Neuvottelija.com

2026-07-15 · AI Lab · Insight

AI and the Productivity Explosion: Who Actually Captures the Gains

Three conversations, recorded months apart, keep circling the same unresolved question: does AI make people better at the jobs they already have, or does it quietly do the jobs instead — and either way, who gets to keep the gain? I want to lay out what my guests actually said, not what the AI hype cycle wants them to have said. Some of it is a forecast, some of it is a personal testimony, and some of it is a deliberately extreme scenario built to force a conversation. I’ll try to keep those three categories separate.

”If you call AI a tool, you’ve already lost”

Start with the claim that set the terms for everything else. In Mikko Alasaarela Tekoäly räjäyttää tuottavuuden, Mikko Alasaarela argues that the whole augmentation-versus-replacement framing depends on how you conceive of AI in the first place:

“Your first mistake is that if you call AI a tool, then you’ve already lost. You’ve lost because if you think AI is an extension of a human, then it can only marginally improve that person’s productivity. But if you think of AI as an essential part of your entire ecosystem that does a big part of the jobs we used to do, then you’re able to multiply that productivity, because you’re not making the human a bottleneck for AI’s ability to execute.” — 00:00

From that framing he builds a specific, and specifically speculative, national scenario: if Finns collectively achieve a three-fold productivity jump by going “AI-native,” and half the population works, that jump translates into one-and-a-half times current GDP — “which in this current world situation would be enough to get pretty close to first place in the world.” He extends the same logic to the public sector: if its productivity also triples, the implied question is whether Finland needs to keep employing three times as many people there as the new productivity level requires. This is Alasaarela’s own arithmetic and his own hypothetical — a scenario built on an assumed multiplier, not a measured one. It is also the single boldest productivity claim across all three conversations, and it deserves to be read as a thought experiment about what full-scale adoption could theoretically unlock, not as a forecast anyone is committing to.

The pace problem

The scenario only makes sense against the backdrop Alasaarela lays out in the full episode, 200IQ Tekoäly ohi ihmisen | Alasaarela | #neuvottelija 307. He describes spending his Christmas break testing every frontier AI model he could get his hands on, burning roughly 700 euros in API credits, and coming away convinced that the pace of capability growth is outrunning institutions’ ability to respond. His benchmark evidence: OpenAI’s o3 scored 13% on the “Humanity’s Last Exam” test when it launched, and internally the same lab had already reached 35% within about a month. He cites Dario Amodei of Anthropic and Sam Altman of OpenAI both pointing, per current mathematical modeling, to an AGI moment in 2026. Context windows have been expanding fast too — Google’s Gemini Advanced at two million tokens, and what he describes as the largest model at the time, China’s MiniMax, at a ten million token window. And he flags a stranger, adjacent trend: by his account, roughly two-thirds of internet content and traffic was already AI-generated bot activity at the time of recording, with forecasts pointing toward 90% by year’s end — a dynamic he says is already reshaping what search and the open web are even for.

He is candid that the money backing this race is wildly asymmetric: the US committing on the order of $500 billion to AI infrastructure over several years, against a roughly €56 million European AI program announced around the same time — by his math, a 10,000-times gap that he says leaves him seeing “no chance for Europe in this race.”

What changes at the individual level

The more grounded part of the same conversation is about what actually shifts in a person’s day-to-day work. Alasaarela describes his own coding practice moving from writing code to building “pseudocode” architectures and specifications, then letting AI generate and test the implementation — because once you’re producing hundreds of thousands of lines, no single person can review it all, so the human role shifts to architecture, goal-setting, and verification. He frames this as a new professional skill:

“Isn’t it the most sensible strategy for you to take the role where you build those processes and frameworks and models through which you can use the artificial to get more out of it than your own brain would allow, and this is, in my opinion, the new mindset we are in […] absolutely insane productivity leaps will come for anyone once they just learn to use models like that.” — 41:54

He backs the claim with an anecdote about acquaintances who are CEOs discussing candidates for a technical leadership role:

“I have acquaintances who are CEOs […] I was looking for, say, a CTO or someone, and then there were these candidates, and only one of them was an AI native. You can guess who I continued the conversation with.” — 44:40

That’s an anecdote about one hiring decision, not a labor-market statistic — but it’s the closest thing to a hiring-market signal either episode offers, and it points toward augmentation being a filter on employability even before it becomes a filter on jobs existing at all.

The €84,000 backpack

The macro stakes get sharper in Pääoma hylkää Suomen | Hollmen Kotamäki | #neuvottelija 365. Mauri Kotamäki, chief economist at Finnvera, walks through his own calculation: with roughly €159 billion in Finnish public spending against about 1.9 million market-economy workers, that works out to a burden of around €84,000 per private-sector employee per year — and he estimates that, left on its current trajectory with aging and healthcare costs, it could climb past €110,000 fairly quickly. Both he and Markus Hollmen are explicit that this figure is a rough, illustrative construct that mixes borrowing, municipal, state, and pension spending, not a clean per-person tax bill — but they use it as shorthand for the scale of the load an aging, small economy carries.

Against that backdrop, I said on air that AI has personally changed my own work:

“I myself, for example, have found AI to improve my own productivity, and I genuinely experience it as a kind of salvation for me — I’m considerably more productive doing all sorts of things with that AI buddy of mine.” — 21:40

Kotamäki’s response to that was measured rather than triumphant: it’s genuinely unclear whether AI, as a general-purpose technology, will be easier for older workers to adopt than past technologies were — research suggests older cohorts adopt new technology more slowly — and if adoption friction bites, the productivity relief he’s hoping for won’t materialize on schedule.

Who captures the gains

The episode’s sharpest turn is about distribution, not aggregate output. Hollmen raises a scenario from a former employer’s wealth-management research: that by 2030, a third of the global workforce could be rendered jobless by AI — a number, he says, banks use to try to justify the trillions being poured into AI infrastructure, since that scale of investment needs a correspondingly large efficiency gain to pencil out. He puts the distributional question directly:

“Of course that could be good for whoever captures that efficiency, but what does it mean for national economies if a third of the population ends up unemployed?” — 53:57

Kotamäki pushes back on the number itself while defending the exercise:

“I myself don’t find it credible that within five years, 30 percent of people would end up unemployed […] but I think these calculations are nevertheless useful to go through — these doomsday calculations — because it forces us to think through the effects of AI, what the mechanisms could be.” — 55:18

Hollmen then floats a genuinely counterintuitive angle for a small, capital-poor economy: since the US is absorbing the sunk cost of the AI buildout — Nvidia hardware, data centers, training runs running into the trillions — countries like Finland may get to ride the resulting productivity improvement without having funded the infrastructure, “possibly” turning being poor into an accidental hedge. He also notes, as a live data point rather than a prediction, that GPU rental prices had already fallen roughly 30% over the prior year, suggesting the “critical, unobtainable resource” framing of compute may not hold. Elsewhere in the same conversation, the UBS–Credit Suisse merger is used as a concrete, present-tense example of the mechanism at work: combined back-office, IT, legal and compliance functions squeezed further with AI, meaning fewer jobs in the merged bank even as private banking outside the monolith is expected to grow.

Where I land

On air, I’ve put myself on the augmentation side of this personally: I’ve said AI has made me considerably more productive and that I genuinely experience it as “a kind of salvation” in my own work. I’ve also argued that delegating to AI belongs alongside rhetoric, negotiation and leadership as a fourth core skill — building on my own “3D model” of delete, delegate, do — because not being AI-native is, in my words, “a pretty bad idea” if you want to stay relevant in the current job market. At the same time, when Hollmen raised the one-third-unemployment scenario, I was the one who pushed the distributional question back at the table: who captures that efficiency, and if a shocking share of the EU workforce loses work to AI, who pays the taxes and the pensions from there. I don’t think those two positions are in tension. Getting AI-native is, on my own account, the individual move that matters most right now — but that doesn’t resolve who wins at the level of the whole economy, and on that larger question, based on what’s been said in front of me, I don’t think anyone at this table actually knows yet.

Source episodes

Every claim in this essay is grounded in the following episodes; quotes carry timestamps linking to the original video. English subtitles and full transcripts are on each episode page.

← AI Lab · Markdown mirror: index.md