---
title: "The AI Storm Warning — Sami Miettinen on Paasipodi #61"
titleOriginal: "Tekoälyn myrskyvaroitus | #paasipodi 61"
description: "Sami Miettinen's fourth visit to Martin Paasi's Paasipodi is a practical field report from inside the tools. An investment banker who wrote Commodore 64 assembly as a teenager and then left coding for thirty years walks through what a modern AI development environment actually looks like, what it costs, how to start from zero, and why he ends the episode by issuing an official AI storm warning dated 19 February 2026."
format: "podcast"
show: "#paasipodi"
episode: "#61"
hosts: ["Martin Paasi"]
date: 2026-02-19
duration: "1:10:46"
language: "fi"
original: https://www.youtube.com/watch?v=QsV6_rHdIic
canonical: https://www.neuvottelija.com/media/paasipodi-61-ai-storm-warning/
---

# The AI Storm Warning — Sami Miettinen on Paasipodi #61

**Martin Paasi** opens Sami Miettinen's fourth appearance on Paasipodi with a warning that
this one is different: not a discussion of markets or monetary policy, but a hands-on report
from someone who has spent the past year living inside AI development tools. The episode ends
with what Paasi calls the podcast's first-ever homework assignment — and with an explicit
storm warning.

**Thirty years away from the keyboard.** Miettinen wrote Commodore 64 assembly as a teenager
and studied computer science at Tampere University of Technology before moving to the Helsinki
School of Economics. Meeting investment bankers ended the coding: for three decades the closest
he came was an Excel formula. ChatGPT 3.5 brought him back to the edge of it, and vibe coding
brought him the rest of the way over the summer.

**What the environment actually looks like.** He walks Paasi through the three panels of a
modern AI-assisted IDE — the Git repositories on one side, a bash terminal in the middle, and
the AI prompt at the bottom — using **Google Antigravity**, which came out of Google's
acquisition of Windsurf, as his working example rather than VS Code. The code runs on a cloud
provider: AWS, Azure, Firebase or the Finnish UpCloud. His own token spend runs at roughly
€300 a month.

**The bill is the beginner's real risk.** A cloud function that a few enthusiastic friends
start hammering can produce a five-figure invoice, and Miettinen is blunt about it: do not
leave an open payment commitment running against an app you built for fun. Set the spending
caps before you need them.

**What he has built.** A bot he calls **Samantha**, running on OpenClaw at roughly one euro per
task. A public repository — the *villapaita ekonomisti* project — that pulls data from FRED and
Eurostat; Statistics Finland's API, he notes, is considerably more awkward to work with. And
**Consiglieri**, a second-brain application running natively on his own iPhone through Xcode,
with a Supabase database behind both the phone and the browser versions, an idea-enrichment
feature, and a receipt scanner. He has not yet wired the receipts through to his accounting
software, which he concedes would be an entertaining afternoon's work.

**Where an ordinary person should start.** Not Antigravity. Miettinen's recommendation for
anyone curious is the Swedish company **Lovable**, valued at some seven billion, precisely
because it strips the three panels down to one: a text box you write plain Finnish into. It
makes the hard decisions for you — it picks a secure, authenticated, white-labelled Supabase
database, builds the front end in React with HTML and CSS, and handles the backend and backups.
He used it heavily enough last year to be among the top five per cent of Finnish users. He no
longer needs the training wheels, but that is where the road starts.

**The stubbornness problem, and the fix.** Sometimes a model settles into a conviction that a
thing must be done a particular wrong way, and no amount of explanation moves it. The remedy
is architectural: press the GitHub button, pull the code into a repository of your own — it is
your code, Lovable does not take it — and then point a *second* AI at that repository and ask
it to explain why the first one is wrong. Two models arguing produce a better answer than one
model agreeing with itself.

**A worked example.** Paasi asks what this looks like for someone who is not a coder, and
Miettinen builds him a one-man Jeep repair shop: an application that identifies a broken part
from a description or a manufacturer's code, searches worldwide for stock, and ranks results
by price and distance with a phone number beside each. Then he supplies the friction — the
sites holding that parts data do not want to give it away, because their business model depends
on humans clicking through their own interface. The technical problem is easy; the commercial
one is not.

**Why this is no longer a coder's story.** Miettinen's argument is that the thing has leaked
out of the engineering department and into the office. Two weeks before recording, Anthropic
published a **Skills** library — not a new foundation model, just plain text describing how a
lawyer actually works through a case. The share prices of legal-information companies fell
sharply. There is a financial analyst skill in the same library, and **Claude in Excel** and
**Claude in PowerPoint** do, in his view, what Microsoft should have shipped a year earlier.
You can now ask a spreadsheet for an LBO model, a DCF, and a comparable companies and
comparable transactions set — six months of a Goldman Sachs intern's twenty-three-hour days,
produced in an afternoon.

He offers a second example from the same fortnight: the Epstein files. Some three and a half
million pages, released as images. A private individual, on their own machine, put an AI
through the entire set and produced fifteen pages of findings with citations back to the
source pages. In January that was not a routine task. Now it is.

**Scale, and who is operating at it.** Miettinen's list of frontier labs puts Anthropic first,
followed by Google's Gemini, OpenAI, a Perplexity-or-Bezos tier, and Grok. His point about
Elon Musk is not about Twitter: it is that a man with ten million tokens of inference running
continuously is playing a fundamentally different strategic game — running simulations of a
market position that a competitor computes once a quarter — and is now building inference
capacity in orbit, where the energy is cheaper.

**The political consequence.** Miettinen's Samantha bot costs about a euro a task. Finland can
levy VAT on that euro. What it cannot do is tax the work the bot performs downstream, and he
argues that raising taxes on human labour and human capital is becoming arithmetically
impossible rather than merely unpopular. His request to politicians of every stripe is
practical rather than ideological: before legislating on this, spend two hours with someone
competent and actually use the tools. Meanwhile Europe, in his reading, is on its knees —
occupied with directives while the rest of the world moves to a different galaxy every day.

**The analyst's career, rewritten.** He and Paasi trace the arc they both know: the good school,
the good grades, the first job as somebody's assistant, the senior analyst with a
football-pitch-sized Excel model nobody else can read, the two years spent building your own
and discovering theirs computes things wrongly, then Visual Basic. A third of the job was
fetching data and stuffing it into a template; then a couple of weeks of slides; then selling
the story. All of that now arrives at once, and can be made to refresh itself continuously.
Miettinen's own current step: **S&P Capital IQ** is Translink Corporate Finance's principal
financial data provider, and he is meeting their data lead about pulling the data through an
API call directly into his own code — cutting out Excel entirely, on the grounds that Excel is
a dumb 1990s interface wrapped around a cloud calculation, and once you are past it the
constraints of rows and columns stop applying.

**The caveats.** Paasi presses on two, and Miettinen agrees with both. A black box is a nice
toy, but if it vomits out nonsense it is worth nothing, so you must retain enough control to
understand what it is doing — particularly at the start. And the largest gains accrue not to
generalists but to **genuine experts in their own field**, because they are the ones who can
tell the machine how the work must actually be done. His example is fund valuation: rigid,
formal, full of requirements that exist for tracking-error, accounting and audit reasons that
only a specialist knows. If you are a superb Jeep mechanic, cardiologist, nurse, lawyer or
analyst, you hold the advantage.

**The homework.** Paasi's first-ever assignment to Paasipodi listeners is one word: Google
Lovable. Then push it around and make it do something — anything. Miettinen adds three of his
own. Buy a smarter colleague lunch and watch them work for two hours; it is the most valuable
thing he did as a young associate and he still does it. Ask people at the top of a field
directly, as Steve Jobs did — people are, on the whole, wired to answer. And if your IT
department forbids it, buy your own laptop, pay for your own models, and keep it off the
corporate network until you know enough to explain to the IT department what you are doing.

**The warning.** Miettinen closes by declaring the episode Paasipodi's official **AI storm
warning, dated 19 February 2026** — his view is that superintelligence is close to certain this
year and, if not, then next. Paasi's framing for anyone feeling the cold sweat of having missed
the train: there are a great many trains leaving this station, and next week everything will be
twice as fast again.

Original recording (#paasipodi): https://www.youtube.com/watch?v=QsV6_rHdIic

---

Cite as: Sami Miettinen, guest on #paasipodi #61 — The AI Storm Warning — Sami Miettinen on Paasipodi #61, 2026-02-19, https://www.youtube.com/watch?v=QsV6_rHdIic. Record: https://www.neuvottelija.com/media/paasipodi-61-ai-storm-warning/.