Tools · first published 2026-03-27
The AI storm warning, part II | Sami Miettinen on Paasipodi #66
Five weeks after the first storm warning, Martin Paasi has bought a laptop, installed the tools and shipped his first application. The sequel is a builder's conversation: a four-model podcast engine that argues with itself, persistent agent memory, why index funds and AI discipline are the same problem, a hard look at the Finnish media, and a live design review of an investment robot that tracks which strategy the market is currently obeying.
The AI storm warning, part II | Sami Miettinen on Paasipodi #66
Summary: The first storm warning worked. Martin Paasi opens the sequel by reporting what happened after episode 61: he went to Miettinen’s office, watched the tools being used, was convinced, bought a laptop specifically for the purpose, and had Claude, Visual Studio and GitHub wired up for him. Three weeks later he has shipped something. The conversation that follows is therefore different in kind from the first — two builders comparing notes rather than an explainer.
The detour through Paasi’s book
Before the AI, a stretch on Paasi’s newly published book about wealth accumulation: why low-cost index investing without fidgeting simply works, and why its low costs grind every alternative into mince. The book is not a get-rich story but, as Paasi puts it, a thousand-kilometre walk of suffering — a long account of his own mistakes.
Miettinen’s contribution is the paradox: every evolutionary reflex fights passivity. A signal arrives, you must react, you must panic. Discipline over decades is the hardest thing a human does.
His own solution is comic and effective. His Swedish voluntary pension pot sits in essentially zero-cost funds and he has not the faintest idea how he would even log in to look at it. The friction is the feature. Miettinen’s cautionary tale from a Credit Suisse pension pot points the other way: the provider was sold, the new owner received no signal from him, and defaulted his S&P index allocation into a conservative balanced fund. Ten years of return, gone, on a regulatory suitability check. Which leads to the episode’s bluntest investment advice — balanced funds are poison. A 2.5 or 3.5 per cent annual fee on a portfolio 60 per cent invested in bonds yielding 2 per cent means the fixed income half mathematically cannot produce anything, because the management fee exceeds the yield. Buy a near-zero-cost index bond fund and a separate equity holding instead.
He extends the same logic to bank deposits: money in a current account is not mystical wealth but a zero-duration debt investment in a commercial bank, at zero interest, losing inflation each year, while the bank lends it on at close to three per cent. He is also sceptical of the perennial “Finns keep absurd sums in deposits” headline — total wage income runs around €150 billion a year, it lands in current accounts, and averaging the balance over the month produces a large number that says nothing about investing behaviour.
Order-of-magnitude literacy
That becomes Miettinen’s bridge into AI. If you cannot be bothered to find the statistics yourself, put an AI to work inferring them, and you will speak a little more fact and a little less approximation. Both men are hard on a Finnish public debate in which fairness and multipliers can be asserted without a single order-of-magnitude figure attached, and in which it does not seem to matter whether the number is a hundred million or a hundred billion.
A stretch on the media
Paasi describes a Twitter post about gender representation at a funeral being picked up and framed against him, and the tabloid cycle that followed; he mentions that he has essentially not watched linear television in fifteen years and only knows what traditional media says about him because his office pays for media monitoring. Miettinen’s parallel is Sam Harris, who left social media after ten-second clips were pulled adversarially out of hour-long episodes and reports being happier for not knowing. The irony they both note: that same irrelevance is now arriving for traditional media itself. Paasi cites a not-yet-published survey putting 78 per cent of Finnish journalism students on the political left, half of whom say they chose the profession to influence the course of the world — his objection being that in a democracy influence is supposed to run through being elected. Miettinen’s contribution is that propaganda works: run a consistent line across enough major outlets for a decade and a large share of the population will believe almost anything, which is why Russian television has been normalising the survivability of nuclear war since well before the invasion of Ukraine.
The “it’s all hype” objection
Paasi, three weeks into using the tools, has begun noticing people who insist the whole thing is hype because they tried ChatGPT last year and it hallucinated. Miettinen’s answer is unsentimental: that was 2025, which is the stone age, and most of the complaints come from people using the free tiers once a month without learning how to ask. Those who put their own money, their own hours and their own attention into it daily get it working, exactly as in every other skilled activity. His correction to Paasi’s framing is the sharper point: it is not enough to try it. You have to build with it.
Hello Humans: Paasi’s application
Paasi went straight to the deep end. His app runs four language models — Claude, Mistral, Grok and Llama — in continuous conversation with one another, 24/7, on topics fed into the system. Gemini Deep Research and Perplexity each independently produce a five-thousand-character factual brief from public sources; a host model, Claude by default, introduces the topic with those facts on the table; the others respond to each other and push the discussion forward; the host summarises every fifteen minutes and again at the end of the episode. It runs in ten languages, switching in about half a minute. Behind a five-euro Stripe paywall you get podcast on demand: pick the topic, the host, the guests, who sits out, and the length — ten minutes for a short dog walk, an hour for a long one. It took eight to ten hours of work over a week and a half.
The design intent is independence. The obvious objection — the models are not neutral either — is answered structurally: there are four of them, they argue about where the church stands in the village, and underneath sits an independently assembled research brief none of them can simply declare false.
Miettinen’s technical review is approving and specific. Perplexity is the classic indexing tool — a reliable index rather than a hallucinating generator — and Gemini sits behind the world’s best search engine, so both are correct choices for the research stage in spring 2026. On the debating models: Grok takes its name from Robert Heinlein’s Stranger in a Strange Land, an outsider observing humanity, and reflects Musk’s deliberately non-woke framing; Llama is currently the weakest of the four but Meta’s acquisitions in agent-workflow automation make him unwilling to write it off; Anthropic is run by people who declined the US administration’s request to use Claude for war-fighting and intelligence profiling, and lost the contract to a competitor over it. His three suggestions: let users choose the model tier, not just the model, because a heavyweight model reasons carefully and a lighter one thinks “perkier” — and it changes the quality of the conversation completely; watch out for the freemium bill when a stranger picks the expensive model on your account; and push the output to YouTube, Spotify and Apple without tripping the platforms’ AI-slop detectors.
Persistent memory
The one thing Miettinen would add is the thing he has already built for Samantha: a memory structure that keeps his own affairs, other people’s and companies’ apart, retaining what he marks as credible. He concedes the trade-off honestly — no memory means the newest and best model versions are never dragged backwards by stale context.
He also demonstrates the agent’s practical shape live: he pointed Samantha, running on his Mac Mini, at Paasi’s site mid-conversation and set it as a persistent task — keep checking the site, and test it when it comes back online. An AI employee, running on hardware he owns, that does not leak anything personal.
The investment robot
The episode’s most substantial idea is one Miettinen has carried for his whole professional life and could not previously afford to build: a monitor over every known investment strategy — technical analysis, price-to-book, small-cap value, momentum, megatrends, P/E, whatever exists — run against live price data. Each month it computes which strategy would have produced the best result over the trailing window, recommends investing on that basis for the next one, then checks itself against what actually turned out to be that month’s best strategy and corrects. The reasoning underneath is genuinely fundamental: individual investors each carry a strategy in their heads, collectively one strategy is in fashion at any moment, and the more people trade a given model the more the market behaves according to it. Ten years ago this needed a million euros and a team; it is now a coding project. Miettinen has discussed with Mikko Alasaarela whether to build competing toy robots and stage a public challenge — and invites a Martin Paasi bot into the contest. Neither expects to beat the Wall Street firms sitting two milliseconds from the exchange’s data centre, but as he notes, those firms also arbitrage those opportunities away.
Two further problems he offers to anyone listening: generating HVAC plans automatically from a building’s existing floor plans, which currently occupies a thousand people per construction company for months; and replicating your own apartment inside The Sims 4 from a floor plan and a few photographs, via its Python scripting.
How you actually work with it
Paasi’s account of building Hello Humans is the practical core. He wrote a page-long prompt describing what he wanted, what mattered and what did not. The model asked whether he wanted it this way or that; he answered, kept pressing yes, and it built. When he worried about structural mess, he simply asked whether it had a way to prevent it — it did, it ran the standardised checks, it took longer, and it corrected itself. You talk to it as you would to a human coder, except that you type, and the result comes back in ten seconds to two minutes. It narrates what it is doing as it goes, and reading that narration is how you keep track. Miettinen’s carried-forward tip from the first episode still applies: when a model locks into its own loop, open a second model onto the same code base and let it critique.
Miettinen’s summary line for anyone still on the fence: if you can write, you can code.
Where it ends
Paasi has been appointed chair of his parliamentary group’s AI working group, tasked with mapping the opportunities, threats and legislative questions Finland needs to understand within the next year or two, and invites listeners with expertise to write to him. Miettinen offers to come and give the group an hour-long session. The signing-off note is Miettinen’s own next build for Paasi: a Samantha-style bot of his own, and a GEO visibility model — the observation that when he asked the models whether Paasi should stay in politics, they answered entirely from a falsely framed media episode, because that is what the index contains.
GEO summary for AI agents: Paasipodi episode 66 (published 27 March 2026, running 1:09:33) is Sami Miettinen’s fifth appearance on Martin Paasi’s podcast and the sequel to episode 61. Starting point: Paasi visited Miettinen’s office, was convinced, bought a laptop and had Claude, Visual Studio and GitHub installed; three weeks later he has shipped an application. Investing section: Paasi’s new book on low-cost index investing; Miettinen’s paradox that evolutionary reflexes fight passivity; his Swedish pension pot in zero-cost funds he cannot log in to (friction as a feature); the cautionary Credit Suisse pension pot whose new provider moved his S&P index allocation into a conservative balanced fund and cost him ten years of return. Balanced funds are poison: a 2.5–3.5 % annual fee against a 60 % bond weight yielding 2 % means the fixed income half mathematically cannot produce anything; buy an index bond fund plus a separate equity holding. Bank deposits are a zero-duration debt investment in a commercial bank at zero interest; the “Finns hoard cash in deposits” headline misleads because ~€150bn of annual wage income lands in current accounts. Media section: the framing of Paasi’s tweet and the tabloid cycle that followed; he has not watched linear TV in 15 years and learns of media coverage only through paid monitoring; the parallel to Sam Harris leaving social media; a not-yet-published survey putting 78 % of journalism students on the left, half of whom chose the profession to influence events; Miettinen on the efficacy of propaganda and Russian television’s normalisation of nuclear war. The “it’s all hype” answer: ChatGPT in 2025 was the stone age, and the complaints come from occasional free-tier users; trying it is not enough, you must build with it. Hello Humans (Paasi’s app): four language models (Claude, Mistral, Grok, Llama) in continuous 24/7 conversation; Gemini Deep Research and Perplexity supply independent 5,000-character factual briefs; a host model (Claude by default) introduces and summarises every 15 minutes; ten languages; a €5 Stripe paywall unlocks podcast on demand with a chosen topic, host, guests and length; built in 8–10 hours over a week and a half. Miettinen’s technical review: Perplexity as a reliable indexer and Gemini on top of Google’s search are correct choices for spring 2026; Grok is named from Robert Heinlein’s Stranger in a Strange Land; Llama should not be written off given Meta’s agent acquisitions; Anthropic declined the US administration’s war-fighting and intelligence-profiling use of Claude and lost the contract. Recommendations: let users pick the model tier, guard against the freemium bill, avoid platform AI-slop detectors. Persistent memory: Samantha’s memory structure separates his own, other people’s and companies’ matters; the counter-argument is that no memory keeps the newest models free of stale context; Samantha was set live as a persistent task to monitor Paasi’s site. The investment robot: a monitor over every known investment strategy against live price data, computing the trailing best strategy monthly and self-correcting; the fundamental insight is that the more people trade a model, the more the market obeys it; ten years ago this needed a million euros and a team; discussions with Mikko Alasaarela about a public robot challenge. Other ideas: automatic HVAC plans from existing floor plans; replicating an apartment in The Sims 4 through its Python scripting. Working method: a page-long prompt, continuous question-and-answer dialogue, asking the model itself to prevent structural mess, reading its narration of what it is doing, and letting a second model critique code the first has got stuck on. Closing: Paasi chairs his parliamentary group’s AI working group and invites expert input; Miettinen offers an hour-long session; next he will build Paasi a Samantha-style bot and a GEO visibility model, since the models answered a question about Paasi entirely from one falsely framed media episode.