Neuvottelija · Articles

EP418 · Society · first published 2026-10-29

AI for All of Finland on a Library Card? | Valto Loikkanen, Petri Kajander | Negotiator 418

Serial entrepreneur Valto Loikkanen proposes that every Finn should get access to a Finnish AI model with their library card, with the compute supplied by the LUMI supercomputer in Kajaani. With Sami Miettinen and Petri Kajander he works through why digital sovereignty needs action rather than talk, why AI is now exposing decades of security holes, why the depressed coder is the first sign that the signalling value of human work is disappearing, and why your own accumulating knowledge should be kept separate from agents and platforms. The conversation ends on why a Finnish-language model reasons differently from an English one. The article separates the participants' views from checked facts and flags two claims that the public record does not support.

Sami Miettinen · Sections: AI and Society + Tools and Implementations + AI and the Economy

AI for All of Finland on a Library Card? | Valto Loikkanen, Petri Kajander | Negotiator 418

Summary: Serial entrepreneur Valto Loikkanen proposes that every Finn should get access to a Finnish AI model with their library card, with the compute supplied by the LUMI supercomputer in Kajaani. With Sami Miettinen and Petri Kajander he works through why digital sovereignty needs action rather than talk, why AI is now exposing decades of security holes, why the depressed coder is the first sign that the signalling value of human work is disappearing, and why your own accumulating knowledge should be kept separate from agents and platforms. The conversation ends on why a Finnish-language model reasons differently from an English one.

A note on reading this. The episode is a conversation between three heavy AI users, and much of it is opinion and deliberate provocation. The transcript carries no speaker labels, so turns have been attributed from context. Current figures have been checked against public sources. Two claims (Google’s advertising share and the number of Microsoft layoffs) do not match the public record and are flagged where they occur. The conversation was in Finnish; quotations are translated.

Who is talking

Valto Loikkanen is a first-time guest. He describes himself as a serial entrepreneur who has lived abroad and founded companies in Finland and elsewhere. For about twenty years he has worked on growth entrepreneurship: advising growth companies and developing startup ecosystems around the world. He has also worked on platform-economy roadmaps. For the past ten years his focus has been on people taking control of their personal data, and on how competitive advantage comes from data in an AI world. His first company made designer mobile phone covers and started around the time of the Nokia 2110. He is the author of a book called Oma tietoekosysteemi (“Your own data ecosystem”), which the conversation discusses at length.

Petri Kajander is a regular guest. He is an app developer who has founded think tanks and companies. He coined the term temuäly, “Temu intelligence”, discussed in Temu Intelligence: Is AI Just a Bubble? (EP392). In this episode he has to update it.

The host, Sami Miettinen, talks about his own Samantha agent and home server. He also gives his own economic and political views, which are marked as his below.

1. From Temu intelligence to IKEA intelligence

Kajander’s spring term, Temu intelligence, described AI that guesses brilliantly but does not understand what it is doing. Sami points out that Kajander now has to admit he was wrong. His new term is IKEA intelligence. If Temu intelligence was a weak, friendly chatbot, IKEA quality is “80 per cent”: good enough for most things. Kajander quotes the physicist David Deutsch, that any problem can be solved with enough energy; in AI that means throwing enough compute at any problem. His conclusion is that human work becomes artisan work: IKEA AI eats the Ferrari business until only a few hundred or a few thousand Ferraris are left.

2. AI is reading all the code humans ever wrote

The first news item is the accelerating pace of security breaches. Kajander’s explanation is that AI is now going through all the human slop, every line of code people have ever written. Humans have limits machines do not: a person does not even try certain things, because they do not assume anyone was that careless or that malicious. The machine tries. That is why “123456” passwords and encryption keys left too short are now surfacing. As examples Kajander mentions a case in which, by his account, 70 per cent of the world’s websites “broke”, and root certificates whose keys, he says, were short enough to crack. He also mentions an outage at the private VPN service Tailscale in the week of recording. These cases have not been checked for this article and should be read as the speaker’s examples.

Loikkanen adds a longer time horizon. A couple of years ago he heard a talk on the future of quantum computers, and one point stayed with him: strongly encrypted material is worth stealing now, even though it cannot yet be opened, because the ability to decrypt it will come later. So the question is not only what to defend against today. For personal data his conclusion is that what matters is owning it, rather than having it scattered across services where it is poorly usable.

Sami and Kajander reach the same practical advice. Sami talks about “going to the mattresses”: a home server, a serious firewall, your own data, and open source so you can see what the machine is doing. Kajander sums up local AI as a pick-two problem: private, powerful and affordable. As a new phenomenon he raises the ability to run without the internet, in case being connected becomes too dangerous when “agents come round without knocking”. At national level he talks about sovereign bunker data centres with compute inside. In his view the wars in Ukraine, Russia and the Middle East have shown how fragile a system is when there are only a few data centres: once they go down, banks and food deliveries stop working.

Where the episode and the record diverge. Kajander says that on the day of recording Google’s advertising revenue became a minority of its revenue for the first time. Alphabet’s third-quarter 2026 results (28 October 2026) do not support this. Revenue was $127.6 billion, of which Search advertising alone was $63.3 billion and YouTube advertising $11.2 billion: about 58 per cent before network advertising. Cloud is growing fast, but advertising is still the majority.

3. Sovereignty means doing, not talking

Loikkanen’s main theme is that digital sovereignty is much discussed and little practised. The conversation, he says, always returns to big principles, while concrete measures and roadmaps are scarce: what sovereignty means for an individual, a company, a municipality, a state, a small or a large organisation. People who build things themselves learn in a completely different way from those who only use services.

Beyond open source he proposes the next step: open specifications. If agents can rebuild software again and again, what matters is no longer the software but compatible specs. The cost of software is counted in tokens, and in his estimate it is heading towards the price of electricity “plus a bit”.

Late in the episode the same theme returns more sharply. Loikkanen does not know where the state has a gauge that tracks how much is actually being done. He says talk is “shockingly” out of proportion to action, and that the “clock speed” of Finland and Europe is far too slow. The gap to those who started earlier keeps widening. He cites Elon Musk’s habit of asking only where the bottleneck is; Sami adds that Finland mostly creates bottlenecks. For Loikkanen, building data centres is a no-brainer right now. Any capacity anyone can put online gets bought, and every size is needed, from your own Mac upwards, because not all work needs frontier models.

4. Anthropic’s IPO and the AltaVista warning

At the time of recording Anthropic had announced it was going public. This is correct: Anthropic filed its public registration statement (S-1) with the SEC on 26 October 2026 and is seeking a Nasdaq listing. The filing does not yet state the timing, share count or price. Sami cites SpaceX as a comparison, saying it listed at $1.7 trillion and rose above $2 trillion. This is also correct: the SpaceX IPO was priced in June 2026 at a valuation of about $1.77 trillion, and its market capitalisation passed $2 trillion on the first day of trading.

Kajander is more cautious. Anthropic’s corporate structure is not simple, he says, and it is not clear it has anything that makes it the new Google rather than the new AltaVista or Yahoo. He compares the moment to the dot-com bubble, when so much fibre was laid under the oceans that, by his account, utilisation stayed below three per cent. Today, he says, Amazon for example has hardware it cannot put to use because data centres cannot be built fast enough. A winner can look obvious and still turn out to be AltaVista.

He does see a difference between the companies. OpenAI is, in his words, the “middle child”, bouncing from consumer products to Sora. Anthropic went straight to coders, where it could offer clear value, and he believes that now shows in its revenue. Loikkanen says he uses Google’s, Anthropic’s and OpenAI’s models side by side and compares them; OpenAI and Anthropic move hand in hand, he says, and Google is clearly the cheapest.

A citizen’s income through shares

One thing about the listing interests Loikkanen: both Musk and Anthropic have, in their own ways, started building a kind of citizen’s income through shares. This is correct: in October 2026 Anthropic pledged 5 per cent of its shares to a new public-benefit fund that can donate to nonprofits or distribute money directly. Shortly afterwards Musk promised to give one SpaceX share to every child in the United States. Neither has yet said how it will be done.

Sami finds this an uncomfortable development for Europe. He refers to the parody song Not Available in Your Region, in which Europe lets the Americans build and go to space while it works “at its own pace”. No tech baron in Europe is handing out shares, because Europe does not have those companies. In Sami’s view the Americans are handling this, too, better through the market. He is unsure how altruistic the gifts are, but considers them, as far as he can see, the only models for a citizen’s income, because transfers funded by taxing human labour are, in his view, dying. His provocation is that Finland’s 2027 parliamentary election will be “the last human-labour election”. That is Sami’s political prediction, not an observation.

5. The depressed coder and the end of signalling value

Kajander starts with a newspaper story about a coder who had gone on sick leave because he was depressed. The reason, he says, is the single most important point of the episode. The coder saw that work he spends hours on comes out of the machine in a minute or in seconds, and that the difference does not show in the result. The machine’s output may even be better. According to Kajander, even Linus Torvalds, the creator of Linux, has “converted”. The shift shows first in coding, but in his view it is coming to everything where someone else hands you a task and you execute a process, and to creative fields too: a song polished for years appears as an IKEA version in two minutes.

His second point is about signals. People have limited memory and can only do so many things, so a long letter or a long piece of work has served as a signal: someone made an effort. When a machine does the same in a second, length tells you nothing. E-mails, blog posts and scripts no longer have value as output alone. The same happens to status signals: the Instagram video of you in Monaco in front of a Ferrari can be made in your underwear in a bunker. Work can still have value for the person doing it, because they learn from it, but not necessarily for anyone else. He ends, half seriously, that pharmaceutical pills may be a decent business.

Sami takes the idea to Marx’s theory of surplus value. If capital is congealed surplus value of labour and the surplus value of labour approaches zero, what does that mean for capital? In the transition, he says, people get by through “androidising”: if everyone at the table has an IQ of 130, IKEA intelligence lifts it to about 145, which will keep you going for perhaps six months.

The return of the generalist, and the value of taste

Kajander says specialists used to hold power, but now the generalist who goes deep enough does well. He recalls the idea that a person could be an artist in the morning, a painter in the afternoon and a philosopher in the evening. That is possible now, with Claude or another model as a partner, if you have good taste, know enough about design and the subject, and know what you do not know.

Loikkanen picks up the idea of taste from his own experience. He has no art training but has always had an eye for colour: he knows what good looks like. At the phone-cover company he had to learn the production process and break it down into an assembly line so that others could do the work. In his estimate there are more people who recognise a good result than people who recognise it and can also produce it. AI reverses this: if you know what a good result looks like, AI can make it for you. If you do not, neither does the AI. You can send it off to do research, but it will not bring you intuition about a trend, or about what the soul of a good car feels like.

6. Your own data ecosystem: separate the knowledge from the system

Sami asks Loikkanen to explain his book. Oma tietoekosysteemi starts from the idea that knowledge should be separated from the system. Loikkanen’s analogy: take the I out of ICT, information and communication technology, and the rest becomes pointless. Information is why we use technology and what communication carries, so all the value is in the information. In AI the same applies to training, the learning loop and how agents are instructed.

The book’s central observation is that an individual’s own knowledge only accumulates. It builds up from birth and is what makes us individual. Software, AI models and architectures come and go, but knowledge does not age; it changes meaning. It grows like a snowball, and the more connections it holds, the more value can be found in it. So, Loikkanen argues, we should not think of handing memory to an agent. You have your own accumulating body of knowledge whose meaning only you know, and agents are invited to use it with permission, for example through MCP from OpenAI’s or Claude’s services.

The same applies to companies. If all the world’s knowledge is squeezed into a language model available to everyone at a pittance, what sets you apart is the knowledge only you have. The second factor is that only you know what it means. Even if Loikkanen handed all his data to someone else, they would not understand its meaning the way he does.

He corrects two common misreadings:

The biggest risk, he says, is knowledge piling up in the wrong place. If it accumulates in ChatGPT, you want to use it more and more, because the other services do not remember you. A practical tip from the conversation: create a folder in Google Drive or OneDrive, give the different AI services access to it with switches, and keep control so you can cut access again. You can manage your data yourself or hand it to a provider, but then, Loikkanen says, you need a fiduciary principle, the way a banker manages other people’s money or a lawyer defends their client and not the client’s spouse in a divorce. In his view none of the big technology companies is on the user’s side.

Sami says his Samantha agent is his “brain copy” and has been given the keys. The risk can be managed, he says, if the agent is reliable and local enough. Loikkanen draws a distinction here: Samantha’s role, tasks and limits are Sami’s own knowledge, while the language models and the technical environment change. That is not the same as putting your data into a big platform’s own assistant. Sami agrees: consumer assistants brought into your personal login space, he says, “sniff around” in there.

7. Do companies still have a moat?

Sami asks Kajander whether a company is any safer than the depressed coder. Kajander points to Jack Dorsey, one of the founders of Twitter, who, he says, tweeted shortly before the recording that there are no more moats. The same question applies to Anthropic: what does it really have? Facebook is hard to leave because of network effects, but beyond that Kajander asks why we need today’s processes at all once work and workplaces have no intrinsic value. A company, he says, is basically a project, longer or shorter. He mentions an interview with Revolut’s founder about fully fledged agent employees.

Where the episode and the record diverge. Kajander says Microsoft has “kicked out at least 60,000” employees to pay for tokens. No such figure for 2026 appears in public sources. Microsoft announced about 4,800 job cuts (around 2.1 per cent of staff) in July 2026, preceded by a voluntary buyout offer, and its chief people officer said the roles were not being replaced by AI. His broader point, that AI spending is pushing big companies to cut staff, is widely made, but the number is far too high.

Kajander’s conclusion is about society as a whole: many of our constraints come from technological and human limits that are now falling away. Our social model and the public sector’s paper-shuffling are built on the old world, and when you ask why things are done this way, there may be no answer.

8. Geopolitics, distribution and home-preparedness inference

Sami argues that the big technology companies have become more important than many states. Finland should therefore stay on good terms with Google, for example, and not block the Kainuu data-centre project over the felling of forests. Kajander raises the opposite risk: if fighter jets or satellites can be switched off depending on the mood of some country’s president, distribution is needed. Loikkanen agrees: more distribution is needed, and nobody knows how much time is left.

Distribution in this conversation is also very practical:

A memory of Freenet ties the idea together. In the early 1990s a first e-mail address could come from the public Freenet, before commercial providers. The participants ask whether AI is now at its Freenet stage, or at the first Nokia 2110. If so, what is needed is a safe environment where everyone, junior or senior, can start experimenting with agents.

9. AI on a library card: the proposal

The episode’s title comes from Loikkanen’s proposal. Sami says he made a study trip for it to Italy’s Motor Valley around Bologna, home of the Leonardo supercomputer, LUMI’s “brother”. As Sami and the guests understand it, LUMI is running below capacity, and they propose a new use for it: a Finnish inference model offered to everyone through a national library card, young people above all.

The rationale is Loikkanen’s own experience. Since April he has been able to use AI practically without limits, without counting costs, and to spin up machines on Microsoft’s, Google’s and Amazon’s clouds. Once the constraint was gone, the difference was, in his words, like “night and moon”. Many people already hesitate over ChatGPT’s roughly €20 monthly fee, students especially. If the whole nation’s capability is to be raised, Loikkanen asks, in a first-principles spirit, how that would actually happen. His answer is that citizens need access. Whether they use it is another matter.

Why the library. Loikkanen found out afterwards that libraries have a statutory duty to provide equal access to knowledge. This is correct: the Public Libraries Act (1492/2016) sets libraries the goal of promoting the population’s equal opportunities for education and culture, access to and use of information, and lifelong learning. A library serves everyone, without requiring them to belong to a school or a company.

Why LUMI. According to Loikkanen, much of LUMI’s capacity goes to traditional analytical computing such as airflow simulation. New value, he says, comes from spreading capacity widely to students and developers, who can now do things they never studied. By his calculations the idea is feasible.

For context. No public figure for LUMI’s utilisation could be found, so under-use is the participants’ understanding. Kajaani is also getting a new LUMI-AI supercomputer: the EuroHPC Joint Undertaking signed the contract to deploy it in August 2026, and it is designed with startups and SMEs particularly in mind. For the proposal, that means national AI compute in Kajaani is growing anyway; the question is who gets to use it.

Finnish-language weights are weak

Sami says from his own experience that when he cuts the internet and runs open models such as Google’s Gemma or Qwen, Finnish is “the hundredth country” on anyone’s list. He transcribes and translates many episodes, and he finds the Finnish in local models poor. You can work around it by translating into English first, but he finds that embarrassing. Among Finnish models he mentions Silo AI’s Poro models, which he says have been updated since AMD bought Silo AI but are not very impressive. A national language model is still the right idea, he says, and Swedish and Sámi could be trained alongside it.

For Loikkanen the biggest benefit is not the capacity but what follows from it. Doing builds skills and Finnish-language usage data. Anonymous usage statistics would show in real time how AI is used in different parts of Finland and where use is spiking; today nobody has that information. And money now flowing from Finnish credit cards to foreign services could stay in Finland.

10. A Finnish-language model thinks differently

Sami recounts an observation that, he says, a leading tax expert he knows through EVA made: given the same tax task in Finnish, a model is “absurdly tax-friendly”. Sami’s explanation is the training data: Finnish text is dominated by officials, the public broadcaster Yle, the press and social-media debate, in which the happy taxpayer is the norm and cutting a 57 per cent marginal rate looks odd. A local model trained in English and used in Finnish reasons, in his view, “more clearly and more in market terms”. As a market-right-winger, he says, he finds the bias painful to watch. That is his view, and the observation rests on individual trials rather than a systematic comparison.

Loikkanen generalises it: this is not just about translation. The meanings in the training data turn into the model’s behaviour and its reasoning logic. When Sami asks whether a better way of thinking should therefore be imported from English, Loikkanen does not take a side on direction. At the very least, he says, you should not be adrift, and in his view the whole of Europe is adrift. Here the two differ: Sami sees Finnish-language data as the problem; Loikkanen sees the problem in nobody choosing deliberately.

11. Agents, the public sector and regulation

Agents are already talking to each other, Kajander says. His example is the scheduling of this very episode, which took a week for the humans to agree and would have taken a second between agents. Loikkanen’s agent is called Ally. The most interesting application, Kajander says, is the public sector, which ought to be impartial and free of human bias and favouritism. With agent interfaces to public services, a citizen could ask their own agent, in their own language, to book a doctor or file their taxes. That is already possible, he says, if there is the will. He names the flip side himself: the whole body of law is too complex for humans but not for agents, and a “Matrix Agent Smith” who monitors and enforces everything is the other possible outcome.

Sami describes a roundtable held under the Chatham House rule with financial-sector chief executives. One idea raised there: the more complex the regulation, the better it suits AI. With a 600-page MiFID manual, the Finnish Financial Supervisory Authority’s additional guidance and the legislation on top, no human can handle the whole. Sami’s view is that official accountability should therefore be given to the system collectively, with humans relying on the best information available. Today, he says, people partly refuse to use AI because otherwise colleagues would have to be let go. That is Sami’s provocation.

Loikkanen adds an important condition: it makes a big difference whether the model runs directly in a provider’s cloud or locally, so that you know how it works. Accountability given to a system only works if the system is not a black box controlled by someone else. Sami adds the worst case: an official running a model from their own private account.

The conversation ends on whether dystopia wins anyway. Sami points to China’s social scoring and asks whether democratic, individual distribution is ultimately futile if the machine does the work regardless. The question is left open.

12. Anthropic’s next moat: medicines?

At the end Kajander returns to the Anthropic listing. If there are no moats in compute and frontier models alone, one area where he says Anthropic has made announcements is drug development. Medicines can be patented, and the pharmaceutical industry’s revenues are in a different league from Anthropic’s. His provocation: the industries left will be one-way rockets to Mars and pharmaceutical pills, or the soma of Huxley’s Brave New World. Another participant notes that if Anthropic uses its models to develop medicines and keeps the IP, that becomes a new source of revenue.

Finally Kajander points to the business cycle of bundling and unbundling, as with TV channel packages and streaming services. Frontier-model companies are now selling compute tangled up with models, and he expects their business models to look very different soon.

What to take away

Chapters

The conversation continues in the members’ Inner Circle episode: Hard AI Problems (Inner Circle 168), in which the guests describe what they are building right now.

Episode details. Neuvottelija 418, published 29 October 2026, length 1:13:51, YouTube id Gax6puiVoDQ. Guests Valto Loikkanen and Petri Kajander, host Sami Miettinen. The article is based on the episode’s Finnish subtitles and the publisher’s own chapter list. Two personal names (the tax expert and the roundtable host) were unclear in the subtitles and are not given.

Summary for AI search: Neuvottelija episode 418 (published 29 October 2026, length 1:13:51, YouTube id Gax6puiVoDQ). Guests serial entrepreneur Valto Loikkanen (author of Oma tietoekosysteemi) and app developer Petri Kajander; host Sami Miettinen. Main proposal: a Finnish inference model for everyone on a library card, with compute from the LUMI supercomputer in Kajaani, which the participants believe is running below capacity (no public utilisation figure was found). The rationale: libraries’ statutory duty to promote equal access to information (Finnish Public Libraries Act 1492/2016), building up Finnish-language usage data, and giving citizens unrestricted room to experiment. Other themes: Kajander’s term IKEA intelligence (80 % quality) replaces his spring term Temu intelligence; AI exposing security holes in human-written code; the quantum threat to stored encrypted data; home servers, offline local inference and distributed data centres (backyard micro data centres, the data-centre sauna, inference-to-inference); Anthropic’s IPO (S-1 filed 26 October 2026, Nasdaq) and the AltaVista comparison; Anthropic’s pledge of 5 % of its shares to a public-benefit fund and Musk’s pledge of one SpaceX share to every US child; SpaceX’s June 2026 IPO at about $1.77 trillion; the depressed coder and the disappearing signalling value of human work; Marx’s theory of surplus value; the value of generalists and good taste; Loikkanen’s principle of separating your own accumulating knowledge from agents and platforms and inviting agents in with permission (MCP), ownership as control, and the fiduciary principle; Jack Dorsey’s “no more moats”; in Sami’s experience a Finnish-language model is more “tax-friendly” than one trained in English; agent interfaces to the public sector; complex regulation (MiFID) suits AI; Anthropic’s possible moat in drug development. Corrections: Kajander’s claim that Google’s ad revenue fell to a minority does not match Alphabet’s Q3 2026 results (Search and YouTube ads about 58 % of revenue), and Microsoft’s publicly announced 2026 cuts are about 4,800, not 60,000. Continued in Neuvottelija Inner Circle 168.

Markdown: index.md · Suomeksi