this: the compression from sixteen to seven to five — that's not just enthusiasm. It's cumulative infrastructure, digital connectivity eliminating geographic constraints. Isn't it possible the restructuring lag is also compressing, just slower than the diffusion curve?
Hugo Vance: That would have to show up in the productive deployment data. And it doesn't — not yet. Picture a freight logistics firm in 2023, ships AI routing across its entire fleet in six months. Genuine adoption. But the insurance frameworks for autonomous routing decisions, the liability law, the retraining of dispatchers — those are still running on pre-AI timelines. The tool diffused in months. The institutional scaffolding is a decade behind.
Lila Soto: That's Mokyr's brakes again, dressed differently.
Hugo Vance: It is. And that, I think, is the honest concession I can make here — the Installation Phase runs faster now. Perez would recognize it. But the trough still comes. And what sits on the other side of that trough is the actual productivity gain. We're measuring the fast part and calling it transformation.
Lila Soto: Which means the headline number — three years to fifty percent AI penetration — is real and kind of beside the point at the same time.
Hugo Vance: Beside the point, yes. The adoption curve compressed. The restructuring lag — well, we don't yet have evidence it has. And that gap is where the next hundred years of prediction errors are already being written.
Lila Soto: Here's what I still can't answer after all of this — and I'm genuinely not sure you can either. If the structural brakes on electrification took fifty years to work through American factories even after the technology was proven, and Brooks found in January 2026 that even pessimists were still too optimistic about social change — are we actually in a different era, or are we just in 2005 marveling at broadband while not yet realizing the real restructuring takes twenty more years? I mean, the history doesn't give us a clean answer. I keep turning it over and I don't find one.
Hugo Vance: No. I don't think it does.
Lila Soto: Yeah. That's kind of where we are.
Hugo Vance: Good question to be left holding, at least.
Lila Soto: mm. Good one to have had the company on.
