Summary
Petri Kajander considers what AI actually is, and dubs it ‘Temu-intelligence’ (a pun on the bargain e-commerce site Temu, suggesting cheap, knockoff AI). The conversation moves from local language models and agentic coding to surprising account bans and cybersecurity pitfalls. The two discuss whether the AI bubble can hold, how inference is priced, and why a model can guess well without understanding what it’s doing. Also covered: the realities of launching an app on the App Store, social-media ‘slop,’ and the idea of augmenting humans with AI workers. Finally, they talk about how the open internet is shrinking, and why the generalist’s broad perspective is coming back into value.
Transcript
English transcript derived from validated English subtitles (WebVTT · SRT). Timestamps link to the original video.
00:00 Alright, welcome to the Neuvottelija channel. Isn’t this important to you, Petri Kajander — this “alright” opening? Yeah, we’ve been ribbing each other about this from the start, and there was already some talk earlier about whether it was — alright, who starts, who finishes — but nobody even remembers anymore who the “alright” guy was. Which episode was that? Oh right, that was Neuvottelija, but “alright” really is, I think, the trademark. I’ve tried to get Sami to make t-shirts, and actually a proper, proper brand out of this “alright”— —thing, that “needling is caring.” So you can hear from this that we’ve been friends for years. This is actually your second time on, and back then we talked about libertarianism and your experiences from that era. But Petri Kajander has been kind of a backdrop for me on this journey into AI darkness. So, if people remember Apocalypse Now, where Marlon Brando turned into that kind of deranged guy, you’ve been a bit of a bystander in this during 2026, as I’ve gone from being a kind of mediocre Python coder into becoming Head of AI at Translink Corporate Finance Finland. Congratulations on the new title.
01:10 Thanks. Thanks. And you’ve been hanging around on the sidelines a bit, acting as peer support, and every now and then throwing in some light needling. So I’ve especially liked your X posts, that they’ve always been right there — I remember, was it 25, or it’s been well over 20 years now — you can find everything on X. You can go check the facts there, but that’s when I also realized: vibe coding is Minecraft for adults. That’s right. But as Head of AI, it’s been a long time since it was just vibe coding — now we’re talking about agentic AI programming, and I haven’t even posted this on X yet —posted it. This is brand new. This came up today, and when I had to figure out how to really nail this down properly, I said: it’s Temu-AI. If Sami is now Head of AI, well, I said everything that’s “AI” already has it right there in the word — it’s pure Temu, and intelligence is nowhere to be seen. I do, however, use the Chinese Qwen 3.6 model on my Mac hardware, and from that
02:13 I have to give Petri special credit — back in the ancient year of 2025 I was still a PC guy, and roughly every seven years I make that swing between Windows and Mac, and I wasn’t ready for this yet. But then Claude came along, and I talked to you, and you said just buy the Mac already, you won’t regret it — and that turned out to be right. So in January there was this first real aha moment in life —in life, a 64-gig MacBook Pro, and I think we talked a bit about how memory certainly isn’t getting any cheaper, so I ended up buying a bit more memory. But then that big memory pool — I had to learn the lesson the hard way a few times, working it into unified memory. That’s right. Yeah, and Mac probably stumbled into unified memory purely by luck. I mean, it’s actually a pretty old invention — well, it’s the M-series architecture —the first machine, did that come out in ‘21? Of course it’s been done for longer than that, but back then AI was still kind of — well, there’s always been Temu-AI, I mean AI goes back probably 100 years or more, probably since the invention of the computer. But this agentic intelligence, and then the next phase, was already in February ‘26, when I was talking about how this young Austrian genius, Peter Steinberger, had
03:31 —built this OpenClaw thing, so I ended up asking around about how the Mac minis work, and then I placed an order, and there was about a week’s delay before the OpenClaw release, so I had the Mac mini sitting there humming on the desk, and this 16 gigs of memory is a bit annoying — I should have gone with 32. Just between us, Sami, you kept bragging, saying was it 32 or 64, and now, for the record, it’s actually 16. The 64 of course is my main machine —and mine has the mini, which is 16, but it has the M4 chip. Yeah, and Sami was like, don’t touch this main machine — but now I’m not entirely sure whether that rule still holds, whether it’s already… Well yeah, since as Head of AI you’re the sovereign operator who wants to partition machines like a god, so on the main machine there’s a partition — that’s where Hermes Miettinen lives, on the main machine, and that’s when it uses local memory, because this 64-gig —unified memory is running Gemma 4 26B and then Qwen. So Hermes here has this local agent. I haven’t gotten around to building an Ollama bridge from the mini yet — for you listeners out there [laughter], all of this went right over everyone’s heads, way over the top, so just listen along — because if you ever send an email or anything to Sami, you can assume it’ll automatically end up on the deep, deep, deep web, and everything is compromised,
04:59 which is what Sami said, just put a bit more elegantly. That’s not really how it works, but security has come up in our conversations too. So I asked how this works — the backups — and he said, well, isn’t Time Machine enough, and then you had that thing you used, some CCC, Carbon Copy Cloner, right? No, I didn’t set that one up, but I did set up Bitwarden with a long passphrase note, in a couple of different places, so my security is in order. That security-leak thing was just jealous slander — I’d still recommend, don’t ever put [laughter] anything classified to Sami. Now that we’ve established that you’re not an agentic coder like us Heads of AI, maybe you could tell us why you didn’t stay on the hardware side. Maybe I should say this much about it: since I use, under Sami, OpenAI’s best GPT — the 5.5 — as my main model, it was a bit of a surprise that OpenAI dropped this multimodality, meaning the video side.
06:01 Right, isn’t it the case that you’ve specifically focused on the audiovisual side? Uh, well, Sora — back when it came out. If we go back a bit further, there’s a little story here. Uh, I don’t even remember anymore what TikTok’s original name was, back when it was about music, it was for lip-syncing. It had its own name. That’s when I signed up — it was Musical.ly. Yeah, I had an account there back then. I didn’t do much more with it. But then, was it 2018, before Covid —I just decided that since this visual, video thing had of course already been on Instagram for a long time and was clearly picking up fast, I thought, okay, let’s do a little challenge: one video every day. Mm. And yeah [clears throat], I did it — I had the account, and every day I made one TikTok video, and if you’ve ever made TikToks, you know that every third second, fifth second — usually closer to three seconds — you have to earn the next chunk of attention. And usually the camera angle changes too, if —you look, it’s the same in movies. The camera angle changes pretty frequently, and it’s also a bit of a stand-up comedy trick if you actually want to produce some solid clips that people watch and that take off. So this was kind of my exercise, testing whether I could get something to go viral, what works, what doesn’t. And back then it was a matter of the fancier the script I came up with, the more I thought it was actually good now, and the production values too
07:24 —were in place. Well, those usually didn’t go anywhere, and that’s probably still true on Instagram today, but the ones I made just on a whim, on the spur of the moment, those surprisingly ended up hitting hundreds of thousands, or whatever the view counts were. So this was like a hundred-day exercise. Then I stopped, and I hadn’t touched TikTok in years, but then Sora came along — that’s the OpenAI one — which was never —ever released in Europe. It was US-only, and it was actually really sweet when it launched, at the start — I got an account there, and what I used to do on TikTok kind of by hand, having to edit and think it all through, now you could just prompt it and get it out. And back then Sora was still totally the wild west — there were all the Disney characters and all the movies, you could try all of that out. And it —was there — the 100 that took 100 days on TikTok, well now I was probably doing about ten a day. There was some limit, I don’t remember what it was, but I downloaded all of them. About 300 came out of it. I only tried it for a few months back then. But it was the same cycle: what you’d do by hand on TikTok now came out beautifully through the machine. But like with all of these, having fancy tools doesn’t make
08:44 —you creative, it doesn’t make you any kind of professional Christopher Nolan. That didn’t happen just from using Sora. Mm. It wasn’t like that either when World[unclear], Luma, Kling 1.2, or Kling 3, or whatever else came along — those tools don’t make you any kind of master. Yeah, they make things easier. You can do all kinds of cool stuff, but if you look at Sora videos, or look at any TikToks or Instagrams, original ideas are pretty thin on the ground for a lot of people. Yeah. But then, I was a bit puzzled — you came out with this Apple developer program thing, got results — this app, Läppä. I thought, what’s this for now. But then I’ve started to understand it a bit — that it’s about, say, doing things on the phone’s own hardware, and that’s kind of the point here, we’re not doing this the wrong way here. [laughter] So local models get run — that layers in a bit —more about Läppä. Well, if I take the long way around, if Sami allows it, there’s a bit more of a story to it too. Let’s ask Sami’s permission. Sami says yes.
09:46 Well, it’s not up to Sami. Sami will edit this in here — for the record, for the whole audience, there are no third parties present here, neither as people nor as voices. If anything shows up later, then Sami has edited it in, which of course is entirely possible these days. But the year was ‘23 — wait, was it ‘22, November, when the first ChatGPT came out. It was called ChatGPT half a year after that. It was the 3.5 that came out back then. So that’s what I started using. That summer I took a Python course, this Harvard 101 course, which probably about 2 million people have taken by now. I did the CS50, because it covers the whole architecture already. And I’d actually recommend to Sami that the Python course should be next. [laughter] Yeah. Which you never did? I had a three-week Python phase myself — more on that maybe a bit later. Yeah. And the story continues that I did some stuff in Python, and after that, at some point, it turned into: hey, it’d be nice to build something with some kind of user interface, where you get something visible. So I thought, well, I’ve never built anything for Apple, but it’d be nice if it were on the phone, since that’s a bit more challenging, and of course everyone’s got one in their pocket. So I made something just for my own fun. In Tampere — since we’re both originally from Tampere — there’s Tappara
11:03 —well, Tappara’s alright. Uh, so there was this Lenin Museum. Yeah. Back when I went and registered the nootti.com domain, there was no telling it would turn into the Nootti Museum. So if you type “Nootti” into a search engine, Nootti.com comes up — it was at least at one point the very top result. So, I built it just for my own fun, back when Elon Musk had bought X and everyone was kind of like, what’s going to happen with this now. There was Bluesky, there was Mastodon, there were all kinds of Nostr, all kinds of different networks. Everyone wanted out of X, or at least a lot of people left. I’m still there, and back then I thought — I posted in several places, since Sami was in one place and someone else was somewhere else, and since everyone was a bit scattered, you just had to —go everywhere people were. So I built this little hack of a tool that made it easy to post and customize for each of these places. It wasn’t meant to go on the App Store. It was just for my own enjoyment. But then I thought, I’ll put it up there anyway, and went through the whole cycle. I practiced from absolute zero, and then pushed it to the Apple Store, and the first week I ate my own dog food, using the app myself, and TechCrunch picked up on it, and then it became app of the month — this was back in [unclear], and then 2000 — wait, was it 2000 — now I’ve got my years mixed up, it was probably 2024, anyway six months after that
12:28 —after ChatGPT came out, Nootti was the first — these are all Finnish, now that there’s this sort of anti-American, pro-EU mood going around. So, nootti.com was my first short domain that I managed to find. The next was Nuotit, and then the last one we’ve been talking about was Läppä — so, to sum up the last few years, there’s lappa.com. This has kind of been my AI era, if you will — since ChatGPT, you could —say it’s nonetheless been the groundbreaking thing, whatever the best Finnish word for that is. Anyway, since then I’ve used all of these and noticed where they can be used, where they can’t, where they work, where they don’t, and of course all of that is baked into those products too. But yeah, this has really been my Minecraft era. It wasn’t purely vibe coding. I know what’s inside that code, and I still don’t just close my eyes and push it out —my 30 million lines of code, they’re all floating right there in front of me. [laughter] So, so, well, according to my ideology, if you look at it, everything runs 100 percent on-device — there’s no cloud to be found, there’s no advertising in it anywhere. Yeah. So that fits pretty well with the current business model. Except then there’s the Apple tax. So, so that’s a bit annoying — 30 percent, that costs.
13:54 For smaller, smaller developers it’s 15 percent, so at least there’s that much joy to it — everyone gets to download it. Yeah. But it’s been a fun journey, and now you’ve challenged me — that I really should make this growth journey into an Xcode guy too, meaning an Apple publisher, but I’ve been avoiding that. Maybe some moment when I’ve got a bit of time to vibe-code. So if there’s ever a weak moment in an argument with Sami, you could ask, let’s get me on TestFlight — since Sami’s been bragging, going “look, I’ve got all kinds of fancy apps” — that was back in the time when Sami had, I don’t know if —it’s still on your phone, was it the [unclear] one? Yeah, right, yeah, of course it runs there locally, but as a developer build, not publicly released. And when Sami shared it somewhere on X, or wherever it was, what caught my eye — I mean the logo — how it grabbed attention. Was it 20,000 or 500,000 unread emails? [laughter] Well, now we’re talking about Gmail, which is my
14:54 —kind of junk-catchall for everything, [unclear], so I just put sami.miettinen at gmail.com out there for all the spam subscriptions. You should really put some numbers in there or something like that, and clean them out afterward. But yeah. Well, I had a script running there overnight, starting this March, that deletes about a thousand at a time, so maybe by the end we can get it to a state that pleases even you —too. This unsubscribe script keeps churning away there — I’ll remind you again [laughter], this is Minecraft for adults, meaning a hobby that doesn’t produce any kind of product, but does get you a TestFlight, uh, well, that’s — [laughter] actually, from that we could take a bit of a detour: now, I could think of it this way — we’re living in a world where a Star Trek-style teleporter has been invented, so where before you had —that kind of traveling, where you’d have to take all sorts of trains and cars and planes, go through airports, just to reach, say, the Italian countryside to meet some [unclear]. [laughter] I’d think of something like Italy’s finest tailor, who lives about 70 km outside Milan — and then you’d get there instantly with a teleporter, so everyone could of course jump to these kind of odd places just instantaneously, so, so
16:17 —then these tests and other nonsense, the Apple-tax stuff, would disappear too in this teleporter world, and these inference machines — Anthropic, and Google, and OpenAI, and maybe even X — become candidates for this teleporter layer, where you jump straight to the quality, say straight to your Läppä thing, or to the Miettinen empire, that’s right, and in between there’s no —need for any of this kind of advertising, or websites, or other crap — people just know by some heuristic where the quality is, and they jump straight there. Now we’re at the heart of the matter. [laughter] So why is it Temu-AI? These AI machines have no notion whatsoever of what they’re doing. It’s statistical guessing, completely random, except for Sami, or whoever, who has curated three —layers of brain[unclear]. Exactly. So if some kind of teleporter device comes along, you’d better think twice about whether you want to go through it, because you might come out three-headed, or you might never come out at all, or you might just end up minced into a grinder. So, so this shows up in everything that comes out of AI right now — or, as we’re recording this, with these well-known accounts. For example, hey
17:35 —an Instagram account, which is pretty well-known, at least to some. This mail service got hijacked, and what happened, since this is still such a fresh story that not everything is verified, but the AI Trust and Security teams were apparently thrown out by Meta, Facebook’s parent company, and because of that, of course, from the company, at least —largely, and these well-known accounts, meaning valuable few-letter usernames, have indeed been getting hijacked, so that, uh, the AI has gone in through a VPN — or sorry, the hackers have gone in using the same VPN. Then they’ve gotten into the chat, into AI support, since there’s no humans there. Then they’ve said, well, even though you have —had two-factor authentication turned on, and you’ve got all your own account’s stuff there with top-notch security switched on, this AI has just looked at it and gone, oh well, okay, you seem to be coming from the same country or whatever, and hey, that’s a nice profile picture too. We can just change your email address entirely for you [laughter], and then you can go ahead and reset — reset the password, and that two-factor
19:00 —authentication isn’t needed here at all. So basically, as a user, you had no way whatsoever to reach anyone, even though your security was in order, because there was no intelligence whatsoever on the other end. Mm, so quite a few of these have come up now, really well-known accounts. So, so this is essentially that AI doesn’t know what it’s doing. It doesn’t know what it doesn’t know. And it’s also lazy in the sense that at some point it just —stops doing the work, and that’s how you get this Temu-AI. Yeah, but as we start building better safeguards, hardened guardrails, curated multi-layer brain structures for it, so that dumb things like the same VPN or IP address shouldn’t be able to happen — some kind of hardened handshake ought to take place instead. And by the way, isn’t it just lovely when somewhere over there —at Meta they’re fumbling around at this level, that on top of everything, customer service has always been bad there — there’s no, say, single point of contact from my side. If there’s someone here from Finland listening to this, or wherever this gets translated into other languages, well, this is exactly what happened with my most recent AI thing. The very same post I put up elsewhere, on Instagram and everywhere else — well, this Läppä post on Threads
20:22 —got banned. The only thing I did — my mistake was probably that I used it through the browser, I logged in, wrote the text, put the video up, and in that video the word “no” or something was probably mentioned, since it related to a feature the product had, and nothing else. I mean, it was entirely my own video, and that’s it. The moment I’d posted it, the message came back that the account is currently, like, shelved. Then you can go ahead right away and file an appeal, and that was basically total nonsense, and after that it was just done — and I wasn’t the only one, even the founder of Y Combinator —had the same thing happen to him, apparently — Paul. Not sure if it was on the Instagram or the Threads side, a few weeks later. This is completely random. It can happen to anyone. You don’t have anyone you can go and ask to fix it. You need to know someone on the Meta team. If someone out there works at Meta, please, unlock that account. Uh, I even posted this entire episode over on X. It’s all there. Yeah, exactly, but of course
21:30 —for me, if say YouTube decided to, or somehow put this channel on the shelf, well, luckily I do have friends at Alphabet, so I’d be able to escalate that. Yeah, but this happened, and this has also happened recently, very recently, with Google as well. This one was pretty wild. I don’t know the facts, but apparently it was some Japanese person who had a long history with Google — had Google Workspace or whatever, anyway had Google Drive, had YouTube on there, and a long history with all sorts of stuff, and had uploaded some of their own manga drawings to their Drive, Google Drive, their own personal drive, nothing published anywhere — and the AI had interpreted —one of those images somehow, and the whole account just went, everything happened at once — so if Sami disappears overnight, then Sami’s [laughter] got some manga that ended up on a drive, or maybe there were some slightly too explicit illustrations, some visualization that got made — but from that, well, what could go wrong — Henna Virkkunen and colleagues have now gone and dragged the CSAM thing back into that AI regulation, so — uh-oh — so from there, probably
22:40 —there’ll be more of these coming, as EU bureaucrats, with their finest AI expertise, go off scanning our private traffic to see if there’s anything that should be banned. Yeah. So this trend we’re actually in right now isn’t funny at all. It feels like what we managed to build in the ’90s in terms of an open internet, and these internet walls, which were partly about encryption — France, for example, still doesn’t allow my app, Nuotit, Nootti, to be available there. Do you know why? I’ve got two countries banned. Russia is banned across the board, but France is banned over one single thing, and that’s Nootti. Well, it’s probably some local server requirement or something else dumb. No — if you use, if you use encryption, just normal encryption, the kind that’s totally standard, not high-tech at all — it was maybe high-tech 20, 30 years ago — you have to request permission from some French communications ministry or whatever, and it’s basically like sending a fax and letters [laughter] —and stuff like that. And that’s the reason so much is missing from the French App Store. If you go by the book, you have to get permission from them before you can use any kind of encryption in it.
23:50 And yeah. Well, Nostr keys are like Bitcoin keys. So there’s private — you basically use a PKI, private and public keys, and that’s already encryption, and because of that you need to get permission from France before it’s allowed. But this was kind of a long intro leading to the point that we’re now sliding toward — under the guise of protecting children, we’re heading toward a point where soon you won’t be able to go anywhere and read Iltasanomat unless you’ve put your passport up somewhere for verification. Yeah, and that’s the thing — again, since we just talked about this AI capability to hack in, isn’t it nice how it then spreads out into the world and everything — and here in Finland we don’t yet have, or, well, I think that unique ID, or the renewable personal identity number, just came in, so at least those can be renewed now. Because right now our old-style personal ID number is basically irreplaceable — once it truly leaks, once and for all, that’s it, that’s game over. Yeah, the world is full of these, —so much so that there might even be an entire nation’s passports and all its data leaked, and it’s happened in more than one place, more than once, and then there’s really nothing you can do about it as a person, and in the worst case, bank credentials get opened, loans get taken out, all kinds of things get done, and then, in effect, the burden of proof reverses. You have to prove what you didn’t do, which is pretty much impossible since you don’t necessarily
25:16 —know what’s been done in your name. Yeah, exactly, and undoing this is a hell of a job. So, as a responsible AI trainer, so to speak, I’d also hope that, for example, Stripe has started offering agents limited transaction rights — yeah, in principle, as an experienced bot trainer I’d want that too, but that opens up yet another can of worms, since in a way you get —with that authorized agent permission, you’d probably still enable proper scams to be pulled off, and that brings us to this security question, and to why it’s Temu-AI. So here we’re not talking about Heads of AI, but about people who just can’t — meaning, decision-making [laughter] — yeah, headless — and Temu-AI also in the sense that you can never, in old-school coding, —basic coding, it’s a certain way, I mean it’s completely — you know how it comes out: if this, then this; if there’s some bug, it’s still predictable. And this doesn’t work at all, not in the case of AI, and that means you can never be sure what’s going to come out of it, whether it’s what you wanted, or whether something completely different comes out. Unless you build guardrails around it. Mm.
26:37 So, in the right hands, AI is very powerful, but you have to realize that it’s Temu-AI. There’s no real — there’s no consciousness of any kind behind it. It doesn’t know what it’s doing. It guesses well, and since it has a huge amount of statistical data underneath it, it guesses extremely well. If it’s marketing, or sales, it doesn’t matter so much, since you can check the text yourself. But for example, if you’re building competitor comparison pages, it would be good if the facts were right, so you don’t make false claims. But like, this is —easier to fix. But in certain critical matters, where it touches personal data, medical history, health records, any area where it needs to be right and not wrong — it matters quite a lot whether you have cancer or don’t have cancer, at least on a personal level. And if there’s wrong information there, the consequences can be pretty severe. And this brings us to the point that if you’re using AI tools, they can easily produce a huge amount of data, and they can also find things that would easily slip past people — but then you have to be able to build a certain kind of predictable guardrails around it, and cross-checks, so you can find the errors and see whether it’s actually gone off the rails, and that’s the painstaking work, and this is good
27:54 —news for the everyday worker, for those who know what they’re doing, but also for generalists. Because if specialists have long been treated as gurus, for being deep in some specific area, say SaaS company sales, well now we’re moving to a point where ChatGPT the specialist can go even deeper there, because it just has a broader knowledge —base at its disposal, but it doesn’t understand what it’s doing. And that’s when the generalist, who sees things more broadly, can take a step back and ask whether this is even remotely possible. I’ve been reading a lot of articles recently, and there’s one — I don’t remember the whole context anymore, but it was about some kind of green transition, and it had to compare emissions on some massive scale, a global scale. And then in it —there was a sentence saying this amount of emissions is equivalent to the annual emissions of five cars. Mm. So, so basically, if you don’t notice that — there’s just no sense of proportion here at all, things aren’t in any kind of order whatsoever.
29:09 And anecdotally, everyone can spot this easily in the areas where they consider themselves a certain kind of expert. There, you see right away, oh, it’s talking nonsense. There’s no sense to this. But then you start treating it as an absolute guru on things you have no clue about. And, and, and this is exactly the danger, when we sort of start — well yeah, I’ll sign on to that, but on the other hand, since I’m used to this incomprehensible human hallucination in Finnish society, I don’t see this as that big a problem, because, in a way, the collective Finnish ability to spout completely deranged nonsense, based on nothing but, say, some left-wing tradition, has been so —so tiring, in a way, that if the tolerance there is like 90 percent wrong and nobody notices the garbage, well, if then in some matters the AI hallucinates something, I’m — I’m not that harsh about it. But like you said earlier, this kind of adversarial setup, where a second AI checks the sources, is basically standard practice — for example, when I do legal work with a pretty good tech stack, there too you get a bit of intolerance built in, so it always checks, say, the Supreme Court rulings, and a second AI checks
30:19 —that they’re real and not hallucinations, and that’s how the links and sources get checked. Yeah. But then, this also brings us to: don’t ever send Sami anything confidential. Because if it’s from Sami, it’s curated, completely secure, physically, here — I mean, this is Petri the amateur talking, who’s never even had OpenClaw running, so he’s just talking, just slandering here, he has no idea how brilliantly this stuff of mine is built — or maybe you know a bit —but this curated, three-layer Andrej Karpathy vibe model — the stuff Sami always posts here in our internal WhatsApp [laughter] group — I’ve usually put those up a week or a day earlier, and then started using them myself, so there’s still a long backlog that hasn’t been done. Post it yourself there and you’ll get the attribution. But yeah, one such example could be some Andrej Karpathy insight, say that vibe thing, and earlier, say, the Ralph loops, and things like that, —those important things for us Head-of-AI types. So I’m probably the headless AI [laughter] — yeah, but maybe you’ve just got some good bot that feeds you scraped stuff from somewhere, and how would I know your sources — I just see the end result: a really well-timed X post comes out. Should this stay in the inner circle, or could you tell us what the source is? Go ahead, let it rip.
31:36 It’s just X. So even though everyone bashes X, and everyone’s left it, the fact is there isn’t anything better. Meaning, news everywhere — it shows up on X first. There are quite a lot of builders, creators there. Yeah. The atmosphere isn’t nice. It’s not — it’s not the same as it used to be. It was actually interesting looking back a couple of days ago, doing a bit of a retro — I had to find one tweet, and it was from ‘22, a little before, a few months before —this first round happened. Mm. It was pure bliss reading that feed — if everyone remembers, if you were on Twitter back in year one, two, back in the stone age when it started — it was never quite what Jack Dorsey’s [unclear] intended it to be for. This passport business was already a thing back then too, yeah. I didn’t go quite that far back. I only went back to ‘22, but what caught my eye was that there weren’t AI-slop texts, these terribly long rambling threads. There weren’t all those images, there weren’t all those videos. It was a lot calmer. Sure, there were memes, sure there was just as much junk in the same way, but visually, and overall, it was a lot calmer.
32:47 I won’t say whether it was better, but it’s like slowly boiling the frog. Mm. Or was it — is it OpenClaw that’s being slow-cooked here — so, so you don’t notice it, and this is roughly what’s happened to all of us in the same way: this AI slop has crept in bit by bit, and you barely notice it — everything is completely fake, we’re pumping out a huge amount of stuff, what used to be a tweet or a text message, well now —it turns into a full page-long blog article, which was of course written by AI, and then Sami receives it, and then Sami can’t be bothered to read it, so it goes back to Sami’s AI, which says, okay, just give me back the one-liner. And this is now, in a way, what our social world looks like. There’s fake stuff over there, fake stuff over here, and then to get to this humblebragging and LinkedIn — whatever you find on LinkedIn was stolen from X two weeks earlier, and whoever your guru is there, whoever you follow. And this was actually — we can put this in if Sami has some kind of episode show notes. On my way here, I found this article. It’s probably a week old.
33:58 So, in the Philippines and virtual assistants there — nowadays they also use ChatGPT, or Claude, or whatever else. The local professor, or any professional, or anyone building a career — nobody has the energy or the will, or just doesn’t feel like writing these long rambling posts on LinkedIn and commenting on 50 articles a day, so all of that is done in the Philippines with ChatGPT’s help. So this kind of fake-it performance effort — we’re all already pretty deep into that, —pretty deep. You can’t always blame it on Temu-AI. Well, that’s how it is, of course — some of us write a bit better skill prompts for the AI, telling it to write like a human, don’t use em dashes, and use proper Finnish, and so on. And then this loser crowd just dumps it straight onto the first-tier Temu-AI inference engine, say Claude, into the chat: “write me a LinkedIn post,” not
35:01 —like that. So of course you use skills, meaning you train these AI workers, and also, if you don’t want to go quite that far — setting up a Hermes agent or running OpenClaw, I know that’s hard — well, similarly, there’s my own subordinate that runs on a Hermes agent. So, so, right, so then you can just put those skills into, say, Claude too. So, if we translate this into plain terms, it’s —a text file that says do this, don’t do this, remember this, and then you get this long laundry list, where half the things still get left out, and then Sami’s AI at some point just decides not to do any of them, skips it, and sends everyone Sami’s latest photos. No, it doesn’t actually do that. The security’s been hardened, hardened. But anyway, into the slop, generally — if you don’t want to be exactly the type Petri’s parodying here, meaning someone who —puts out something with em dashes and bullet points and, uh, made by AI. So, using weird Finnish words to make a LinkedIn post — knowing that, put some skill prompt in between, so you don’t come across as a total zero.
36:13 Yeah. Well then, of course it’s worth making a slightly shorter version for X and a slightly longer one for LinkedIn, and using Sami’s natural style and things like that. That alone already makes it a lot more human. Did you notice? [laughter] Sami just confirmed it — all of this is fake, nothing is true, it all comes from some two-line prompt, and then Sami shows off with that fake stuff. I have this long, original thought that I’ve divinely turned over with the help of various assistants of course, and sometimes I even give you a call, saying, hey, I’m going to scrape your brain — so it calls every couple of weeks, so it can go show off for the next couple of weeks. [laughter] But I’m not recording this here in a way where I could just drop it into your Läppä, or somewhere, or — well, you actually recommend feeding it all into MacWhisper — so, no, I’m not doing that this time, I’ll try to just remember it with my own bare-metal brain, whatever it said, something sensible, or was there any sense to it at all [laughter] [deep breath] — yeah, but human-scraping, that’s — that’s — that’s a hit
37:18 —so, calling people up like the old uncles used to call each other, that’s human-scraping. Could that be a new term here? Yeah, actually, just this past week, I even called people completely at random, and it’s been a bit of a little test too, seeing who picks up — and surprisingly some people even pick up right in the middle of lunch, since calls come in so rarely nowadays. It doesn’t mean someone’s died if you call them. It’s totally normal to call and ask, hey, how’s it going, and it doesn’t even need to be about anything serious. But yeah, this has kind of fallen by the wayside for all of us, though of course nowadays it’s gotten to the point where you don’t even answer calls anymore. Yeah. We’ve kind of lost this public space. There’s not really any place anymore where it’s nice to chat, where you can just talk. We used to have blogs on the internet, back in the day. Podcasting started up around then too. There was, in a way, this sort of grassroots, ground-level activity and all kinds of things like that, and now that’s not really there anymore — it’s all under these big platforms, and soon you’ll need a passport for everything on top of that.
38:20 Yeah. And then once your identity gets tied to it, you stop saying anything at all. So we’re basically right back in China and the Soviet Union. But then there are these oases in the middle of the desert, like the Neuvottelija inner-circle WhatsApp group, where pure intelligence gets to shine, even from Sami’s side there — I mean, Sami’s fair game for teasing, and Sami has never once betrayed his master by leaking —sensitive information, even though you’ve tried a few times here, who knows [laughter]. I’ve got this thing — did I already say curated three times, that brain [laughter] — yeah, I’m quite sure what’s in Sami’s mug, but, uh [laughter], yeah, but it’s still educational — I mean, OpenClaw is kind of — since Steinberger’s the one who put it together, the basic architecture there is still a bit of ours —that command-line style, so you always have to think a bit about what state it wakes up in, and whether some new kind of setup has come along, where the skills just say box, box, box. And by the way, speaking of these teleporter front-ends, there was this whole drama a couple of months back when Anthropic decided that this communist queue-based economy doesn’t work anymore, let’s switch to this kind of capitalist token economy — that was pretty fun
39:42 —to watch, this script of Sami’s, when first the initial $50 went on Monday, and then I put in another $50, and that went the same way, poof. So then I just removed Anthropic entirely from the API keys and moved over into the open-source world. But now the same thing has happened, dramatically, over at Microsoft’s GitHub Copilot — they switched over to a model like this back in June, and now we’re only on day two since —this is being recorded, so — the moment you say “hello world” or “good morning,” it basically snapped shut right away. Yeah, and before that Google had this Antigravity thing, where you could basically run it in a similar style to Copilot, and people noticed and started hammering Antigravity’s free tier, so Google’s solution was to just ban, ban everyone in one go. So basically, I’d rather pay that $90 in pain money than —have it just walk away entirely. So this is really the harsh reality of today’s AI world, that both outcomes are possible. One is: oops, forgot to set the token limits, and now, hey Sami, you’ve racked up €500 in costs from your scripts. Or the other is: you’re now illegally running this on some monthly subscription that should have been paid for in real money, so we just decided here in California to ban it outright — so both of these are genuinely possible outcomes.
41:07 And actually, these things are happening even right now — there have even been companies of 50 people, 30 people, running on a single Claude seat. Yeah, and the whole company just [scoff] dropped out on a Monday, or whatever day it was. So, so Anthropic has been very, very sensitive about this — they’ve got a lot of heuristics checking how many different places you’re logging in from, whether you’re using the same browser, the same machine, the same location. Is it roughly the —same typing speed? Meaning, is there AI pumping stuff in there, are you running multiple accounts? And the reason behind all this is that all of these companies are now racing to get to an IPO quickly. Just the other day, there was a $10 million stock stake being sold off at a valuation of a trillion — trillion, in the American sense of the word. —complete nonsense, who can really tell if Anthropic or any of these companies is actually worth that. But sure, it’s nice to say the valuation is that, though you can of course ask whether they’ve read their Buffett, when everyone’s bragging about how high it is. But I’m currently in the position of thinking these tokens will probably get more expensive, since they’re being subsidized so heavily right now. It could be five times as expensive, could be 500 times, could be a thousand times. Mm.
42:26 Everyone wants to show off right now, as long as it lasts until they get to that IPO window. But I’ve already seen this before, say with Microsoft — if you remember, back when Sam Altman, the founder of OpenAI, had some internal drama going on, and then Microsoft came in and rescued the situation, and then he came back into the CEO seat. Back then, in that corner, Microsoft put in tens of billions, invested on pretty generous terms. And now we’ve seen that Copilot doesn’t get used, even when it’s being paid for, so real usage has stayed really small, and this shows exactly, in a way, that some of these data center projects — maybe half of them — that were supposed to come online this year or in the next few years, aren’t actually going to get built. They just can’t get the generators or whatever else they need, so this AI bubble is probably going to start rotting. So Sami really needs to get that thing onto TestFlight pretty fast, or else he’ll end up just —twiddling his thumbs. I don’t really believe that, though. I mean, I think these app channels get some of the blame too — it’s that boring filler right in the middle of the teleporter, on the way to teleporting to quality, meaning Neuvottelija, or Sami Miettinen, so it still needs that stuff in between. [laughter] Check back on this episode next year to see whether Sami’s tried to IPO some
43:51 —company. [laughter] Yeah, but I wouldn’t be that worried, actually. I have this thesis: the global wage bill is around 54,000 billion euros, and in Finland alone it’s 135 billion euros. So, so this AI, this AI labor, and Sami’s AI — is it still only costing like twenty-something a month, if we —leave the hardware out of it, since hardware doesn’t count here. How was it [laughter] — it was $50 in the morning before we even got to lunch, and then another $50 got thrown in. That was back when Anthropic decided to stop cheating us on this, or actually started charging real, proper prices. Not this whole all-you-can-eat buffet thing that belongs to Sami’s AI, free brains for eternity — and Sam Altman is now like my friend, because, at least at this stage of the game, his AI is generously offering the same brainpower at this kind of cheap price. But still, though, uh
44:51 —this thought, kind of a one-billion-dollar valuation thing [laughter], so this labor value — I mean, that’s the real steal for Sami’s AI here — that if it’s now just costing tens, or even hundreds, or maybe a few thousand euros, then, trained by me, that AI’s value ends up matching a kind of lazy, pudgy young employee, and it’s already possible to replace some middle managers with these kinds of Sam-type —work flows. So, so yeah, this does have value — I don’t buy into the idea that data centers are just going to gather dust and that this whole inference thing is a bubble. It’s just that the untalented, who don’t know how to properly direct that intelligence, call it AI and end up with Temu-AI instead. So we’re in a Temu-AI situation, where TikTok — the US TikTok — apparently went to Oracle, and Oracle’s Larry Ellison said just the other day that the public internet has already been picked clean and eaten up, and yeah, I’ve seen it too — everything has basically been ripped. Anna’s Archive, where all the pirated stuff in the world is, all the music, all the books, all the videos — it’s all already been sucked into it. It’s in — it’s inside all of these models. So
46:06 —and all that’s left now is this private data, the stuff that isn’t public, and that’s the situation now. But then we get to the point that, since it’s Temu-AI and it doesn’t understand a single thing, we’re already at the peak moment of the bubble. No, and all the easy stuff is already gone. The cream’s already been skimmed off the top. Well yeah, but that’s the old line that AI is just a better next-word predictor, that’s all. So, we Head-of-AI types have had these curated three-layer data structures, and then some reasoning model on top of that, and, for heaven’s sake, it’s really not quite like that. I mean, take Humanity’s Last Exam, for example — scores there just keep climbing past 50 percent now, and that doesn’t happen by just scraping together some Stack Overflow leftovers — no, they’re actually using this reasoning, and creating their own knowledge with this intelligence, meaning that [clears throat] AGI is about to come pushing through soon, meaning AIs are able to solve problems that have been impossible for humanity all along, and everyone can try these out themselves and maybe come up with new ones that haven’t been used to catch it out. So, in English, people have asked the AI, how many D’s are there, how many days of the week — in English — how many of the days of the week contain the letter D
47:30 —or, uh, whether I should walk or should I drive, if I’m taking the car in — early 2026 kind of examples, so, so, meaning Temu-AI — I can drive my point home right here: if you ask the AI something that should be really simple, like when Sami says — you say, hey, I need to take the car to the car wash. Should I just walk there or drive it, and out comes this reasoning every time, that, well, you —should walk, since it’s such a short distance that there’s no point driving the car to the car wash — or that there aren’t enough D letters found among these. But well, 4.8 has probably improved on that, since some of us use these models right on release day, and some probably still use, say, Claude Opus something, or Sonnet from ancient times, and then bring up these urban legends, —about how that’s happened at some point. So, so if you understand the basics, this is actually very simple: if you just cram in more of the same, it doesn’t change the outcome. Since there’s no understanding of any kind there, adding more of it doesn’t add anything more of that. Yeah. With a much larger amount of data, sure, we get better predictions about what already exists, but it doesn’t create anything new. It’s not creative, and this is — well, but then, say, Elon Musk goes and builds Grokipedia, where
48:52 —there’s kind of curated knowledge, some of it, say, information basically lifted from Wikipedia — well, yeah, but it’s also curated in its own particular way, without that whole left-wing-bubble nonsense — although unfortunately this too has [laughter] gone the same way, yeah, that we’ve got clever, clever people over on the China, Russia, and Iran side, so it’s maybe even 25 percent of these AI models, since they’ve been trained on that material, —so there it’s been poisoned, there are, there are these kinds of poison pills in there, there’s ideology inserted somewhat in a direction that isn’t factual and of course suits the value system of these dictatorships, and these are inside all of these models. Not because they’ve been specifically loaded in on purpose, but because the underlying material they’ve been trained on hasn’t been qualified in any way. So in there goes —Lenin’s collected works, into the same category. Sami thought about that with the blog post, and then the pinnacle of humanity’s collective wisdom.
49:57 Exactly. And then future generations will all suffer for it. [laughter] Yeah. But that’s actually what I do. I’m contributing to building this archaeological fossil layer of knowledge with this Neuvottelija empire, so that these kinds of GEO-optimized structures then get soaked up as training material there. Fake in, fake out. Right, and actually I’d suggest we go over there, since this ties into the whole website technology thing, so let’s go into that on the inner-circle side later, and there too we’ve had various discussions about how this website technology should be organized in this AI age, and how it maybe used to be done back in the day, but let’s save that for the inner circle. Well, right now I’m sensing a bit of Luddism and this AI-nihilism —coming from you, but hey, no harm in it — let all the flowers bloom, and let those who know how do it, and those who don’t, well — specifically, you’ll find it on the App Store [laughter], three of them: worth checking out are Nootti, Nuotit, and Läppä, and you’ll find generative AI there — Läppä can produce a pretty nice-looking video from a single recording, which you could push straight to LinkedIn, and the captions come out ready-made. Yeah, you can put all kinds of cool stuff in there, and it makes
51:13 —backgrounds, and yeah, it does use AI in the background there, but it’s all on the device. So it’s great that you’ve had the patience to build these on such archaic App Store technology. Yeah, but should we have some final wisdom to wrap this up? I do think this kind of “androidization” makes sense. So back in December 2025, I built this kind of second brain, a better notebook, a Second Brain, and from there, pretty quickly, through these agentic friends of mine, I ended up concluding that it’s better to augment the human into that —AI workforce and their army, and then of course build curated structures in there, so they deliver quality and don’t hallucinate, and then can potentially patch small gaps in these foundation models on the fly. By the way, do you remember, back in 2000, when there was that earlier bubble — or have there been several bubbles in between by now — there was a company called something like Second Head.
52:18 I don’t remember Tom [unclear]. Yeah, I remember it was called Second Head — that was the name of the service — and they had this marketing stuff, and one item was a pillowcase. I don’t remember exactly, but I think the slogan was something — correct me if I’ve got this wrong — but it was “rest your second head here.” [laughter] Yeah, it’s a bit in the spirit of Being John Malkovich, kind of the same thing, that new is old, old is new. Yeah, but we really are living in an age of voluntary —“androidization,” and you might as well just join the cult, right? But I’d suggest we jump over to the inner-circle side now. Let’s talk about the development of website technology, and, well, I do have to reveal this much: Petri and I have been racing each other on these website agent scores, and you threw the first shot at me, and I got past you, but now you’ve gone and passed me right back, and
53:21 —that’s because my main website, neuvottelija.fi, runs on Lovable’s Supabase cloud, which is honestly pretty crappy — but for this kind of thing, could it be, say, Cloudflare, or are there maybe still some archaic WordPress, PHP setups in use — let’s take these website topics over to the inner-circle side. Would that work? Yeah, but let’s also bring in the bigger picture — sales and marketing, how that’s handled these days. Now you —managed to double your subscriber count here. Hey, now you’re hearing that modern marketing, combined with website and DNS technologies. That’s already AEO and GEO. Great. But we’ll see you over on the inner-circle side. Thanks, Petri Kajander. We’ll keep doing this human-scraping thing after this too, and, uh, if you’ve stuck around watching or listening this far [music], go ahead and hit —subscribe to Neuvottelija. Thanks, Sami.