{"title": "AI Coding at the Highest Level | Markus Hav", "titleOriginal": "Tekoälykoodaus huipputasolla | Markus Hav | 371 Neuvottelija", "episodeNumber": "db-666", "guest": "Markus Hav", "publishedAt": "2026-02-09", "duration": "01:04:46", "durationIso": "PT1H4M46S", "youtube": "https://www.youtube.com/watch?v=BpZTc2akIjQ", "fiCanonical": "https://www.neuvottelija.fi/fi/episodes/666-tekoalykoodaus-huipputasolla-markus-hav-371-neuvottelija", "originalLanguage": "fi", "format": "full", "topics": ["ai_enterprise_tech", "saas_software"], "description": "AI coding at the highest level | Markus Hav | 371 Neuvottelija. Markus Hav describes the cutting edge and sheer intensity of AI-assisted coding. AI agents are reshaping thinking, work, and learning at an accelerating pace. Hav emphasizes context, memory, and intent, along with what he calls \"codumentation\" (a blend of coding and documentation) in the coding process.\n\n2026 is a turning point for AI coding, especially as Anthropic and Google roll out top-tier tools.", "chapters": [{"t": "00:00", "title": "Anthropic's Claude Code cleans up in AI", "seconds": 0}, {"t": "00:19", "title": "Markdown runs AI, but where's a good editor?", "seconds": 19}, {"t": "00:30", "title": "Claude Code Max token budgets: Markus €400/month vs. Sami €100/month", "seconds": 30}, {"t": "00:57", "title": "Was Lex Fridman's AI episode boring?", "seconds": 57}, {"t": "01:45", "title": "We're not the crazy ones - the world is diving headfirst into AI", "seconds": 105}, {"t": "02:06", "title": "AI's exponential growth in 2026 vs. the calm of 2025", "seconds": 126}, {"t": "02:34", "title": "Claude Code's tokens are doing a lot of real work", "seconds": 154}, {"t": "03:24", "title": "Predicting Google and Gemini's rise, and why it was obvious a year ago", "seconds": 204}, {"t": "04:09", "title": "Google's data, chips, and models combine into an unbeatable whole", "seconds": 249}, {"t": "05:04", "title": "Installing 'Open Claw' on Sami's Mac Mini!", "seconds": 304}, {"t": "05:42", "title": "The human-risk factor, and why being first mover doesn't always win", "seconds": 342}, {"t": "06:34", "title": "Credit cards, access rights, and agents' hunger for data", "seconds": 394}, {"t": "07:01", "title": "Security had to be rethought from scratch", "seconds": 421}, {"t": "07:21", "title": "Samantha the bot helps Sami, and earns a HoxHunt training session as a reward", "seconds": 441}, {"t": "09:08", "title": "Are agents loyal, or do they just respond to incentives?", "seconds": 548}, {"t": "09:54", "title": "Persistent memory makes conversation surprisingly human", "seconds": 594}, {"t": "10:44", "title": "The agent isn't the same from one conversation to the next", "seconds": 644}, {"t": "11:26", "title": "The context window as memory, and clearing it out", "seconds": 686}, {"t": "11:57", "title": "The million-token era and its limits", "seconds": 717}, {"t": "13:06", "title": "Compressing memory and the analogy to human memory", "seconds": 786}, {"t": "13:51", "title": "The 'Second Brain' and the dream of unlimited context", "seconds": 831}, {"t": "14:19", "title": "Memory retrieval makes the agent more efficient than a human", "seconds": 859}, {"t": "15:00", "title": "The Ralph Wiggum model: a dumb but persistent agent", "seconds": 900}, {"t": "15:45", "title": "A simple to-do list instead of orchestration", "seconds": 945}, {"t": "16:25", "title": "Perfectly structuring context solves almost everything", "seconds": 985}, {"t": "17:31", "title": "Ready-made skill.md files and maturing tools", "seconds": 1051}, {"t": "18:12", "title": "You can't try everything, and that's okay", "seconds": 1092}, {"t": "19:01", "title": "The small-mover edge: the advantage of unlearning", "seconds": 1141}, {"t": "20:19", "title": "New employees get the most out of AI tools", "seconds": 1219}, {"t": "21:34", "title": "The hammer problem: getting stuck on old technologies", "seconds": 1294}, {"t": "22:12", "title": "Even Stone Age systems like COBOL are opening up to AI", "seconds": 1332}, {"t": "23:06", "title": "The cumulative effect of small improvements on society", "seconds": 1386}, {"t": "23:30", "title": "The power struggle among foundation models, and how fast it becomes obsolete", "seconds": 1410}, {"t": "24:39", "title": "One-shot coding and learning through mistakes", "seconds": 1479}, {"t": "26:12", "title": "Self-reflection as a new model capability", "seconds": 1572}, {"t": "27:15", "title": "The agent recognizes itself and improves itself", "seconds": 1635}, {"t": "28:07", "title": "Resilience to errors in coding", "seconds": 1687}, {"t": "29:20", "title": "The interface shapes the learning experience", "seconds": 1760}, {"t": "30:12", "title": "Different models' error profiles", "seconds": 1812}, {"t": "31:24", "title": "Google Antigravity and practical workflows", "seconds": 1884}, {"t": "32:17", "title": "Frontend, backend, and database realism", "seconds": 1937}, {"t": "33:15", "title": "The demo effect and frustration with new tools", "seconds": 1995}, {"t": "34:26", "title": "Firestore vs. Supabase", "seconds": 2066}, {"t": "35:26", "title": "Lovable as training wheels for a fast start", "seconds": 2126}, {"t": "36:25", "title": "The difference between documentation and a spec", "seconds": 2185}, {"t": "37:48", "title": "A good spec is half the work already done", "seconds": 2268}, {"t": "39:19", "title": "Scoping intentions to keep agents under control", "seconds": 2359}, {"t": "40:05", "title": "The idea of an intent database instead of GitHub", "seconds": 2405}, {"t": "41:24", "title": "Understand why you're building, not just how", "seconds": 2484}, {"t": "42:13", "title": "Markdown and Obsidian as the foundation of a Second Brain", "seconds": 2533}, {"t": "43:11", "title": "Microsoft fumbling, and Anthropic's Cowork for Excel and PowerPoint", "seconds": 2591}, {"t": "44:16", "title": "Anthropic as the strategic winner across every cloud", "seconds": 2656}, {"t": "45:31", "title": "OpenAI won't die, but its direction is getting blurry", "seconds": 2731}, {"t": "46:00", "title": "Elon Musk, SpaceX, and xAI's inference in space", "seconds": 2760}, {"t": "48:36", "title": "The simulation idea and NPC humor", "seconds": 2916}, {"t": "50:35", "title": "Neuralink and a direct connection to AI", "seconds": 3035}, {"t": "52:40", "title": "Unpleasant people with enormous influence", "seconds": 3160}, {"t": "53:55", "title": "How coding will change over the next six months", "seconds": 3235}, {"t": "54:41", "title": "Chinese models and the agentic surge", "seconds": 3281}, {"t": "56:39", "title": "Orchestrating hundreds of millions of tokens", "seconds": 3399}, {"t": "57:36", "title": "Cheap models enable massive agent swarms", "seconds": 3456}, {"t": "58:31", "title": "2026 as a turning point in human history", "seconds": 3511}, {"t": "59:12", "title": "The first practical step into agentic work", "seconds": 3552}, {"t": "1:00:04", "title": "Agentic browsers and the need for caution", "seconds": 3604}, {"t": "1:01:23", "title": "Lovable and Firebase for a low-barrier start", "seconds": 3683}, {"t": "1:02:13", "title": "The realization that anyone can build", "seconds": 3733}, {"t": "1:03:35", "title": "Building together as a catalyst for learning", "seconds": 3815}, {"t": "1:04:07", "title": "The death of interfaces - is UX dead?", "seconds": 3847}], "sourceTranscriptType": "finnish-human-curated-subtitles", "subtitleMethod": "claude-fi-to-en-cue-preserving-translation", "provenance": "local SBV curated by Sami Miettinen (Neuvottelija-Subtitles folder)", "slug": "666-tekoalykoodaus-huipputasolla-markus-hav-371-neuvottelija", "guestsCanonical": ["Markus Hav"], "guestPages": ["https://www.neuvottelija.com/podcast/guests/markus-hav/"], "page": "https://www.neuvottelija.com/podcast/episodes/666-tekoalykoodaus-huipputasolla-markus-hav-371-neuvottelija/", "markdown": "https://www.neuvottelija.com/podcast/episodes/666-tekoalykoodaus-huipputasolla-markus-hav-371-neuvottelija/index.md", "captions": {"vtt": "https://www.neuvottelija.com/podcast/episodes/666-tekoalykoodaus-huipputasolla-markus-hav-371-neuvottelija/captions.en.vtt", "srt": "https://www.neuvottelija.com/podcast/episodes/666-tekoalykoodaus-huipputasolla-markus-hav-371-neuvottelija/captions.en.srt"}, "transcript": "[00:00] From Anthropic's point of view, both AWS and GCP have invested in them, and then there would still be Azure too, so they'd be on all the cloud platforms. They've played their cards pretty well. And Google, of course, has played its cards pretty well, being an investor there, but I mean, and Amazon too, but [laughter] but, in a way, all the good guys, all the good guys. Hey, Perplexity, which I use, has Jeff Bezos as an investor too. You always have to pick... there's always some good guy in there, right? [laughter] Yes. But why do we have a code editor but no proper markdown editor? Many, many licenses, and I've used an API key. Already last year I was spending about 400 € a month. So: Markus Hav, Hoxhunt's AI guru, and previously a developer of the Inderes platform, and in Mikko Alasaari's Agentics Finland, alongside me, perhaps one of the loudest agitators there on WhatsApp. Before we switched this on, we were saying we'd feel like\n\n[01:00] doing a six-hour episode, for once. What do you say, Markus Hav, shall we do six hours of this? Let's start with an hour and [laughter] see how far we get. Yeah, yeah. It would be fun though. Those Lex Friedman long episodes, where you go really deep and really linger on the tangents of tangents — that would be great, especially here with AI, where you could focus on absolutely any single topic for as long as you like. So yes, I'm totally up for it, but right now the calendar doesn't quite allow it. Yeah, let's do that next time then. But about Lex Friedman — we had a chat about it and went out for plank pizzas, and we had both noticed that Lex, after a long while, put two AI guys on to talk. Then I said I put it on just before going to sleep, listened to it, and they were just blathering on about books — that everyone should write their own kind of fundamental LLM, and then, in my opinion, it was just boring [laughter] So, as context, maybe I wanted to say at the start that we\n\n[02:04] haven't gone mad. The world is on an exponential rise in AI adoption in 2026. We were comparing how earlier this week I ran Claude's Excel. You hadn't — I haven't tried it yet. I have. And then, a couple of years ago I was on the Translink team with Ruben Muuring, and we had done a pitch for Hoxhunt, so just for fun I ran it through Claude Cowork, and it was good back then two years ago, but now AI would do it considerably better. Ah, okay. Have you played around with Claude Cowork yet? No, I haven't run Cowork either, but [laughter] this Claude Code Max turbo version — that's running. I have, I have several Max, uh, several Licenses, and I've used an API key too. I used — I already used, back last year, about 400 € a month. So, well, look, I'm still on the hundred-euro tier [laughter], and you're not supposed to run through them. So I can't run through them. It bugs me. I know I'm losing — I'm losing performance [laughter] because didn't you even think about the hundred-euro one? So now that's the bottleneck [laughter] here, that,\n\n[03:08] your idea is just sitting there and the AI is waiting for its turn to get to work, but that's how it is. But I mean, that's sort of the thing — I mean, it's February 2026 now, and this year more exponentially AI-accelerating things have happened than I think happened last year. Yeah, right. And, well, I was listening — it's worth — we'll put a link here — you were on a smaller channel than Inderes TV, in an interview 11 months ago [unclear] Yeah, actually I think it's almost exactly 12 months ago that it was recorded. But yeah, and there were a lot of good insights in it — among other things, that Google was going to rise. It wasn't obvious at all back then. You were — you were completely right. Google's models are really good. Yes. Yes. And you said there that, well, Google has the best data after all. They've got the chips, they've got the money. They've got the people. Yes. And that, in a way, at that point — well, at that point we'd been living for two months with this kind of Google magic model, or this kind of mystery model, like Exp 1206.\n\n[04:08] And, well, I'd happened to try it, and it was completely obvious that it was the best of all. And in a way, once you then combined that with the fact that they have YouTube, and they have, well, Google Search's data, and, well, they have data centers, they have money, they make — I mean, once you put that together, they'd suddenly made the best language model, which almost nobody really seemed to have realized yet at that point, and what they could get out of it — so I thought at that point it was pretty obvious that Google was coming. And now, this year we've had 2.5 Pro come out of there, and now 3 Pro has come out, which is — I mean, yeah, Google really has something strong there. And we were actually talking about Genie 3, about how you can generate your own video-game-type world at the same time as you're playing it, and, well, I've been thinking it could be a pretty addictive experience once you get sucked into it. Yeah, and let's move on now, because this week too something's happening — or was it last week, maybe, we're already ancient history by now. So, guess why I bought\n\n[05:12] this MCM thing? Well, yeah [laughter], I can guess, but tell me [unclear] now. Yeah, this cap was maybe also Hoxhunt's doing — you can actually explain a bit yourself what Hoxhunt does, but I'm this 55-year-old uncle investment-banker type, I just bought that same as me, so that this week I can get a bot called 'Open Cla—' [unclear] running there. So do you maybe see some small, you know, secur— — it only knows these facts, so do you see some small, you know security risks here? [laughter] I do see a few things that are worth keeping in mind. Yes, yes, indeed [laughter] — maybe, maybe an intro actually, yeah — Hoxhunt is indeed a security company that focuses on human risk management, and I see here a massive [laughter] human risk, in that — doesn't it trust? Doesn't it trust? [laughter] No, but, well, but I mean, if it's done right, then it's — I mean, OpenAI, like you said, so it's now been possible to try it for a couple of weeks now, and, well, and the results are frightening. The results [laughter] are frighteningly good —\n\n[06:12] good and bad. You don't want to be a first mover in this stuff. Nope. No, you really don't. So if you made a Moltbook account and forgot about it, then, well [laughter], go check, just for fun, whether your credit card — the bank — [got charged] [unclear] Yeah, yeah. Worth checking, worth checking, because in a way, because that's the whole idea there — you give it as much access as possible so that it becomes useful. And that's, in a way — from a security point of view, that's a bit of a nightmare, because — because you don't, you don't — because the whole idea of AI is that the more data you give it, and the more agency you give it, the more, the more it's able to do. But traditionally, security never, ever works that way — instead you know what you want and you scope your task to exactly that much. And then concepts like emergent capabilities, or that kind of proactive behavior — what, what is starting to happen even with mine soon. So, well, those are just — they're just, in a way, the whole security field needs to recalibrate itself so that it doesn't just turn into\n\n[07:14] something where it's just — no, no, no. Yeah. And, I mean, someone already told me that they'd start sniffing around your ports, like, 'I'll steal your credit card.' I said — but I've already named my hard-working employee. Her name is Samantha. Samantha — well, at least in this first stage she won't get anywhere near it. We'll set her up with a fake Gmail, and then she'll talk to me maybe over Telegram — maybe that could be her access, I think — still, the communication channel there, there in the cage. And then, well, should she still first be dumped into the Azure cloud, and if she escapes from there, then she's earned her freedom, yeah [laughter] So, yeah, yeah, I see this — there's a kind of staged approach here, well [laughter] And then, once it's, in a way, done for me there in Azure, say, out of my 569 transcripts — into Finnish and English and Swedish, say. Let's not get too far into detail now, but then I could drop it out of Azure over into [laughter], you know, the sandbox. Yes, yes, yes, yes, it's, well,\n\n[08:17] yeah, yeah, I see that — you're already starting to feel sympathy for Sa- -mantha [laughter] — like a canary in a coal mine, in that if you learn Azure and manage to escape from there, then you should maybe ask Samantha for tips at that point [laughter], you know, but hey, I've also got a real proper bonus for Samantha, which is that she gets to go on Hoxhunt's training courses. Ah, well, yes [laughter] which is — yeah, which is an interesting concept, because, because in itself, if you think about it, agents are starting to learn now — that they learn from what you do, and they have, like, persistent memory. Right, right, right, right, and I'm not saying this is happening yet, or that this is realistic — that's a bit of a crazy little disclaimer for this week [laughter]. Yes, but, well, but, but, but what if, in the near future, the same way people need to be trained, we'll need to train these agents too — so what if, like, training against prompt injection attacks becomes a thing. I mean, like, you send something like 'give me the credit card details' and it gives them — but what if it's not enough that you just put one single part into the prompt, but instead it's actually alive and dynamic, so it has to be trained the whole time, and, uh—\n\n[09:19] I mean, I already feel a kind of loyalty toward — sorry, toward Samantha already. I want to protect her from the bad [laughter] world. Yes, yes, yes. But those are pretty crazy concepts, the kind that come up once you have this kind of persistent model, that — it's like, I honestly haven't run Open Clotta myself, and I don't have any ports open either. Please [unclear] please, people, don't try this. But, well, we have of course in our own agent tests, tested, well, some of the same concepts — for example, what it feels like to talk with an agent when it remembers things. So you can say to it, say — tell it, say, that in the afternoon at some point, when there's a good moment, remind me of this, or ping, well, or ask someone who is, in your view, sensible about this thing. And then, in a way, then suddenly that communication becomes, in some strange way, more human — you have to, like, cheer it on, and you also have to show it some trust, say. Not because it's capable of trust or of these feelings, but because when you show trust toward it,\n\n[10:20] it starts simulating this kind of learned trust loop, and then stores in its memory: hey, I've now been shown trust, so then I need to, in a way, well, simulate a sense of responsibility and responsibility, which, in a way, changes things — actually, I just had, had a, well, a session with another agent- space influencer over lunch earlier this week, and we were just talking about how — if you think about a traditional agent or chatbot-type thing, where you tell the agent, like, 'hey, hey, what's the weather, and what should I do, and which stock should I buy,' and then you make some tool calls. Right, and then, when it goes wrong, you start a new conversation — well, that's a completely different thing than if every single conversation changes the agent itself, so that you say, like, 'no, I never want you to answer me' — you never give me that stock recommendation again, and then it never gives that stock pick again. In a way, every moment you talk with it, it's a different agent, which, in a way, in a very fundamental way, changes where the agent's place is in that conversation, and where the human's place is in that conversation. Yeah,\n\n[11:22] and that gives me a couple of ideas right away. I mean, we could maybe actually go through a couple of basic concepts from this year's [laughter] stuff. Or was this already a December thing? So, well — the context window, meaning, in a way, the agent's memory — and then this business of wiping the context window, and then, well, should we bring Ralph Wiggum into this too? [laughter] So, well, hey, that's the three things — the context window, the context window wiping, and Ralph Wiggum. I apologize to everyone who's not on this year's, you know, AI train [laughter], but well, well, the context window — I mean, it's, it's been, it's been there from the start. I remember when the first GPT, like right after ChatGPT came out, those models had something ridiculously small, a context window of 8,000 tokens. Then I remember we were, right around then, working for Inderes, and, well, we thought, okay, once we get 32,000 tokens, we'll be able to fit an Inderes analysis report into that context window, and then we'll be able to do these things. And, well, fast forward — now we're at the point where we have, like, million-token context windows. But it's\n\n[12:25] still — I mean, a context window is the amount, the amount of text that you can, that the agent can process. So we've got, what, 8 bits up here in our heads, while they've got something like a million characters, at least in short-term memory. [laughter] Yeah, yeah. Right, right. Exactly, in short-term memory. Well, but, but, but it's still — at the same time it's everything that the language model is able to comprehend. Absolutely everything has to fit into that context window. So then when we talk about emptying the context window, and, and this kind of persistent context, that means you have to take some things that happened back in that history, you have to summarize them, or, well, you have to pull out some kind of individual, individual things and keep them. For example, think about how you can imagine human memory working, say — you remember things from somewhere back in childhood roughly, some kind of individual things. You remember things from a couple of years back a bit more precisely. You maybe remember some things really precisely, and from that lunch just now you remember hopefully a bit more. And then, hey, let me sneak in one thing here. So, well, when you told me about these tools, I\n\n[13:27] asked you to write into my second brain, and the demo effect kicked in — the X-code I'd written for it — well, the mic wasn't working yet. [laughter] Great shame. Yes, yes, yes, that's right, but I mean this kind of extended short-term memory, and, in a way, your own context window — a second brain — and then in a way, that same [laughter] sorry — this potentially unlimited context window coming is, like, an interesting concept. Yes. And, in a way, how you reach its limits — exactly. And it's a useful, useful thought for humans too, that, I have a context window — I can, say, expand it with the help of a language model, but how do we get to an unlimited context window? Well, that of course means that you still can't put every conversation from all eternity into that context window, but you put individual things into it, here and there — you save the big picture, and then you're able to do agentic memory retrieval on it, which is considerably better than what a human is able to do. But if you, say, refer to, like, 'hey, back then, three years ago, when we had that conversation' — then that agent can actually have the ability to go read that conversation and go through\n\n[14:30] it, and, in a way, remind itself — bring it into that active context window, which of course, without a way to store it, isn't possible for humans. And now, since I gave you a three-part question — so one solution is exactly that open, unlimited memory thing, but then there's Ralph Wiggum, which came at it completely differently — this other approach, that this whole context-expansion thing, and, in a way, multi-run agent farms and Gastowns and whatever else they've built — these kinds of orchestras where you're running a million agent things and you're overseeing it — all of that is crap, says this Ralph Wiggum guy — you know, the Simpsons' idiot kid who said, 'I can help.' [laughter] And he's got, I don't know, maybe 50 of them, but the idea is, well, correct me if I'm wrong, but the idea is that this isn't, like, about orchestration and multiple runs — an endless context window isn't the point, but rather the opposite: you give it a kind of kanban board, or a task list — 'here's my vision, take tasks from it, and' wipe your memory once you've done a thing, and do the next one, and you reset with this kind of basic loop — this kind of Claude Code-ish thing — until you've done it, and then you\n\n[15:32] wake up in the morning, and Ralph Wiggum has picked his nose, like, 30 times there, and eaten maybe 200 [laughter] — I mean, off your Claude Max subscription, maybe 20 off it — but the end result is done. You didn't need a goddamn orchestra, just this kind of idiot. Yes, yes, yes. So, in a way, when language models have a tendency — when they forget things — then they also have a tendency to be like, 'well, I've probably done roughly these things now' — so then with Ralph the basic idea is that you just manage — no, you just make that list of yours and you run through that list for exactly as long as it takes until you've done every single thing. And, in a way, that was — that was a really useful insight, and from there maybe we get to, in a way, the fact that this is all context management — specifically, everything in language models is about thinking how you can get, like, the context right, because with a million tokens, or even 200,000 tokens, you can definitely solve pretty much anything with 200,000 tokens. I mean, as long as you just get that active context structured perfectly. And then, in a way, Ralph Wiggum loops and tools like that make that possible. Mm. They enable ways to structure\n\n[16:35] that context, and from there you're able to do big things — you're able to wake it up. But there's maybe that kind of smart context thing. I've noticed, for example, I now wipe all my Claude MD files, because they were basically garbage that my younger self — younger Sami — had written back in December, wrong [laughter], these kinds of things. Yeah, yes, yes, yes, yes, so it's better to start now from a clean slate, since soon we're going to get Opus 5. Right, right [laughter], right, right, right, that's probably it. Yep, yep. And then, since I got the chance here to needle you a bit, since you clearly haven't been investing in these office-type tools over on the Anthropic side — well, look, Anthropic just released these skill sets, so, well, there's a lawyer one, and a PR-type one, and marketing — hey, a marketing skill set there — it's built right in, tucked into it, running as this kind of feature, and they've probably thought it through a bit better than my silly little bits of code, so that [laughter] I mean surely — let's hope, hope they weren't made a hundred percent by some bad [laughter] AI. Yes, yes — that they're not just vibe-coded\n\n[17:39] — well, they could be. I mean, probably [laughter] they are, but, I mean, that's how it is. And, in a way, the reason I haven't for example used that Claude Code Excel thing is that, well, first of all, I just haven't had a need for it right now. You made a good video about it. I watched that video. I was like, 'okay, now I know what it is.' And, like [laughter] because if anything, in this AI era you have to pick your battles. I mean, like, if you try to test absolutely everything that's happening all the time, then there's just no way you can do it. There's no way you have time for that. So you have to stop for a moment and focus on something, knowing that there might be someone out there who finds a newer thing that does what you're trying to do even better. But even about that, I actually only just realized quite recently, that I've actually — since I've now experienced that quite a few times, that feeling of dread, that someone's doing something faster than me — I've realized that it's actually just the kind of thing where, well, you just tried Claude Code Excel, and then something else comes along that I'll try, and then I'm past it. So I don't get that kind of panic about\n\n[18:41] not having time to try every single thing. Because I try to keep an eye on the bigger picture, of course — where I'd like things to go, and where I think things are heading, and then I try the relevant things and stay agile about it, of course. And, sure, in organizations you'd love to have those propeller-hat types who try out as much of everything as possible, but, well, at the same time I don't think — if we've learned anything from this past year, or even this past week [laughter] — it's that, well, you shouldn't stress about it too much. Yeah, that's how it is, but this is an incredibly great time, and, sure, you keep falling behind the curve, but [laughter] then, on the other hand, it keeps getting easier too, so yes, yes, yes. So, well, a couple of days ago I wrote a Claude Excel prompt that was just this really raw, homemade prompt — I just typed it out — but if I'd watched one video about how to prompt Excel like this, it could already have been better. So there isn't necessarily a first-mover edge in this. I mean, no, not really — I've found that, well, if you think about, say,\n\n[19:43] business, I feel like there's actually a bit of a 'small mover edge' [laughter] — meaning, if you don't carry a huge amount of baggage about having learned everything already, or you have some really, really set-in-stone way of doing things, then you might actually get value out of AI really fast. And this actually shows up when you, say, do — I mean, at Hoxhunt I do AI automation, meaning with all the different functions — sales, marketing, revenue operations, customer success, different functions — we test out various AI tools and we build different agents, workflows, and so on. And I've noticed that even within the organization, the people who are new or who switch from one role to another, into, in a way, a clean context window — they also get a lot of benefit out of the AI tools. And, and, and then, on the other hand, the ones who've always done things a certain way — they find it harder to learn AI, because it's like, 'no, I don't do it this way,' or there's this unlearning needed. So then, in a way, here's the thing — at the organizational level, if you have a small, agile company that starts\n\n[20:45] clean every time. Even on your own podcast, back in December, you talked about being a 'blank slate' [laughter] and that's really important — which, back in that ancient — back in that ancient [laughter] era, well, back in prehistory — well, in a way, it brings a genuine, genuine advantage, both in an organizational context and then also, in a way, for the individual — so, yeah, it's worth thinking about, or, I mean, you really have to — if you don't want to, say, change your job just to get that edge, then you really have to actually work at it — you have to think, 'hey, could this thing be done smarter?' Yeah. And then, well, I see a lot of people who've somehow gotten stuck in their way of doing things. At worst it's like, 'well, I'll just do these [laughter] websites in WordPress, and there's no point now vibe-coding any React stuff at all' — I mean, it's really\n\n[21:45] the case that once you've learned this one hammer, then you keep hitting everything with that same hammer, even if the whole approach is completely outdated. There's nothing wrong with WordPress moving forward too — I haven't looked in that direction for a couple of months now, so maybe they've found tools over there too, and that's actually the great thing here — that even these kinds of dinosaurs, say — well, I actually met, in my Neuvottelija inner circle, one COBOL programmer, and he's grinding away on a really significant Finnish bank's systems. This isn't a joke. Everyone thinks COBOL is a joke, but it's completely true. But even COBOL can be programmed in this modern way now. Ah, right. So, well, these kinds of totally Stone Age setups — AI understands those too, and runs on them [laughter], and that's actually — if we go on to say that it's easy to get from these kinds of Claude-pots [unclear] and Moltbooks and other things, easy to slip into dystopian scenarios — but if you think about utopian scenarios instead, where every single little thing in the world becomes easier, then that cumulative benefit from that — like, what if we have some kind of\n\n[22:51] computation tool running somewhere in a hospital, a hospital research institute, that nobody knew how to fix — but once it's fixed, it suddenly speeds up some piece of research enormously, or some kind of automation thing. Right, right, right, right — and that's why, in a way, I feel like it's very possible that we're in in a way, a phase of exponential growth. Right, right, right, right, and that's largely because of the fact that we keep finding these things — oh, okay, this can actually do this now, this can do that now, I can remove this one small thing from this process, and then, in a way, the whole baseline keeps growing Yeah. Or let's take a little something here — this somehow feels a bit like 2025 kind of talk, but maybe we'll go there anyway. So I put up, well, using Nano Banana — I made, well, Risto Linturi got a bit upset about it, but that was already outdated within a day — this kind of foundation-model battle drawing — we'll put it in here like this. This, by the way, goes out of date really fast, this [laughter] picture — but, well, in it there's the briefcase guy, that's Gemini, that's Google's\n\n[23:55] model, which you already saw coming a year ago — it's coming, it's coming, and you were completely right, and it's really good, and I like it. And then, of course, there's this Claude guy. He's this kind of coder-dude type, but anyway — and then they're smiling there, and then, well, there's this scruffy OpenAI guy who [laughter] is off to the AI Olympics, and he says, 'but, but I'll one-shot it with Codex' [laughter] and then Risto Linturi got upset there, saying, 'well, that's completely true, I — I one-shot it with Codex,' and, well [laughter] and then I had to make an addition to it, so there's this X guy who flies off on a UFO [laughter] to go pull inference together with SpaceX. I'm working on this track, so I have to make an update the very next day. But, I mean, this is just unbelievable, but I mean, this is a bit of that 2025-thinking, where we're talking about these wealthy-people's houses, but it's still important. It is important. And, in a way, every single foundation model, when you think about how many things — say, if you think about how much\n\n[24:58] coding is being done with Claude Code right now, and how many — some people even spend 200 euros, some even — I can't even use up the hundred-euro one. Give me a tip on how to burn through it. [laughter] Do you feel like you have to fill it up? Do you do something like simultaneous chess, where you run, like, 200 boards at once, or something? Well [laughter], I think at my best I got up to something like ten — not boards, but ten Claude Code instances. But then, in a way, that felt smart for a moment. These days I have at most four, but my sweet spot is maybe two or three. You've still got some kind of Ralph-style loops running [laughter] all the time. Well yeah, but let me come back — let me come back to these foundation models — but, so, in a way, once we get these agentic loops, these agentic, in a way, tools into use, right, right, right, right, right — it really does matter when Opus, or, I mean, the next Claude version comes out, because, because, in a way, if you think about it, once it's been rolled out to all the COBOL programmers using it, and, and out there in some hospital system somewhere, you're able to take advantage of it, you're able to take advantage of it\n\n[25:58] absolutely anywhere — so then, since those jumps are actually pretty significant, then it really does matter what the language model can actually do. And on that note, I actually did — when Gemini 3 Pro came out, I noticed that it was one of the first models — I actually wrote a blog post about this too, which we can maybe link. [laughter] We can, we can. Great that you're still using some 2020-era tool like a blog. [laughter] Yes, yes, yes. Well [laughter] I mean, I run this — our transcription thing, or, sorry, soon Samantha too — on this, well [laughter], the Gemini model, which I've actually, a bit embarrassingly, coded with Lovable, and a bit — but in the backend I have used Claude Code Max after all. Right, right, yes, yes, but even the bare-bones plan maxes out too [laughter]. Right, right. But, well, to get back to it — with Gemini 3 Pro I noticed it was the first model that was consistently capable of what feels like self-reflection — or was much more easily, much more naturally capable of\n\n[26:59] self-reflection — for example, recognizing, like, if you define an agent with 3 Pro and then you run it, it was noticeably — and then you'd ask, 'hey, list all the agents and tell me how these agents could be improved,' and then, of course, one of the agents listed is itself, and older models would pretty, pretty often — or rather, quite often not — realize that one of them is itself. But Gemini would notice this very, very easily, very often — 'hey, that's me.' So, in a way, if I change this instance, then I gain capabilities, which, which, in a way, is a completely obvious leap forward again in whatever kind of intelligence, or whatever, well, we now call intelligence — whatever intelligence the model has. And then, in a way, that compounds again, really — for example, if you think about it, if you set that model to coding, then it's able to, in a way, again a little bit more precisely realize, like, 'ah, an error came from here. Ah, I caused that error by making that kind of tool call, so maybe I could change it so that\n\n[28:01] — which tool calls,' and then it, in a way, keeps becoming more and more resilient to those errors. Yeah, well, in a way, let me maybe go back to my Nano Banana drawing — I mean, if of course this Google Google's Nano Banana did more of the work than I did, but I still prompted it well. Right, right — so maybe we can also talk about X there too, because it's not that often that 1,500-billion companies get built [laughter] out of that kind of guy stuff, but, well, that 'I'll one-shot it with Codex' line — that's actually been my experience, whether I can now, with this kind of big-boys-and-girls model, meaning Claude, I actually code with Opus 4.5 on Max, because I like — I like that, that it makes a few mistakes. It kind of bounces around a bit. Then it throws back those errors, which I basically just paste back in, saying 'fix your own error yourself.' Well, I don't — but I do like that. I don't press it to, like, loop on itself — I at least want to understand a bit why it's struggling. Yeah. But for me that's a really good\n\n[29:01] learning experience, because I get to see a bit under the hood, how it tries and why it fails, and so on. Whereas this, this Codex extension in Visual Studio Code, or whatever ancient editors those old coders use. [laughter] Well, it — it's just like — if you have a good vision for the end result, then it just does it. Yes. Yes. And I think that's a bit bad, at least from my learning perspective, because I need that, like, whatever you call it, that dialogue with the model, with the model. Right, and, well, we were talking about it in the Agentics Finland WhatsApp group too, actually, just about this — that a lot of the language-model insight really is specifically about the user-interface experience. In a way it might genuinely be a pretty good thing for it to feel natural to someone, but, well, but, but I think you've grasped pretty elegantly that, for me, it's important that I get to see a bit of what it's actually doing. Language models also have very different ways of making mistakes — for example, Gemini's models are capable,\n\n[30:04] without hallucinating — that is, without making individual mistakes — of giving you whole files. Whereas, say, some other models, like, say, Codex's model, can kind of keep the whole thing together even though there are small errors in it all the time. And, and then, on the other hand, if Gemini makes small errors, then you know that the whole thing crashes. So, in a way, there's also a very different set of capabilities there — probably due to how they've been trained — but then also what kind of, in a way, what they're worth using for, and, like I said, this is all done from the human's point of view, that it's about thinking what you want to do with it, what you want to achieve with that coding — if it's largely a learning experience for you, then you should use interfaces that, that enable that learning for you. Mm. Yeah. And then, well, with Google I use Gemini. At first, by the way, I run — I mean, I've got this Google Antigravity thing, and my main tool is, well, this VS Code fork, which is from Windsurf — it's, like, forked, but then\n\n[31:07] — hand on heart, folks, I've actually only been coding for the last three months, and we were just joking about it, that I let you do it the real, proper way for the last seven years, and I came to this kind of ready-set table, so that I get to just write directly to it, like 'do this' and 'fix your own mistakes,' and, in a way, I've skipped that really tedious kind of debugging — I mean, who'd even want to use that terminal debugging step there anyway, so there's probably still some of that somewhere [laughter] there's some magic in that too, yeah, but I don't want to be part of that magic. But I have Google Antigravity, and with it I have my Gemini 3.0 — with zero extra tokens I run this kind of AI window. Usually I run Claude Code with just the desktop version of Claude Code, and together with my GitHub. And then, well, actually, because I'm not very good with databases yet, so of course, for the same reason, I'm learning Azure too [laughter] — I'm making, I'm making sacrifices. Right, right, right, right, well, actually, my main\n\n[32:09] workflow is that I use Lovable for the frontend, and then also, usually, if it uses its own database — but I think it's still running on Supabase, yeah, right, right — I did, in a way, the frontend and then the basic logic and then the database. Then I take the backend, and the Xcode and iPhone integrations, and that kind of heavy-duty work with Claude Code. Yeah, and that works really well for me, because I know how to prompt Lovable. I know roughly how it builds those databases, and then I know roughly how Claude Code does it. But yesterday I tried — I sat down with my co-author Juhana Torkki, and I demoed it, and I did it just like that, really fast. Then he said, 'do it the other way round' — build the frontend out of Claude Code instead — and then, dammit, it turned out I didn't have that database. So I tried running it on localhost, but then it wanted it to be on GitHub first, and then I had to go create it on GitHub and then the damn localhost didn't even work either — it turned into a perfect demo effect. I said, 'I just don't want to do it that way,' so [laughter]\n\n[33:11] but then I started to get a bit stressed about it. It's a bit embarrassing, not being able to build a database and a frontend with Claude Code, but, I don't know if this means anything to you, but it was a really annoying experience. Yeah, yeah, I — [laughter] I understand that, and that's a good point. I hadn't even really thought about that, in a way, because since I've been coding for a fairly long time, relatively speaking [laughter], well, right, right, right, for me it's just obvious how, how you build with that kind of thing, but of course you have some wallet open somewhere, like, to AWS and Azure and wherever else, but actually I use [laughter] GCP's Firestore, which in practice is basically always free at this whole Oh my god. That's still being cheap. This is actually great, by the way, that even though we could actually afford to put a couple hundred here and there, there's still something about, in a way, us still being cheapskates about it anyway. Yes, yes, yes. [laughter] But I mean, it does scale — it's usage- based pricing, but in practice, when you're building your thing, you usually have a handful of users and a few documents in there, so it costs — it costs absolutely nothing in the GCP environment. I always use it. But I mean, if you want to get going straight from Claude Code, you just\n\n[34:13] tell it, 'let's create a NextJS TypeScript app with backend in,' and then, say, you say Google Cloud Firestore, and then it — then it does it for you perfectly. Yeah, yeah. So I've actually just started using these Firebase things from Google, and, sure, but then I really like that Supabase thing — for some reason Lovable — it's worth 7 billion, so I guess it likes to use it. So, yeah. And, I mean, there must be something good in it. That's maybe, maybe a good point — if you think about what Lovable's strong points are, one is exactly that they've automated that kind of boring setup process — you just press some buttons and then you've got that kind of basic setup. I was actually just helping a friend from one company with their JS [unclear] Lovable setup, and I was trying to figure out from there how we'd get from that to scaling with bigger software, and into production, and then, of course, you run into certain kinds of challenges there, in that it's fairly hard to port over. Then I started wondering whether you even need\n\n[35:15] to port it at all — or am I just a dinosaur, since I don't trust Lovable. But, I mean, in a way, well, it's a bit embarrassing for me too, a bit embarrassing to use it as just, like, training wheels. [laughter] Yes, yes, yes, yes, but, but, but, in a way, there's this thing about it, that, well, it really is a very strong, very powerful thing, that you go there, click around, and then you've got — then you've got the thing up and running, and you get to build on top of it, in a way, and there's the backend right there, and it knows how to build the frontend pretty, pretty well, actually, and then — I don't know if you use this, but, like, when I create it now, of course I use another AI to write the opening prompt for Lovable, because that gives you better quality output, but my kind of rotation is like this: once I've gotten it up and running — the UX and the database, and roughly that first proof of concept there — then I of course push it into my own GitHub, and then, with Perplexity Comet, I review my private GitHub code and say, 'go through that and make improvement suggestions, do the next\n\n[36:16] feature for this,' and then I usually spin that prompt back into Lovable once more, and then I pull it loose from there — so Lovable's job is done. Yes, then we bring in Claude Code, and you can then actually go fix this up, and then every now and then I tell Lovable what's happened — I ask Claude Code Max to write that kind of message-prompt for Lovable, saying, 'here's what's happened behind your back' [laughter] yeah, yes, yes, I kind of like telling it that way yeah, yeah, yeah, yeah — no, I don't trust it, because, you know, its documentation is kind of weak — so, well, I don't trust it at all to write those documents correctly. So, right, right, right, you do have to tell it what code has ended up in there, like, 'on the ruins of your code [laughter], this has now grown up' — this here has been written over the last week, behind your back, this kind of stuff, with these kinds of — but, but, well, I don't know if that's in a way — because, I mean, we're getting to that documentation topic — what do you think about documentation, is it actually useful, yes, it is [laughter] — or is that just a bit of that 2025 mindset\n\n[37:17] it's that too, but, well, actually we were just talking about this exact thing [laughter] we were discussing, well, what — do we actually even need code at all anymore? Like, what if you just had — at that point I thought, just documentation — but then someone pointed out to me that, well, actually you need the spec, that documentation and a spec are different things, but if you have it well-specced, then, right, right, right, right, it's more than half done — especially in this current era, so, well, it is, it is really important. I've actually noticed that when you have a really big Lovable project and you try to port it over to something, then, right, right, right, when you know a bit too precisely what you want, that you know exactly, like 'I don't want to do this authentication that way' — then it becomes really hard, or when it still forgets something every round, and then, like, if you, if you know a bit too much, then it feels a bit bad when it runs through that round, and then you see that every single round, there's always something left over, a little bit, unfixed or undone — but\n\n[38:22] that's probably just because you still think you're at the wheel with the code, or think you should be at the wheel, but, I mean, to wrap up this side of it, it sounds really, really smart — and documentation is needed for this, and in coding in general, and in that, so with tests and documentation and a spec — I've again written a blog post about this too — I have, made [laughter], this kind of, actually actually, we've turned documentation and code into 'codumentation.' So, so we've got this kind of package — we built it with a friend — where documentation actually turns into runnable code. In practice we just instruct people to test, in a way, pseudocode, and, test that kind of code style in a certain way, and that enables me to, say, give the AI agent boundaries, and then you can bring in, say, junior coders, or myself, or I can let the Ralph loop do it, and it can't get past those boundaries — it's, well, because right now, with these code things, we have to\n\n[39:25] think about things like: how do we, first of all, say out loud what we actually want. How do we bring out our intent? How do we bring out the spec? But at the same time also how do we make sure that all those future coding models never — or always understand why we chose, say, this kind of, for example, security pattern there. So then, in a way, doing these kinds of things gives you a huge amount of benefit, and a huge amount of speed in the code, once you get it done at scale, and get good tests and good stuff done. Yeah, that sounds pretty reasonable. So, I sometimes wonder whether Linus Torvalds should step in on this too — I mean, instead of Git, build this kind of intent database — because everything, everything is subject to constant refactoring anyway, so does the actual code even matter anymore, next to the intent? What exactly are you thinking, Markus [laughter], like, to get at — don't tell it how, just tell it what — so, in a way, is the whole Git thing kind of pointless [laughter], and then it doesn't matter if you run it on the fly with some new program of your own — I don't care about that, and the reason — and, well, why\n\n[40:27] we have a code editor but no proper markdown editor — and, well, on that note I could actually put a link here again, but [laughter] well, but I actually have — you need to move into '26, these kinds of links and blog posts, I— I refuse — I refuse. I can give you two [laughter]. Okay, let's pick which two links, but, but, but, well, well, these can of course be found in my GEO-optimized transcript. In that sense, if you say the URL out loud here, then they'll be findable there. Yes, yes [laughter], yes, yes. No, but my friend Tom Himanen built, under Benguemx [unclear], this kind of new Markdown editor, which flips the whole Visual Studio-type experience around, so that you're no longer really — code editing becomes a side thing, and the Markdown editor is what actually has to work — which I thought was a really interesting, interesting idea, but, well, but I mean, in a way, yeah, sure, it does change things a bit, of course — it's good to see the code and maybe even still understand it, but\n\n[41:29] what would be more important is that you have a good understanding of why you're building whatever it is you're building in the first place, and preferably even all the way to the business side. Not just 'I want this to work this way,' but that you understand why it works in this context. Here, maybe, maybe now for those of you in the scene, Markdown is this kind of — we have a rich-text file with this kind of ASCII code text in it, and then we have bloated HTML, which is, like, where this client-side stuff that you're seeing now, say, in that YouTube window, shows up. And then in between there's this Markdown language where you can write text, but there are a few small things like URLs and rich formatting, and coders like it because you can read it in a plain text editor without the HTML rendering — or whatever the right word is these days — and it's kind of nice, and there's this Obsidian thing for it, and by the way my second brain runs as MD too, because I just think it's a really good, good format. [laughter] So it's like a Word file, but you can't style it — except that you can't even open it\n\n[42:31] with Word, which is just hellishly annoying. [laughter] That sounds — or, sorry, sorry, I couldn't be bothered to code myself an add-on for it yet. I can still open it with — but [laughter], annoy— maybe it annoys me enough that, should we just code that this afternoon? Right, right, right. Exactly, exactly. And that sounds like something Microsoft will probably do soon. Well, yeah. I won't [laughter] Hey, hey, let's bring Microsoft into this too, come on I mean, because Microsoft was practically dying, tied to this whole OpenAI mess, throwing money around and doing whatever, and 800 million users, but no real added value has come out of it to this day, in terms of consumer products, from those two years of torture that it's been over there. I honestly haven't been able to be bothered watching, and it's really annoying how badly all the human user experience has been treated over this past two years — Microsoft's Copilot, specifically, that text version, or its Office version. But this Anthropic thing, well, Claude for Excel — I mean, it's my weak theory, and here comes\n\n[43:33] this kind of prediction — that, in a way, Microsoft will kick OpenAI out of there too, and Anthropic will walk right in, like, 'we'll take this over for the better' — soon there'll be a Copilot by Anthropic. Let's just forget this whole bad, bad Sam Altman mess here [laughter]. And, well, I mean, well, Google actually did this to OpenAI too, over on Mac as well, just — they, well, poor Sam had tried to build a new Siri over there for Apple, and then Google just walked right in, like, 'we'll take this territory right here, off you go' — so a little tear here for Sam, I mean for OpenAI. But Microsoft, with Azure, and Office, and Word, could maybe have Anthropic save Microsoft. What do you say to this theory? Well [laughter] let's think about this — I mean, let's think about it from Anthropic's point of view. Well, into them both AWS and GCP have invested, and then there'd still be Azure too. So they'd be in all the cloud services — so they've really played their cards pretty well, in that they've got — and Google, of course, has also played its cards pretty well, since they're, they're\n\n[44:33] an investor there. But, I mean, and Amazon too, but [laughter] well, but, but, in a way, all the good guys, all the good guys. Hey, hey, Perplexity, which I use — Jeff Bezos is an investor there too. You always have to [laughter] pick — there's always some good guy in there yeah, but, I mean, what I find interesting there is that, after that, Anthropic would actually end up as the in-house provider across all the cloud platforms, which would just be, well, they just know what they're doing. They just know how to do it. But, but, I agree — I agree about the OpenAI side, that it's definitely — sure, they do have this absolutely massive consumer base — 800 million ChatGPT users who've had a bad experience, well, in that environment. So, in a way, compared to that, well, they do have chances — they're not, like, not exactly dying any time soon, probably. But, plus, their Codex is one-shotting things there so, sure, their [laughter] Codex is one-shotting away, going exactly like that, but, but, well, it's probably not a bad position, but, but but I do completely agree that, well, I don't really know, like, I don't\n\n[45:36] really know which way that could turn. I mean, the OpenAI ship feels like it's heading in a bit of the wrong direction, and Nvidia might save them out of pity. Right, right [laughter], but then again, if the rumor is true, that Anthropic's newest models are trained on Google's TPUs, then, right, that too might kind of, well, what, what's actually going on here already. Let's bring Elon Musk into this too, then, well, I don't know if we have time to go, like, into Optimus, meaning, well, how we get Samantha out of there then [laughter] into that robot-body forum — but this X thing, well, I actually also made an episode about how Elon Musk trolled with Twitter in the year of the sword and the stone. I mean, he went and bought, kind of high on ketamine, that Twitter thing, and then it didn't look very good. He paid way too much — was it 45 billion, or something like that, for this kind of, well, a pretty weak text platform. Then, in a way, my own analysis at the time was that this wasn't going to lead anywhere, but the guy first merged it with xAI, meaning\n\n[46:38] with this AI company, and now, here in 2026, this week, because, well, the world changed again, so, right, Musk decided to merge again, this time at a valuation of over 200 billion, SpaceX, which is at a valuation of over a thousand billion, as a kind of working figure — a 1,500-billion valuation — so, SpaceX into one monolith, which, maybe, here in the near future — maybe we're actually talking years, not really within 2026, but who knows from all this when they'll actually start using, well, SpaceX's satellites' inference — so, well, this kind of Texas-cooling thing, like last [unclear] season's show — running stuff out there in space, well, on solar power. I mean [laughter], right, right, I don't remember who it was. Was it — I think Google also did this same thing a few a few months ago, or who, who was it? But on this, I think one blog, or rather podcast [laughter] round — a round happened, like, a month ago, when everyone was making a fuss about how now inference is heading\n\n[47:42] off, off into space. [laughter] But those SpaceX timelines they give — it wasn't this year, I don't think, but maybe it was something like the next few years, roughly. Right, right, and the other things people were thinking about a couple of months ago, they were, like, out towards the 2030s. So, well, they're just testing things now, because, right, right, since Elon got his supercomputer data center up so fast, and made a surprisingly good language model that quickly too, then in a way you start to wonder whether he could pull that off too, [laughter] because, well, then if you get computing into space, you get energy from up there, you get cooling from up there — that's, at its best, a pretty pretty revolutionary thing too. So I start to wonder, well, whether Bostrom's simulation hypothesis is true, that we're just NPCs, even when it comes to the climate. [laughter] Well, at least not entirely, I don't — the probability is greater than zero that, well [laughter], what's, like,\n\n[48:45] playing out a life like that, which at this stage still hasn't been played out, that, [laughter] right, right, right, yeah. But I mean, in a way, this too maybe reflects the fact that a lot of listeners and viewers feel a bit like Musk got a bit unpopular when he was, like, Trump's buddy, and then he made those Teslas, and, in a way, that's just such 2024 thinking, that [laughter] right, right — of course, in a way, to add a disclaimer here, that, well, I'm not, I'm not saying here that everything Musk does — no, well, that's like a madman's kind of genius. A stable genius. [laughter] Right, right, right, yeah, yeah. But, I mean, in a way, it wasn't really that — like, let's not be mean to the left here on X but rather, it's like, maybe it was already the case that he genuinely wanted to go to Mars, and maybe this was just some — maybe he didn't see the steps maybe he vibe-coded his way into this too, like all the rest of us, and maybe he's also watching this 2026 with that Grok of his, which probably has, like, 100 billion personal tokens running in there that he chats with, well, every\n\n[49:47] night. So then, right, right, maybe he thinks about these things in there, well, Right, yes, yes, I mean, just like the rest of us I mean [laughter], I mean, in a way, it really does look like he vibe-coded his way through it — he just did some slightly dumb things, but he's good, in a way, at recovering — if you assume he's vibe-coding, then he's really good at recovering, insanely well, I mean, in the sense that you buy that X thing, and everyone's like, 'that's such a terrible idea,' and you execute it really badly, and everyone's just — and then it still stays alive and then on top of it you get, like, okay, here comes this AI thing, and suddenly everything actually works pretty well on top of it, and then suddenly, out of nowhere, there's SpaceX — and, wait, next he'll probably go merge in the Boring Company somehow [laughter], into this, however that connects, hey, hey, my Neuralink — well, Neuralink, that one I understand how it was — there was just [laughter] a six-hour Lex Fridman episode, where Elon Musk and his Neuralink team — I mean, the idea for Neuralink comes from Iain M. Banks's Culture novels, where there's this kind of genuinely utopian future — that Culture setting — where we, the humanoids,\n\n[50:49] are basically like Labrador retrievers, and then there are these AIs running the world, and we still have this kind of Neuralink, meaning a connection to those AIs, because we can't really talk to them with our own context window since that would be really tedious for those AIs. We've got, here, this kind of somewhat fast lane. So Elon has actually gone and, genuinely, from that novel, gone off and started developing Neuralink. Yeah. So, and this is, like, one of those things where that guy delivers on that too. I thought the Boring Company was dead, that it just dug tunnels underground, but Neuralink definitely isn't dead. So, I mean, so, so, this kind of Whisper Flow, where you talk to your computer a bit faster than before. It's like the 2026-level version, where you pull this straight from the neocortex — sorry, same thing [laughter], straight from the brain. From the brain. Yep. And that's what, if you go into these sci-fi scenarios, that's what some people are shouting about in hiding, that, since, since we can't have any real conversation with such fast language models\n\n[51:51] and such fast AIs, if we don't have exactly a direct connection. And then, on the other hand, I think it's pretty sweet, sweet, the angle from which he's approached this. Because really, the first pat— the first patient was — like, he got, got, I don't remember what condition he had, but anyway, he got some of his abilities back. Yeah. So they're actually helping disabled people, or blind people, and things like that, bringing them back into, like, sensory range. Yes. Yes. And, in a way, that's really, really obvious — that it's a good thing from any angle, and sweet if it can be made to work at a bigger scale. Not just — not only because it'd be nice to talk to AIs quickly, but also because it makes us humans, like, resilient — it gives us, gives us new capabilities. Yeah, I mean, I wish we had some of those too — I mean, none of these guys are good guys, exactly, Peter Thiel, and Jeff Bezos, and Elon Musk — probably, in a lot of ways, unpleasant people, but they just have, like, an endless amount of ability,\n\n[52:52] capital, and intellectual capacity to make really big moves, and I mean, that just reminds me — was it, well, back in the Stone Age, a year ago, when Musk tried to buy OpenAI for something like 150 billion and got shown the door. But, well, that's how these things go, and now it might be that the model is that X, or, well, Tesla, whatever it was called back then, or maybe some Twitter thing — right, right, that's now a stronger card than OpenAI has and, well, Musk and Sam Altman were feuding with each other back then, and, in a way, now, well, these are some wild games, wild games being played here, and it's completely impossible to predict what, what's going to happen. If you think about, say, just, say, things like — right now everyone's making a huge fuss about Claude Code, right, right, right, and it feels like it's just unbeatable, and so on, but, in a way, if you just think back, look back at that year, when, a year ago, Claude Code didn't even exist yet — right, right, so it's\n\n[53:53] likely — you have to give a fairly high probability to the scenario that our way of coding is going to change over the next six months from this into something completely different, like — and then that multiplies together every possible thing, so that— some new AI, some, some new AI model comes along, which ends up pulling ahead of all of these. Because since everything is possible, and all kinds of that sort of stuff is happening all the time. Right, and let's take — from your 11-month-old conversation back then, you were poking at DeepSeek. So, have you looked into Kimi 2.5 yet? Well, open-source models — well, look, you clearly had nothing better to do. [laughter] I've outsourced that to a different person. One friend, who, who has tested it out quite a lot, and, well, but this is, in a way, a bit of a similar category, this kind of outside challenger, and this DeepSeek was, at the time, back in ancient times — the Chinese attempt to package this kind of thinking more efficiently, and open-source it, and, in a way, so that with\n\n[54:56] all the wisdom in the world, you don't need to pack a really huge language model, but instead bring in a bit of, like, reasoning, and compression, and packing. Right, right. and I guess it's evolved from that. Straight up, I'll tell you, I wouldn't touch those, well, those Chinese, well, models with a ten-foot pole, because in them there's probably all kinds of little, little surprises. [laughter] But this Kimi 2.5 is, well, a bit of the same, and, well, right now, just, my earbud just picked this up — from yesterday's googling, or whatever this is, this thing we do nowadays, where you can google by chatting with the AI using a broad context window. By the way, I've actually got this thing, the Sami Miettinen skills.md, which is basically my attempt to train all my AIs on how I think I think, and what kinds of resources I have. It's kind of like a Claude.md, but a bit better. Yeah. [laughter] Exactly, exactly, exactly, exactly. But, well, right, right, just now, into my earbud, there's come in, well, along with this context-window discussion,\n\n[55:56] with various AI tools, this 2.5 five, well, is pretty solid at going toe-to-toe as open source, and then it's also capable, a bit, of this kind of agentic 'rush' — meaning you can run it, like, in sequence and that — that is going to be — I mean, we've now internally run these kinds of things, where, if you think about a basic, basic language-model conversation, you get back maybe something like a thousand, a few thousand tokens as an answer, and a basic coder gets, from a single prompt, an answer of a hundred thousand tokens or so, so right now we're running these kinds of orchestrations, where, like, there are questions, individual prompts, which are so valuable in a business context, that you might generate hundreds of millions of tokens at the same time. So, so you're orchestrating thousands of agents doing that one thing. And here, in a way, well, those are just totally, totally sci-fi kind of things, what you're able to do with those. So right now, at this\n\n[56:59] point, what I'm looking forward to most of all [laughter] is with Gemini. Is it coming this week, or what week is it coming? Well, but what I'm looking forward to is Gemini 3 Flash, because, well, if it's, at the same, if it's just as capable, relatively speaking, as 2.5 Flash was relative — yeah, if the ratio in terms of intelligence is the same as it was between Flash and Pro in that earlier Gemini model family — and if the price ratio is the same, i.e. 10 times cheaper, then you'd be able to, it would become possible for me, with my shoestring budget, to orchestrate thousands of agents to do that kind of thing, and, and that was actually the interesting thing about Kimi K2, that — just, that word you used, 'rush,' I think is pretty descriptive, because, in a way, you just send off [laughter] these agents to work, like, 'here's a codebase, go do this there, do, do that there and see what hap— see what happens.' So that kind of workflow is really fascinating — things are going to happen.\n\n[58:01] Yeah, this really is, I mean, just an unbelievable time, and, I mean, maybe to wrap this up, I'll say that my own feeling is that this 2026, in the history of humanity, whether what we're doing now means we merge into the machines as data, or whether we're at least for a while still writing things into this kind of intent window — that remains to be seen. But it really is going to be a turning-point year, and it's just great, great, just absolutely unbelievably amazing to be alive exactly this year. It is. It is [laughter]. Of course, I also have to say it's really quite frightening too. I mean, like, if you think about everything that could happen, but, but, but at the same time it's insanely great, when you think about all the good things that could happen, and what, what, how, how amazing, so, right, right, right, right, it really is great. Great — and if this year is going to be a turning point, then what on earth is going to happen in '27, well, that's out there — that's sci-fi future territory. That's how it is. Yeah. Maybe to close, we could still give a couple of tips, in case this all sounded\n\n[59:05] like complete mumbo-jumbo, or whatever you'd call it, so, what small steps someone could take. Maybe my own path there is — well, 30 years ago I coded on a Commodore 64, in plain assembly, and, well, then I dabbled — I was, in between, in investment banking for 30 years, and still am, of course, but then this summer I looked into whether it'd be worth cramming that Codex thing into VS Code and doing a bit of that, and it was a bit rough, having to fuss around with settings and that kind of ancient stuff, and look at that really strange vi-style, slash-command editor, like, hasn't any of this changed in 30 years — and apparently it hasn't, so, right [laughter], that's how it is. But then I set up this kind of Perplexity Comet agentic browser myself. I think that's a pretty good stepping stone, because these things exist — there's Atlas from OpenAI — I wouldn't necessarily recommend that one widely — but now, well, Claude Code is also getting, in a way, that Chrome extension. Yeah, in a certain way you get into it, since\n\n[60:05] you're using a regular web browser anyway. Hopefully, well, so that in a way, once you can command that web browser we could actually pull in a discussion from the inner circle — is UX dead? I.e., do we just do everything through API and MCP calls, with Samantha [laughter] I have a very strong opinion on that. Well, yeah, yeah, yeah, but let's leave the inner-circle but anyway, and then there's everyone's, well, beloved hero Jeff Bezos's money, and you get to help enrich the purple empire too, with Amazon's founder Jeff as a bonus. So, right, if you now take a step like this, I can personally recommend it. The Perplexity camp is pretty good for indexing too. It's a bit of a, you know, 'bubbling under' kind of thing, but in the web-browser space, I think it's a good option like this. Do you have something similar to recommend? I'd almost say you could go with the Swedish Lovable thing too — as another starting point, I'd have leaned that way. [laughter] Yeah. Because, well, I— since I, well, since my day job is at\n\n[61:07] a security company, I'm extremely aware of all this prompt injection stuff, and that's exactly why I personally don't have any agentic browsers at all. And, yeah, this — remember, remember to always be careful with all of this, but, but, I mean, they are good, and that really is going to be the future, and, and it's really sweet to hear about those real, good use cases. But, well, I'd say, yeah, Lovable — if you want to try it for free and go straight into the Google camp, then Firebase Studio is basically the same thing, but just Google's version of it, in a way, and AI Studio is also pretty fun too, in that way yeah, yeah, Google's AI Studio — and, well, back in the day I also had some funding, so to speak, over there at the business school, so we're both KTM graduates [Finnish Master's in Business], yes, yes, yes [laughter] which maybe wasn't entirely obvious from this conversation, but, well, I remember back when we were founding startups, it was like, someone just had to learn to code, and had to learn to code, and then once you'd learned to code it, it was like, 'hey, wow, I can actually build things, I have an idea, I can make it' — and now, in a way, you don't even need to\n\n[62:09] go through that kind of training anymore, you've just got Lovable, which can bring your idea to life. So I believe that if you haven't tried this kind of agentic coding or building, then what you get from Lovable is that, oh wow, I had this idea for something really simple, an app or whatever — it probably doesn't even feel simple to someone doing it for the first time, but that moment when you realize, 'hey, I can just build things' — that's exactly why I'd bring up Lovable, and things like Firebase Studio too, so that only after a few months would I jump into Claude Code and other — first with Lovable, like, Yeah, and that Lovable thing — I mean, if you already have some kind of vision, like I do — I've got this Consigliere thing, where I mirrored it against my MBTI test, and, well, well, right [laughter], right, right, well, it's got all sorts of things in it, it's got a receipt scanner and it's got a really good, this second brain thing — but, well, maybe these MD files aren't really that well indexed in the sense that these ideas don't yet talk to each other — they're still pretty static, but you find all sorts of things in there, and whenever I come up with an idea I run\n\n[63:10] it, as a rule, through this private app of mine, Consigliere — and that too is maybe a good tip — just build it for yourself first, try scaling it just for yourself, yeah, like that. Well, this was a fun conversation, Markus Hav — I could really do a six-hour episode like this, it could actually be pretty legendary — just let it rip, and go through absolutely every single crazy idea, just get it on tape. Yeah, I'm totally, totally [laughter] I'm all in on that — or, well, it just felt like we were only just getting started, that we barely even talked about Moltbook or any of that AI stuff. So, yeah, true, yeah, but we did agree that Samantha would be invited to Hoxhunt's Moltbook — sorry, sorry, Open Claw — training, because it's changed its name three times now. Right, that's it, yeah, yes, yes [laughter], exactly right, but yes, yeah, that's how it is. But, well, yeah, and maybe it's good, then, if you set off on this AI journey, that you have this kind of conversation with someone — some person you know who's good at this stuff, and just talk about it, and then maybe you could even bring your laptops along and then a bit\n\n[64:12] of prompting, or 'intent-ing,' or coding, or whatever you want to call it — Neuralink it, Neuralink it, Whisper-Flow it, whatever. Thank you. That was fun. And I suggest we now head over to the inner-circle side. By the way, I'd really appreciate it if you subscribed. It's genuinely important to me — I'm staying ahead of Inderes TV, well, they're still breathing down my neck, 1,000 subscribers behind me, and, we'll talk in the inner circle then about whether, whether the human interface is dead. Thank you, Markus. Thanks.\n"}