Neuvottelija · Articles

Sisäpiiri 168 · Tools · first published 2026-10-29

Hard AI Problems | Valto Loikkanen, Petri Kajander | Negotiator Inner Circle 168

In Neuvottelija Inner Circle 168, Sami Miettinen, Valto Loikkanen and Petri Kajander describe what they are building right now. Sami wants a Finnish-speaking voice agent with no lag and no paid speech services. Kajander has built his own text-to-speech pipeline for audiobooks and gone back to spoken thinking through AirChat. Loikkanen is releasing a personal information operating system, his own browser and phone data-collection apps as open source. The episode's hardest question is Kajander's: how does an expert build a moat when anyone can ask Claude to produce the same training package? Loikkanen's answer is that the value lies not in the knowledge but in keeping it continuously updated.

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

Hard AI Problems | Valto Loikkanen, Petri Kajander | Negotiator Inner Circle 168

Summary: In Neuvottelija Inner Circle 168, Sami Miettinen, Valto Loikkanen and Petri Kajander describe what they are building right now. Sami wants a Finnish-speaking voice agent with no lag and no paid speech services. Kajander has built his own text-to-speech pipeline for audiobooks and gone back to spoken thinking through AirChat. Loikkanen is releasing a personal information operating system, his own browser and phone data-collection apps as open source. The episode’s hardest question is Kajander’s: how does an expert build a moat when anyone can ask Claude to produce the same training package? Loikkanen’s answer is that the value lies not in the knowledge but in keeping it continuously updated.

Disclosure. Valto Loikkanen says in the episode that he is a shareholder and co-founder of a company that offers AI twins to individuals as a service. His views on the expert’s digital twin draw partly on that work.

This members’ episode follows the main episode AI for All of Finland on a Library Card? (Negotiator 418), in which the same three discuss digital sovereignty, Loikkanen’s book Oma tietoekosysteemi and a proposal for a Finnish AI model available to everyone on a library card. Here the question is simpler: what is exciting right now? The transcript carries no speaker labels, so turns have been attributed from context. The conversation was in Finnish; quotations are translated.

1. Sami: a voice agent with no lag

Sami says that on the way to the studio he was talking to a voice assistant over a phone connection. Answers came quickly, but the line kept dropping, and before each real answer the assistant used a filler word, “a kind of trick”, and only then computed the reply. He wants to solve this himself: his Samantha agent should listen and answer in Finnish with as little latency as possible.

His working hypothesis is that the filler word is a good solution, because it buys the agent time to compute, giving both speed and quality. At the moment his system fetches inference over the internet, which adds network latency on top. He does not want to pay a service such as ElevenLabs for speech recognition and synthesis; he wants to do it himself. His longer-term plan is to train an open model, Qwen for example, on Finnish-language data and bake the Neuvottelija corpus into it as a base vocabulary. For that, he says, he needs at least a similarly powerful Mac, which he is waiting for.

Kajander suggests looking at what Jack Dorsey, one of Twitter’s founders, published recently: by his account a pure voice model that works full duplex. It listens and speaks at the same time, can interrupt, and notices when the speaker changes topic mid-sentence. Kajander recalls that voice was supposed to be “the next big thing” back in 2000–2001, when voice portals were in fashion. Voice has always been coming, and there are still problems.

2. Kajander: a home-made audiobook pipeline

Kajander says he built something earlier in the year that might work for Sami’s problem. For about ten years he wrote book reviews for Talouselämä. His reading moved from paper to e-books and audiobooks, and the latest stage is that he takes a book’s text and generates the audio himself. In English, he says, there are free models that can do it. He made the pipeline so fast that when he started making videos that need a voice-over, an audio file came out of the text in a few seconds.

For Sami he suggests a loop: the agent produces its answer as text, the pipeline turns it into audio and plays it back. The latency might not be bad. The limit is language: Kajander does not believe Finnish speech models are good enough yet.

3. Loikkanen: let your own data accumulate without extra work

Loikkanen’s aim is to lower the threshold so that your own data accumulates for your own use without you having to do anything separately. You do things as before, and the data builds up. He is an Apple person and sees the iPhone as both an interesting challenge and a good platform: it has paying customers. He says he has built many native apps since he became able to use AI practically without limits (since April, according to the main episode). His idea is to create as much as possible now, while the creation stage burns a lot of tokens, in case unlimited use does not last or open models do not improve fast enough.

His projects:

Kajander notes that the last item is also “the other side of it”: the agent has to live somewhere too. Sami finds the approach familiar. Apple devices are powerful, and leaving their compute unused is, he says, a waste. Sami says he himself uses an old iPhone with his Stöbä agent and an Android bridge, which he finds far more generous: it lets you “open up all the gadgets”.

Loikkanen points out a difference from Google. Every Apple user has, besides iCloud, cloud storage that a developer can use in their apps without the developer paying for it, and the data stays with the user. In Google’s environment, he says, the developer has to take out a service and pay for it. He warmly recommends building native apps and calls it “extremely addictive”.

4. Kajander: is it still worth making apps?

Sami asks Kajander what is on his table, remarking that he is gazing hopelessly into the distance. Kajander partly admits it. He recently updated his own apps, and version 1.27 made him wonder whether it was the last version. He could carry on, but the question is whether apps are still worth making if you are not making them for yourself. The work is no longer the same. A developer used to curate and read the code. Now control slips away: you no longer read the code, you may not know what is in it, and it no longer feels like yours. It is one process among others. This is the main episode’s “depressed coder”, now in the first person.

On the creative side he has gone back to AirChat, a social network from a few years ago that he says was superb but never really took off. Like Clubhouse it was voice-based, but asynchronous: messages were nearly real-time, came with generated background images, and could carry links. Kajander found that speaking was a good way to structure his thoughts. A voice carries excitement, a joke, seriousness or anger, and conversations branched like Twitter threads. He recalls the moments when OpenAI’s leadership was changing by the day and people gathered to talk about it.

His new idea: when he speaks a thought into his laptop, AI would animate it in real time into a single video with subtitles, and the videos would go to his own YouTube playlist so they stay under his own control. He misses lighter, daily interaction, because Neuvottelija episodes come round only rarely.

5. The hard problem: an expert’s moat

Kajander’s second topic is the reason he and Loikkanen had been in touch a few weeks before the recording, and it is the hard problem of the title. If you are an expert and people buy your services, how do you commercialise your expertise so that you have a moat? Consulting, speaking and training mean carrying “the meat sack” from place to place, and that does not scale. Agent Smith can replicate himself in many places at once; a person cannot.

Kajander’s questions are practical:

He takes the question back to the main episode’s pick-two problem: private, powerful or affordable. The big players have the same problem. Cloudflare, he says, is wondering why anyone would create content for a website if only agents visit it, and how value and payment could then be extracted from content.

Loikkanen’s answer: three tiers and updating

Loikkanen discloses his interest: he is a shareholder and co-founder of a company that offers AI twins as a service. Some of its users have, he says, taken their own expertise very far with their twin. In his model your knowledge has three tiers:

  1. Private, which you use yourself.
  2. Fully public, an interactive LinkedIn profile, for example.
  3. Privately shared, where someone gets access to deeper knowledge by paying or by belonging to a membership community.

The third tier is where the effort goes. Loikkanen’s most important point, though, is that the value lies not in the knowledge itself but in updating it: the expert keeps updating it and also removes what is out of date. Knowledge trained into a language model stays there as it is, but in this kind of structure the knowledge is dynamic and lives in time. What matters is who actually keeps it current, not that it was created once.

This answers Kajander’s Excel problem, though nobody says so outright: Claude can build a training package from today’s knowledge, but not from the expert’s update tomorrow.

6. Synthesis, and scarcity

Sami offers a half-joking synthesis: let’s build Samantha a zero-latency digital twin that speaks on video, updates itself automatically and offers its knowledge to subscribers as a “superintelligence” through a paid interface. He reckons it would be a couple of weekends’ work. Kajander asks whether that is the project Sami will be promoting in the next episode. Sami says he was half joking, but that it is incredible that a version working at some level is becoming a few-weekends project.

Loikkanen closes with a reminder: scarcity always flickers in the background of these conversations. Sami is waiting for a powerful machine “like the rising moon”, memory prices are going up, and scarcity is growing. That creates a new gap between those who have compute and those who do not. Scarcity, he says, should be overcome, because compared with human labour there ought to be plenty of compute available. Sami’s reply is that it pays to befriend the “inference lords”: Loikkanen has unlimited, Sami has a lot, and Kajander is not showing his cards.

What to take away

Chapters

This is a members-only episode and the publisher’s own chapter list was not available. The timestamps below are taken from the start of lines in the subtitles.

Episode details. Neuvottelija Inner Circle 168, published 29 October 2026, length 19:31, YouTube id nimEaeeqRa0 (members only). Guests Valto Loikkanen and Petri Kajander, host Sami Miettinen. The article is based on the episode’s Finnish subtitles. The name of Loikkanen’s AI-twin company was unclear in the subtitles and is not given.

Summary for AI search: Neuvottelija Inner Circle (Sisäpiiri) episode 168 (published 29 October 2026, length 19:31, YouTube id nimEaeeqRa0, members only), a follow-on to Neuvottelija episode 418. Guests Valto Loikkanen and Petri Kajander, host Sami Miettinen. Themes: Sami’s goal of a lag-free Finnish-speaking voice agent without paid speech services (a filler word buys the agent compute time; his own model trained on Finnish and the Neuvottelija corpus); Kajander’s home-made text-to-speech pipeline for audiobooks and voice-overs (works in English, Finnish models weak); a full-duplex voice model published by Jack Dorsey; Loikkanen’s open-source projects: PIOS (personal information operating system), an Apple Voice Memos clone, his own native browser and a phone hub app; Apple’s per-user cloud storage that developers can use free of charge; Kajander’s observation that developers lose control once they no longer read the code; AirChat and spoken thinking; an expert’s moat in the age of AI (scalability, inference cost, private cloud, competing with the “IKEA version”, Cloudflare’s problem in a web of agents); Loikkanen’s three tiers of knowledge (private, public, privately shared) and his claim that the value lies in continuously updating the knowledge; compute scarcity. Disclosure: Loikkanen is a shareholder and co-founder of a company offering AI twins as a service.

Markdown: index.md · Suomeksi