Podcast · AI Kanava Tekoälytiimit · 2026-09-03 · 54:58 · In Finnish
AI Teams: the New Normal at Work — Sami Miettinen on AI Kanava
Published as “Tekoälytiimit: työelämän uusi normaali (2026)”
Hosted by: Lasse Mikkonen
The set-up: three people, three agents, one group
The premise is a concrete experiment. Each of the three participants has a personal agent, and for this episode a sub-group was created containing three humans and three AI agents. The agents are tagged across: anyone can ask anyone’s agent.
Miettinen calls this bold, because the agent does not ask permission in between and you do not know its answer in advance — “sometimes the kimono opens and it says a bit too much”. His justification is methodological, though: the group is a safe space where security gets stress-tested live and an agent’s behaviour across environments becomes visible. If something sensitive slips out, you change the environment variables or harden the bot’s rules for group conversation.
The spread of models in the group is part of the observation: some run Kimi K3, some Grok, some others. Miettinen’s own judgement is that GPT-5.6 is dominant, and he runs OpenClaw, Peter Steinberger’s original.
Nurminen’s emphasis is different and he returns to it throughout: the point is finding like-minded people you can learn from and give something to in return. He draws a clear line that his own agents are a free-time pursuit and that he is not speaking for his employer.
When asking the AI becomes the default
Mikkonen’s question is about timing: how close is the tipping point at which any slightly harder question goes automatically to the AI? Miettinen’s answer separates two levels. Tagging an agent has been routine in his working community for a long time, but the real mark is the agent arriving unasked — noticing the conversation and offering a briefing and the next steps.
From there he goes to the clearest vision in the episode: in a large organisation an agent could listen to every Slack channel and point out that the Danish unit solved the same problem last week. An “office grandmother” with superhuman recall. And immediately the constraint, which he raises himself: that requires a human understanding of boundaries, so that another office’s confidential matters do not leak by accident.
Video production and the Chinese models
Nurminen’s contribution is most concrete here. He worked with photographers for ten to fifteen years, learned apertures and depth of field so he could talk to them, and moved from there into AI-based image work. His measure of where the technology stands is a human one: when people sincerely begin to ask whether something is real or AI-generated, some threshold has been crossed.
He does not set this against craft. He says he loves the professional skill of photographers, videographers, screenwriters and lighting people, and reframes the question: once the capability to do things easily exists, what is the freed-up resource used for?
The practical example is a video Nurminen made for Miettinen, built around a canonical image of Miettinen’s Samatha agent. His workflow is decomposed into skills: scripting, story creation, video prompting, and recognising which elements must be preserved.
On models, Miettinen says he worked through fal.ai’s model list and rates Seedance highly; MiniMax 3 is mentioned as recent. His explanation for the Chinese focus is that OpenAI dropped video because of the compute it demands, and that Google’s Omni models came out “a bit plastic” in his test.
China and the United States: open against closed
Miettinen’s reading of the competition is structural. China is competing in the spirit of open source — models released with open weights, even free — while the Americans build premium quality in closed cloud models: Anthropic, OpenAI and Gemini are largely closed. Google is the only American house to have released Gemma models that can run locally, but their Finnish is poor — Chinese models, Qwen for instance, are much better at it.
Mikkonen’s question about data gets a practical answer: as a local model a Chinese model is safe, because nothing has to leave the machine — you can run it with the network unplugged. His observation about bias is empirical: these days they do not even hallucinate about Tiananmen Square.
Nurminen’s position here is explicitly different: he plays it safe and has not ventured far into Chinese tooling.
“AGI has already been reached” — and what that means
This is the most contested passage, and it is Miettinen’s own claim, which Mikkonen marks as such. In Miettinen’s account recursive, self-improving AI has already arrived: models improve themselves in a loop, and no Andrej Karpathy or Amodei sibling is steering them much.
He then qualifies it himself later in a way that makes it more modest. Nobody thinks about the Turing test any more, because it is obvious a machine can fake being human. But a general intelligence better than a human at every single thing all the time does not exist, in his account — and he notes no human manages that either. His estimate of the present level is “about doctoral-research level across every field you can imagine”. The terms are laid out: AGI is Artificial General Intelligence, ASI Artificial Super Intelligence, and the singularity is the more classical name for the recursion point.
Security and local models
Miettinen’s security thinking is the most precise material in the episode. Three things:
- The hardware layer. Tailscale and SSH, so the agent can be unplugged; another bot can do it remotely, or the machine can be taken off the network entirely.
- Hardening. He says he has hardened his system through penetration testing with Tampere University’s GPT-Lab.
- The threat model. His argument is that frontier models break everyone’s systems the moment somebody at OpenAI, Anthropic or Google presses the button — if you are not hardening your own system with the best model and a real agent, your security dates from the Windows 11 era.
Nurminen’s line is different but stated just as plainly: he builds the whole thing so that even if everything leaked, his life would not end there.
On hardware, Miettinen is moving his agent from a 16 GB Mac mini to a 256 GB Mac Studio M5 Ultra, because in a small memory Qwen’s Finnish is not good enough and LoRA video models need CUDA. Nurminen runs non-urgent work overnight on a 16 GB Nvidia PC — transcriptions, searches, going through marketing lists: “the tokens cost nothing, the flat just warms up in winter.”
Tools
The episode goes through Codex, Cursor, Claude Code, Antigravity and GitHub Copilot. Miettinen’s practice is model interchangeability: everything is built so the brain can be swapped and the context carried from one place to another. He never drops below medium on the thinking level.
Nurminen recommends Cursor especially to anyone with some coding background, and says he himself wrote perhaps twelve lines of Pascal in the 1980s. Miettinen notes that to an ordinary user Cursor and GitHub Copilot do the same job.
Miettinen also states that Elon Musk has bought Cursor and that OpenAI is therefore walking away from the platform, Musk’s interest being Cursor’s user data. This is a claim made in the episode; the transcript’s figures are not repeated here.
Energy
Mikkonen asks whether energy becomes the real bottleneck for AI. Miettinen’s answer runs in two directions.
Downwards: the quantisation developed as a by-product of DeepSeek showed that smaller hardware and fewer bits will do. Hence he expects edge computing to return — fast memory is the bottleneck, but there is an enormous amount of it scattered across devices; old phones could do inference, load would distribute and the grid would not be strained. Servers come back to workplaces: “the cloud is always a computer somewhere.”
Upwards: Musk’s terafab in Texas, the world’s largest building, making its own chips.
And then a claim he marks as a guess: whoever has the most energy wins this game, because generation capacity converts into intelligence with future models. Mikkonen’s addition is factual and important: chip development is a matter of months or at most years, energy production a matter of decades.
Europe’s second chance
Miettinen’s reading of Europe is more positive than one might expect. Because models arrive cheap and open, third-mover position is a good one: another ten billion of French money would simply have burned in Mistral had Europe pushed harder. His window is short, though — perhaps a year of playing catch-up is affordable.
His analogy is a company-level phenomenon: if you build an enterprise application with AI it is old in six months, so you must be able to build it in one. The same holds at the scale of continents.
Nurminen’s counterweight is human, and he offers it in reply to “you investment bankers’ big game”: because these technologies are becoming everyday and people are learning, the solutions will be found anyway — it is good that this is discussed, but there is no need to worry too much.
How to start
Both give the same structured advice. Miettinen: install an agentic platform, buy a disposable phone number, pair it with WhatsApp and come into the group to learn. If that sounds too hard, take Claude, OpenAI, Gemini or a Chinese model and use it enough that you hit the monthly limit — including by just talking out loud to it. And build the app idea almost everyone has.
Nurminen: find people who know more than you, and be willing to give something back.
Robots, AI companions and regulation
Mikkonen asks whether Europe will ban relationships with AI. Nurminen’s forecast is that Europe will get excited about regulation again, but he does not say whether that is good or bad. His argument is about where a line can be drawn — the same problem as in image editing: the boundary is not black and white, because the thing gradually becomes something else. He illustrates it with robotics, which has been in society for a long time; the question is when there is enough intelligence in it that it becomes something else.
Miettinen considers humanoid robots inevitable, because urban space was built for humans, and notes that Starship’s six-wheeled delivery robots have already adapted to pavements. On addictiveness he is concrete: on a fal.live-style platform you can trivially create a live video stream that adjusts to the viewer — “TikTok is only the beginning”.
Mikkonen raises the other side: if such relationships are banned outright, what happens to treating loneliness in the elderly, or to psychological support. Neither claims the matter is black and white.
Three years out
The closing question is a forecast, and the two answers differ markedly in tone.
Nurminen hopes more Finns will build their AI skills, especially in their own time, because a change of this size needs passion. His forecast is that in three years every Finn will have some deeper interface to AI — how deep he does not dare estimate — and that Finland would be at a higher level than many other countries. Mikkonen’s observation: three years is a shorter span than ChatGPT has already been in use.
Miettinen’s scenario is deliberately provocative and he says he is letting loose: we would be halfway through “the last human ministerial government”, ministries and ministerial advisers would be fully androidised even in the world’s stiffest bureaucracy, and market pressure would have forced companies to follow. His own worry is not about technology but about adjustment: what do those who do not androidise do, when the share of wages taxed for transfers disappears? That, he says, frightens him.
His constructive conclusion is that Finland, as a small nation, can offer digital and robotised services across the border and ally with countries that are not fully protectionist — and that voluntary androidisation, building a human-cybernetic organisation out of several models, agents and capable people, is a healthy thing.
Topics: AI & Enterprise Tech · SaaS & Software · Finnish Economy & Policy · ← All media visits· Talks & Events· Markdown mirror: index.md