Neuvottelija.AI

EP392 · Tools · first published 2026-07-02

Temu-grade intelligence: is AI just a bubble? | Petri Kajander | Negotiator 392

This is a summary on Neuvottelija AI. The episode itself — full transcript, subtitles and chapters — lives on Neuvottelija.com, which is its canonical home.

App developer Petri Kajander and Sami Miettinen work through what AI actually is. Kajander’s term is "temuäly" — Temu-grade intelligence: it guesses superbly and understands nothing. The weightiest part is not the bubble argument but what happens once moderation and customer support are automated — an account taken over by getting an AI support agent to bypass two-factor authentication, and an entire cloud account deleted over private drawings. Plus local models, the realities of the App Store, Gell-Mann amnesia, and two opposed views on whether inference pricing is a bubble. Both speakers are committed parties here, and that is disclosed.

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

Temu-grade intelligence: is AI just a bubble? | Petri Kajander | Negotiator 392

Summary: App developer Petri Kajander and Sami Miettinen work through what AI actually is. Kajander’s term is “temuäly” — Temu-grade intelligence: it guesses superbly and understands nothing. The weightiest part is not the bubble argument but what happens once moderation and customer support are automated — an account taken over by getting an AI support agent to bypass two-factor authentication, and an entire cloud account deleted over private drawings. Plus local models, the realities of the App Store, Gell-Mann amnesia, and two opposed views on whether inference pricing is a bubble. Both speakers are committed parties here, and that is disclosed.

The guest — and a connection worth stating up front

Petri Kajander is an app developer who builds on-device software with no cloud and no advertising. His apps come up repeatedly during the hour. Sami Miettinen, for his part, describes having gone from a middling Python coder to Head of AI Finland at Translink Corporate Finance.

Readers should know this before the content: neither man is an outside observer. One sells apps, the other advises on AI for a living, and they are friends who have followed each other’s work for months. That makes the episode unusually concrete — these are things they have done themselves — and it also means every position should be read as an interested party’s position. Neither of them claims otherwise at any point.

1. “Temu-grade intelligence” — why the word is better than it first sounds

Kajander’s term carries the episode, and it is worth separating from the joke.

“Everything that’s called AI is already in the word: it’s pure Temu, and it has never seen intelligence.”

The pun is cheap. The claim behind it is not, and it recurs in every section:

The model guesses superbly and does not understand what it is doing. With a vast body of statistical material behind it, the guess lands extremely well. But there is no awareness, and nothing that knows what it is doing.

Kajander draws a practical line from this, and it comes from his own work:

The difference is not in the model’s quality but in who checks. From which comes the most usable instruction in the episode: when a tool produces a lot of material quickly, you need predictable safety nets and cross-checks to find the errors — and that, in Kajander’s words, is “the painful work”.

Miettinen adds his own practice: a second model checks the first one’s sources. Adversarial checking is, on his account, a baseline requirement when the material is legal.

The counter-argument the episode makes against itself

Towards the end of the section the opposite view arrives, and it is not reconciled:

Miettinen argues that reasoning models, and the knowledge they generate themselves, are a different thing from mere next-word prediction, and that AI is beginning to solve problems that have been impossible for humanity. His examples are the classic trick questions — counting letters in a word — which a reasoning model now answers correctly, and cases needing ordinary common sense: ask whether you should drive the car to the car wash and the model tells you to walk, because the distance is short.

Kajander does not concede. His answer is structural rather than anecdotal: more data produces better predictions about what already exists, but it produces nothing new. It is not creative.

The disagreement is left open, and it is the most honest moment in the hour.

2. The weightiest section: what happens when there is nobody on the other end

This part is worth reading even if the rest of AI discussion bores you. It is not a forecast but three cases, and together they form an argument.

Case 1 — two-factor authentication bypassed through AI support

Kajander’s description is precise, and he marks part of it himself as not yet confirmed.

Well-known, short and therefore valuable usernames had been taken over like this: the attacker signed in from the same VPN or IP address as the real user, went into the customer support chat — which has no humans in it, only AI support — and got the password reset without two-factor authentication.

What makes the case significant is not the attack but the absence of a defence:

The user had no recourse at all. Their security was in order, two-factor was on, everything was right — and none of it mattered, because there was no intelligence on the other end to explain anything to.

Kajander also proposes a fix, and it is concrete: a hardened handshake in situations where the request comes from the same address or otherwise looks like credential reuse. The episode does not assess it as an expert would, but it is a proposal rather than a complaint.

Case 2 — a word in a video

Kajander says he published a video of his own to a platform through a browser. The video concerned a feature of his product, and it probably mentioned one particular word. The account was suspended immediately. He says he published the whole sequence of events openly, and asks aloud in the episode that if anyone on the platform’s team hears this, the account be reinstated.

Case 3 — private drawings in cloud storage

The heaviest example, and Kajander tells it about somebody else. That person had uploaded their own manga-style drawings to personal cloud storage. Nothing had been published. An automated interpretation flagged one of the images, and the entire account was deleted at once.

What the three cases say together

Kajander does not leave the examples unconnected, and the conclusion is the most important thing in the episode:

When both enforcement and appeal are automated, the burden of proof reverses. You are expected to prove what you did not do — which is practically impossible, because you may not even know what you are suspected of.

He applies the same structure to the Finnish personal identity code: the old code was irreplaceable, so once it leaks the situation cannot be repaired. He notes that a renewable identifier is now arriving.

The argument does not depend on what you think of AI. It is about what happens to due process when the decision is made by a system you cannot explain anything to.

A regulatory side-path

The section also contains an observation that is checkable and rarely heard: Kajander says his app is not available in France because of encryption regulation — going by the book, using strong protection requires official permission. The other blocked country is Russia. He notes that a key pair is already encryption: using public and private keys is enough to bring an app inside the rules.

This is not verified here, but it is a specific claim that can be answered.

3. The bubble — two views that do not fit together

The title question, and the answer is two answers.

Kajander: yes, and tokens are too cheap

His chain of reasoning is clear:

  1. Everyone is rushing to list. He cites a report in which a small block of shares was priced at a valuation absurd in its order of magnitude.
  2. Inference is subsidised. Token prices are artificially low, and he expects a multiple-fold rise — fivefold, five hundredfold, a thousandfold; he does not claim to know the multiplier.
  3. Data centre projects are being cancelled. He says he has heard that a significant share of the projects planned for this and coming years will not be built, because the generators or other inputs simply cannot be had.
  4. The public internet has already been eaten. He refers to an industry chief’s statement that the public web has been worked through — the cream is off, the easy part is done.

All four are unsourced claims or second-hand information, and are marked as such here. As a whole they still form a coherent thesis: the price will not hold, because it is not the real price.

Miettinen: no, and the reason is the wage bill

The counter is the only calculated argument in the episode, and it is worth reading because it says nothing at all about model quality:

The global wage bill is, on his figures, of the order of €54,000 billion, and Finland’s around €135 billion. An AI worker costs tens of euros a month. Even if substitution is only partial, the comparison is so large that current pricing is not self-evidently too high.

The figures are given from memory and are not verified here; the order of magnitude is plausible. The weak point in the argument is the one the episode does not address: the size of the wage bill says nothing about how much of it is substitutable — and that is exactly the question on which the bubble turns.

Neither gives way, and the episode does not pretend to a consensus.

4. Tools do not make a master

The easiest section to carry away, and it rests on Kajander’s own experiment.

Before the pandemic he ran a hundred-day challenge: one video a day. The finding ran against expectation:

The more carefully scripted and produced the video, the worse it usually did. The ones made on a whim took off.

He later repeated the same thing with generative video tools. The conclusion is identical and he states it plainly: a tool does not make anyone Christopher Nolan. It makes doing things easier; it does not produce original ideas. He applies the observation to the present too: looking at content made with generative tools, original ideas are thin on the ground.

The same theme produces the sharpest observation about public discourse in the episode:

The slop loop. What used to be a tweet or a text message is now an A4-length piece written by an AI. The recipient cannot face reading it, so he feeds it to an AI to summarise. Neither end touches the text any more.

Miettinen gives his own example, and it is honest: he says he tries to recall the substance of a conversation from his own memory rather than running the recording through transcription.

Attached is an observation worth sitting with: virtual assistants working in the Philippines now use the same models. Both ends of the chain have automated, and nobody in the chain necessarily reads anything.

Kajander’s cultural note on this is that it is not a technology problem but the disappearance of public space: phoning has fallen away, and there is no longer anywhere to talk without an agenda.

5. Gell-Mann amnesia — the best single concept in the episode

Kajander names the phenomenon directly, and it is the most usable thinking tool here:

You read a piece about your own field and see immediately that it is nonsense. Then you read the same source on a subject you know nothing about, and treat the writer as an expert.

He applies it to AI, and that is what makes it new: the same mistake is easier to make with a model, because the text is fluent and confident across every field at once. When the tolerance is that ninety per cent is right and the rest goes unchecked, the difference is invisible precisely where it matters most.

From this comes his conclusion about expertise, and it runs against the usual expectation: the generalist’s value rises. The model goes deeper than any person, but it does not understand. Someone who sees things broadly notices whether this is even possible — and that is a check depth does not replace.

6. Platform dependence in a developer’s daily life

A shorter but concrete section, and it concerns anyone building on top of somebody’s tools.

Against all of it, his own approach is consistent: everything runs on the device, there is no cloud, there are no ads. That is also what makes his criticism more credible than a complaint would be — he has paid for the choice in availability.

7. Local models and machines

A practical section, useful if you are considering the same. The episode mentions a laptop with 64 GB of memory and a separate small desktop machine for running agents, plus running local agents on your own hardware. The recurring point: memory is the constraint, and its price is not falling.

The section also contains a security remark dressed as a joke — that you should not send Sami anything confidential. It is a joke, but it points at the right question, which the episode leaves open: when agents read the mailbox, what is the status of confidential information?

8. Poisoned material and the fossil layer

A closing observation that is a forecast but a precise one:

The models were trained on material that was not filtered for quality. That is why ideology and false information sit inside them — not because it was chosen, but because nothing was chosen.

And the consequence, which is uncomfortable in the context of this very series and therefore worth marking: content optimised for search engines and for AI sinks into the training material. Kajander sums it up in three words — fake in, fake out — and Miettinen takes it as a question about his own publishing.

Claims presented as claims

What survives the episode

  1. “Guesses superbly, understands nothing” is a usable line. It tells you where checking is optional and where it is not — without requiring a position on what models are.
  2. When both enforcement and appeal are automated, the burden of proof reverses. Three specific cases make it concrete, and the argument holds whatever your view of AI.
  3. Gell-Mann amnesia gets worse with models, because the fluency is identical across every field — which is why the generalist’s value rises.
  4. A tool does not make a master. The hundred-day experiment says the same thing that generative tools are saying now.
  5. The bubble question turns on how much of the wage bill is substitutable — not on the price of tokens. The episode does not answer it, and that is the most honest thing it can do.

How the episode runs

Markdown: index.md · Suomeksi