---
title: "How the Laki.ai MCP Brings Finnish Law Into Claude | Aleksi Paavola"
titleOriginal: "Finlex suoraan Claudeen – näin Laki.ai MCP toimii | Aleksi Paavola | Neuvottelija 396"
episodeNumber: "396"
guest: "Aleksi Paavola"
datePublished: 2026-07-18
duration: "01:15:03"
youtube: "https://www.youtube.com/watch?v=ain1HvF4o7Y"
captionsVtt: https://www.neuvottelija.com/podcast/episodes/737-finlex-suoraan-claudeen-nain-lakiai-mcp-toimii-aleksi-paavol/captions.en.vtt
captionsSrt: https://www.neuvottelija.com/podcast/episodes/737-finlex-suoraan-claudeen-nain-lakiai-mcp-toimii-aleksi-paavol/captions.en.srt
originalLanguage: "fi"
topics: ["ai_enterprise_tech","saas_software"]
subtitleMethod: "claude-fi-to-en-cue-preserving-translation"
provenance: "local SBV curated by Sami Miettinen (Neuvottelija-Subtitles folder)"
fiCanonical: "https://www.neuvottelija.fi/fi/episodes/817-tekoalyagentit-muuttavat-lakialan-mita-ihmiselle-jaa-aleksi"
canonical: https://www.neuvottelija.com/podcast/episodes/737-finlex-suoraan-claudeen-nain-lakiai-mcp-toimii-aleksi-paavol/
---
# How the Laki.ai MCP Brings Finnish Law Into Claude | Aleksi Paavola

## Chapters

- [00:00](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=0s) Aleksi Paavola, language-model and AI expert
- [01:33](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=93s) Gemini 3 and Google's investments
- [02:27](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=147s) Why Laki.ai MCP is a big improvement
- [03:30](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=210s) The challenge of legal documentation, and an NDA comparison
- [05:06](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=306s) Finlex's data opening up, and Laki.ai
- [06:36](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=396s) Linking documents to one another
- [07:47](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=467s) Legit's Claude legal skills, and Aku Nikkola
- [09:01](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=541s) What an MCP is: the USB-C analogy
- [10:18](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=618s) MCP versus a traditional API
- [12:41](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=761s) What an AI agent actually is
- [13:54](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=834s) The difference between a language model and an agent
- [15:57](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=957s) Always use the best model
- [18:17](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1097s) The Bitter Lesson and the shrinking human role
- [20:35](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1235s) The shrinking harness in software development
- [22:08](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1328s) Legora and Harvey aren't very good
- [26:27](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1587s) Claude is unmatched in Finnish law
- [28:00](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1680s) Chamber of Commerce arbitration and closed data
- [30:22](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1822s) Skills explained, and Finland's legal skills
- [32:48](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1968s) Two models debate each other
- [35:12](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2112s) Lovable's smart defaults
- [40:13](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2413s) Fable returns, and the value of model intelligence
- [45:14](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2714s) The personality differences between Claude and GPT
- [46:35](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2795s) OpenAI's data: Codex displaces chat
- [50:27](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3027s) Ditch Copilot, pick an ecosystem
- [53:03](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3183s) Desktop apps and the future of the browser
- [55:22](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3322s) Who is replaceable by AI
- [58:11](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3491s) Samantha, the WhatsApp agent, and memory
- [1:03:29](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3809s) An agent army on a basic license
- [1:08:00](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4080s) Self-sufficiency and local inference
- [1:10:11](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4211s) A Laki.ai recommendation and a light security caveat
- [1:13:14](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4394s) EU servers, the AI Act, and the wrap-up

## Summary

Aleksi Paavola is an AI builder and technology entrepreneur of more than ten years who has trained language models for teams ranging from Google's Europe organisation to a group in Saudi Arabia. In this episode he opens up Laki.ai MCP — built in an open-source spirit — which connects Finlex's legal sources directly into an AI agent and makes reliable Finnish legal research free. The conversation runs from the basics of MCP all the way to the question of whether humans still add value to increasingly capable AI. Sami and Aleksi compare how Claude, Codex and Copilot work, dissect the expensive closed products Legora and Harvey, and consider why, in Finnish law specifically, a general-purpose model beats the specialised competitors. They close on Samantha, agent armies, local inference, and whether Europe still has any say in the age of AI.

## Transcript

*English transcript derived from validated English subtitles ([WebVTT](https://www.neuvottelija.com/podcast/episodes/737-finlex-suoraan-claudeen-nain-lakiai-mcp-toimii-aleksi-paavol/captions.en.vtt) · [SRT](https://www.neuvottelija.com/podcast/episodes/737-finlex-suoraan-claudeen-nain-lakiai-mcp-toimii-aleksi-paavol/captions.en.srt)). Timestamps link to the original video.*

[**00:00**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=0s) All right, welcome to the Neuvottelija channel, Aleksi Paavola. Thanks so much. And we didn't really know each other a week ago, we'd seen each other's conversations over at Agentics Finland a few times, but, you'd done such a great piece of coding that I thought, get this guy into the studio right away — and luckily it worked out. So could you start by telling us a bit about your background — what you do and where you're coming from? Yeah, absolutely, great to be here. My name is indeed Aleksi Paavola, and I've been in tech for a good 10 years — actually probably closer to 15 by now — as a tech entrepreneur basically my whole career, and I've founded and sold a couple of tech companies, and for the last four years or so I've been building AI stuff, and also doing software development and software consulting. Now mainly around

[**01:02**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=62s) these language models, on top of and around them, all sorts of things — I've done quite a lot, been in Saudi Arabia running trainings, and I've also trained Google's Europe FINA team on language models, [chuckle] and then I've also been in Oslo representing Google's language model cloud solutions at a seminar with over 10,000 people, and there's been all sorts of interesting stuff. Cool — actually I had Topi Manuas as a guest. He moved to Singapore and he's been doing Google's cloud sales, and we talked about Google's multimodal solution. Now I'm well, even though we probably both use Claude more, and Codex — but somehow the Gemini folks invited me along there, when they launched Gemini 3.0, we were at this really exclusive event — it was champagne — and they released the model an hour before the global

[**02:03**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=123s) release, and I've always had this soft spot for bribery, so Google feels like a really nice company, and they have in Finland, in Hamina, real investment going on — yeah, that's actually where in Hamina I was running this training — Google's Europe FINA team gathered in Hamina for a seminar like this, and we went through a bit of the basics of language models. Hey, that's awesome. You're honestly a bigger deal than I thought — I thought you were sort of roughly on my level, but now it's getting a little scary. But anyway, the reason I invited you here wasn't about Google at all — it's that you've built this open-source project, this better Model Context Protocol tool called Laki.ai, and I started using it right away, and it works great. Before that I had something else filling that gap, something called Oik, but that was clearly worse. So you've built this better tool, but explain — what exactly is Laki.ai, and what is MCP? Right. Thanks a lot for the kind words. It's

[**03:06**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=186s) come up from other places too, actually — that listeners should really go and try this out for sure — but, um, this project actually got started with Ilkka Linkoneva and Juuso Vesanto — we tried to solve this challenge, where we'd identified that language models were going to produce a huge amount of this legal documentation. And then people would still need to read and understand it. So we developed a solution for that. This was already a few years ago now. And our hypothesis was that if we fed into our system, say, 100 NDAs, a lawyer could then highlight some section in Word from a new NDA that they've gotten from a client, and then see, from the ones they'd previously approved themselves, um, the corresponding section in those NDA documents that's semantically closest

[**04:09**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=249s) to it, and from that easily approve it, like okay, I've approved this kind of clause on basically the same terms before, so this is definitely fine. Well, the surprising challenge that came up here was lawyers' ability to express things in a million different ways. [chuckle] So these ended up just full of redlines. So now when you — like in that test we had, we had about 50 of these different NDA agreements, and when we took a new NDA and you highlighted some section in Word, it was always just red with redlines. The phrasing and the terms somehow these lawyers had worded so differently that the tool didn't end up being very useful. Well, around the same time this pretty fortunate thing happened,

[**05:10**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=310s) from the perspective of Finnish citizens, actually — Finlex, its data became openly available for everyone to use. So [chuckle] this had been — and this was, this was March 2015, I think it was, and before that, I remember there was some big software house that was doing an AI project with Finlex's data, and it fell through over this licensing issue. I think Alma Media was holding onto it too, too tightly, but then in March it got freed up, and our idea was that we'd build an interface — like a better interface — for Finlex's data, um, with no commercial purpose at all. Just mainly because we could, so we'd do it, and we figured this was going to be a hugely useful service for citizens, of course, and so we built Laki.ai, launched it back in 2015 sometime in summer, and it's still — it's just a

[**06:14**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=374s) traditional web service for Finlex's data. So it has laws, statutes, there are court cases, government bills, and what sets it apart from Finlex itself is that these documents are all linked to each other. So when you open a specific section of law there, for example, you immediately see, okay, this is linked to these government bills, and then this is linked to, say, these seven court cases. Mm. So that's basically what we built, without any bigger plan for anything beyond that. Well, now, fairly recently, um, we had this idea — well, could we build something for an agent on top of this, and [clears throat] we looked into it a bit and concluded that yeah, this would definitely be possible, and that's what got

[**07:17**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=437s) this Laki.ai MCP started. And what that actually means in practice is that now, with your AI tool of choice, you can connect it to this Laki.ai database, and now your agent can use all this legal data that we have there in the Laki.ai database. Yeah, yeah, it got straight into my Claude, uh, tech stack — though there was a bit of fiddling involved, actually, let's just say, right at the start we'd invited — along with Aku Nikkola — from Legit here — he, well, he wanted to spend, understandably, a long stretch of summer vacation afterward, so he couldn't make it here, but around the same time he'd released this — Claude has this legal skills thing, which is kind of Anglo-Saxon know-how, and he'd then

[**08:17**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=497s) built, with his team, a kind of skills library that has the special features of Finnish law — starting from the language, like how a section works, and so on [chuckle], and he published it. I took that into use too. There was a bit of fiddling involved, actually, because for some reason Claude had changed the connectors a bit, so I had to go down to the command-line level to shove it in, which was a bit annoying. But when you combine these two skills — this skills library, which in my opinion is actually pretty good, I'd recommend that too — and then this Laki.ai, you get a really nice combo for Finnish legal needs. I don't know, have you had a chance to test Legit's skills yet? Yeah, definitely have. And maybe I could add a bit more, since you asked, to open up a little more for listeners — MCP isn't necessarily super familiar to everyone. So it's an open-source protocol that Anthropic — the company behind Claude — built and designed, and it stands for Model Context

[**09:21**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=561s) Protocol — that's what MCP is short for — and in practice it's a way to connect different applications for agents to use. So maybe the closest example would be something like USB-C. So if you think about your computer, and you want to plug a keyboard into it, you plug the USB-C cable in between them, and then, simsalabim, you can use that keyboard with the computer. So this is basically kind of the same idea — when you connect your Claude, or soon ChatGPT too, which we have coming for Laki.ai — then your Claude doesn't need to hunt for legal sources on the open internet anymore, which is pretty difficult and easily leads to errors — instead it can now use this database

[**10:23**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=623s) we've built. Yeah, that USB comparison is good. So as I see it, MCP is kind of a more refined thing than a regular API call — Application Programming Interface — and the problem with those was that everyone called them a bit their own way, like pulling data from HubSpot into Word or whatever — so it kind of created this sort of port standard, and it eats up some bandwidth, and it's a bit too complicated for really fast calls, but it opens up all the interfaces really nicely, so — did I get that roughly right? Yeah, pretty much. It's just that, um, behind MCP there's basically always APIs. So it's a protocol built on top of those APIs, designed specifically so agents can use them efficiently. And yeah, what you said is completely true, that there's been [chuckle] — you said there are some challenges too, and that's completely true. And

[**11:23**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=683s) MCPs have gotten a lot of criticism exactly for filling up the agents' or language models' context window, but these have improved quite a bit, and what's actually improved the most, is agents' ability to actually make use of these different MCP services, and it seems like — I myself was fairly skeptical at one point about the future prospects of the MCP protocol. But now I'm a bit more positive again, it seems like maybe we're finding solutions where these agents can now efficiently make use of these MCP connections and services that are hooked up via MCP. Yeah. And in a way, since the level of inference, or AI capability, keeps rising all the time, it's not really a problem to build even really weird pipelines. Like, AI will build basically any kind of, uh [laughter]

[**12:27**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=747s) ancient pipe between some old database and a Cobol system, as long as you just shove it in there, because they're just capable of that — but those are fragile, and MCP is a standard, and standards are good for a reason, because they evolve too, in that way — and it's nice to hear that it's also become more efficient, that the standard itself has developed along the way. Yeah — you say 'agent' a lot, and actually I myself had a little chat about this over lunch earlier, about whether agents — in the sense that I want to understand, like, this same kind of AI employee — are they really that useful? Maybe we should save that conversation for later. I think we have a bit of a different view on that. Um, because personally, what you might call an agent — like some kind of thing sitting next to an MCP, some subroutine call — I wouldn't even publicly call that an agent. It's just some subroutine running there. Not an agent at all. Whereas these agents are more like same-level AI employees, almost god- like beings that can do high-cognition work,

[**13:31**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=811s) not just sit in between two things. I don't know if you want to jump in on that already, [laughter] I can jump in a little, um, to the extent that, if we think about the direction ChatGPT and Claude have moved in, they've become a lot more agentic. If you think about how, say, the very first ChatGPT, back in November '22, that released version worked, it was always just — you'd write some prompt, the prompt got sent to the language model, and the language model sent back some answer. But now — and the way I think about it — the difference between a language model and an AI agent is that in an agent the language model is used in some kind of loop. So if we think about, say, essay writing, a non-agentic way of writing an essay would be that you have a prompt, and out comes the finished piece the moment you feed the prompt to the language model, whereas in the agentic

[**14:35**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=875s) process it evaluates its own work along the way. So it writes a first draft, say, and then it checks, okay, maybe this second paragraph needs to be strengthened a bit more, or the ending is missing some punch — so now ChatGPT and Claude have moved in this direction, so I think in a way you could even call these chats agentic systems already — but of course it's true they're not agents to nearly the same extent as, say, Claude's Sonnet, or Codex, or Claude's Cowork, or Claude Code. Yeah, but that's honestly, scarily, always moving toward that same point — for anyone who hasn't yet heard the term — this kind of open AI employee that has its own three-part memory structure and very advanced context management tools — so we actually have to get into this context concept, because you said the MCP call eats up that precious context, and

[**15:36**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=936s) we're already at million-token contexts, basically as standard models — Claude's and, um, OpenAI's base models run on a million tokens, so it's less and less, in a way, that short-term memory eating problem here that's disappearing, and if the MCP model gets even more efficient, then this is a vanishing problem overall. How do you see, by the way, this — um, I've actually had guests who like to run the 'dumb' models on purpose — like, it's not a bug, it's a feature — that they run this 'dumb' inference on purpose because it's easier to understand, and it's actually good in a way, that there's not too much creativity or harness layered on top of it. Do you have a philosophy on that? Well, yes, yes [laughter], very much so, maybe, maybe as a basic piece of advice I'd say that people should, on a general level, pay for these AI services, so that nobody should be using the free versions of these for work — because paying gets you a lot of value for your money, and these are still

[**16:40**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1000s) relatively affordable, at least for now — though I do expect that prices will definitely go up, and have partly gone up already — because even Anthropic's and OpenAI's Enterprise plans aren't exactly cheap anymore. But generally speaking, I think a good basic recommendation is to always use the best model. And that's kind of something people should just know — it's like modern general knowledge, that you know — if you think about it, if you've chosen to work in, say, the Claude or Anthropic ecosystem, then you should understand and know what the difference is between, say, Sonnet, Opus, and Fable is. M. And then, as a baseline, you could always — if you have some difficult task, [clears throat] you use the best model. And then again, if you have — if you're just asking, hey, what are some good lunch spots nearby, then it really doesn't matter whether it's some super-smart model. You could ask that to basically anything. Yeah, yeah. That's how it is — then we can

[**17:44**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1064s) move on to local models, where you're not allowed to do any tool calls or internet searches — so it's kind of very rigid, and in that case it can be really important to know in what way it's mechanically rigid, because then it can be fully reproducible, because in a way, maybe this hallucination that the better models always get criticized for — the more harness or control layer you have in between, or agentic loops, or Ralph-loop envelopes, or whatever cleverness you build in between the model and yourself, the more hallucination loop comes into it too. So in a way, one solution is to strip all of that away, to just run it directly on bare metal in some task that's outside your own interest, so you know exactly what the model is going to do. Yeah, right. There's actually a great essay on this — I don't know if have you read it, or come across it — the Bitter Lesson? No, I haven't, so tell me more about this legendary Canadian

[**18:44**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1124s) professor who wrote, years ago now, this essay on AI, where he argues that the value humans bring to AI systems is, sort of, really limited over time, and he gives good examples. So if we think about chess, say — originally in chess, people thought that humans could bring value to the AI. So basically, with this harness you could make an even better AI. But then, from chess or chess history, we know very well that a human can't add any value to a chess AI like that. A chess AI on its own is much stronger than AI plus a human. And

[**19:46**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1186s) the AI can't be helped at all by a human giving it tips for that kind of chess AI. And self- driving is a good example too — there, first they tried doing sub-optimization, building [clears throat] small pieces at a time. Andrej Karpathy has actually talked a lot about this, about how Tesla's strategy — which a lot of people have already copied by now — of training the car to drive so that the neural network handles the driving has turned out to be a much better strategy than trying to tell it in specific spots, sort of steering the AI using human intelligence. And now I believe we'll see this same thing, and we already have, in a lot of language-model-based systems now. If we think about software development, say — in software development, at first there was Cursor, which was this traditional

[**20:51**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1251s) integrated development environment — an IDE that developers were used to using — and when AI was brought in and integrated into that IDE, for a long time it looked like, okay, this is going to be the way — that humans can add value at various points along the way in software development. I myself was actually really skeptical about that when Claude Code came out, whether it could really be true that a purely text-based interface could work on its own. Whereas over in Cursor there was this tab completion — you'd write the start of a function, and then the AI would predict what kind of function you wanted to write. Pretty quickly, though, it's gotten to the point where I don't even remember the last time I did a tab completion like that. So it's really shifted toward these... these harnesses have kind of shrunk from that, and I believe this

[**21:55**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1315s) trend is going to keep going, and now if we look at say what OpenAI's or Anthropic's employees say — what they talk about is exactly that they're constantly trying to strip that harness down. Mm. That is, letting the model take on a bigger and bigger share of it. And if you think about the legal field from that angle, I'm actually really skeptical about, say, Legora or Harvey now. They're in the same position Cursor was in a while back, sort of. They've tried to build this kind of software around the legal side, so they could add some value at certain points. Mm. But I think we're already at the point where if you have Claude and you connect it, say, with Laki.ai to

[**22:55**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1375s) reliable Finnish legal sources, then for pretty much every legal-related task you'll actually get better answers from Claude than from Legora or Harvey, which have built a lot of harness between the model and the user. Yeah, that's a harsh way to put it, but hard to disagree. Personally, I actually only started coding properly again this year myself. I was a good coder as a kid, and I've been an investment banker for 30 years now, and I haven't lost a thing in those 30 years, because those backslashes are still sitting there in the terminal waiting. And actually, I do use autocomplete for that — when I SSH between different computers, tab-complete always finds, like, what did I even name this machine, so I still get autocomplete out of that, but otherwise I'm pretty much full slash-goal mode here — I mean, it's public sector, make-no-mistakes territory, just slash-goal — but somewhere there's

[**23:59**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1439s) a line, where your goal is so complex that you can't just do that — this is like an extreme example, something like "make no mistakes, fix the world" or something huge — so somewhere there's still a human path running through it, taste and intent and the steering that's involved, and — well, somewhere, in driving, or in editing a simple legal document, that works, but then if it's some multi-year, complex tangle of shareholder agreements between, say, your five family offices' portfolio companies, then, well, it doesn't quite work like that — there I believe a thick harness still has value, because it is just, somehow, so multi-dimensional that on a single dimension you can't play it. Um, so this kind of task-by- task, one vertical task at a time — yeah, those are getting chewed through constantly, so once enough of them are done and the system has learned, then it can pull them together holistically too.

[**25:02**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1502s) Yeah. And then there's maybe that one, sort of, dimension where Harvey and Legora are trying to compete, which is the interface. They have these UI components that you don't have in Claude, say, but I predict this will stay temporary, because generative interfaces seem to be coming on pretty fast right now. So soon you'll just be able to tell Claude, hey, I have this use case, what would be a good interface for this, and it'll build you that interface on the fly. Yeah. So the room for Legoras and Harveys of the world is going to get genuinely pretty tight. And it's really interesting, because they've sold it so well up to now — it's been sold incredibly well, the price is really steep, the contract terms are pretty long, but this is a really

[**26:03**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1563s) interesting situation in the sense that if we think about those law firms that have committed to these and built their whole process around Harvey or Legora, then along comes some smaller firm that adopts Claude instead — and especially if we're talking about Finnish law, right now in Finnish law Claude is just flat-out superior compared to something like Legora or Harvey, because you can't — in Legora or Harvey, as far as I know at least, you can't actually get Laki.ai hooked in there. Oh. Yeah. So it's that closed off. And they don't have an equivalent — they bought this Swedish startup, Kurigal or was it Kur Data, which basically has done in Sweden the same thing we've done with Laki.ai in Finland. So in Sweden they have that capability, but in Finland they don't have that capability. So legal research in Legora or Harvey in Finland

[**27:06**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1626s) relies purely on a web browser, and we know exactly what level that's at. Now a good example of this is, say, Harvey's own case too, where by default they've actually banned the model from doing this web search, and there's a very good reason for that [laughter], and now you've got this appallingly expensive product whose legal research relies on that web search — so is it really worth paying for? Well, neither of us is a lawyer, so of course we're of the opinion that it's not worth paying for, but — I do have friends though, who do just fine living off that lawyer's know-how, that corpus, and expertise, and their closed systems — and if you have, say, a top Finnish law firm, then their own internal system, with thousands of past cases in it, or you could use training material — you can still find that unique expertise there, the case-specific kind. I'd say — even though neither of us is

[**28:07**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1687s) a lawyer — Finland has weak courts and a strong Chamber of Commerce arbitration system. And this arbitration system is partly a bit like the American case-law type of thing, where, say, one or three arbitrators make a legally binding decision that binds the parties just like a judge's ruling would, if they've agreed to use the Chamber of Commerce's dispute resolution method for it. And they're not public, as far as I understand, so that body of information partly lives in law firms' own databases, because they're parties to these cases. Their partners might even be the arbitrators themselves, and it's kind of kept out of that whole open-data world of yours. I don't know if that rings a bell, that world — this kind of Helsinki Chamber of Commerce-type exception. Well, yeah. I mean, the challenge here, if we think about these law firms, what's

[**29:10**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1750s) difficult about it is that now a couple of years ago I was wondering which direction this was going, these general-purpose AI tools. Would they become so easy to use that anyone could use them, or would it get more complex? And I think this has clearly gone toward getting more complex. So now law firms traditionally maybe don't have much IT expertise to begin with, let alone this kind of AI IT expertise — which means that actually getting the benefits out of it, whether it's something like Legora or Harvey. Let alone something like Claude — to actually get the power out of Claude, it's exactly like you said, the key role there is played by your company's data, built up over the years. And now you need to get that into Claude's hands in a sensible way. You need genuinely good

[**30:16**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1816s) skills — which, as we touched on a bit earlier in this conversation — came up. So, skills means you can teach these language models different kinds of tasks. You could think of it like this: take doing a somersault, say — and say we have some young child, who doesn't know how to do a somersault, then the somersault involves certain things you have to do — so once you've taught that child the somersault skill once, then it can, sort of, whenever it needs to do a somersault, just pull out that somersault skill. And this is kind of the same thing — now you can teach these language models different skills. It could be, say, NDA commenting. Mm. Right, but you'd need to build it — and this was actually a really great opening move in that direction. This Claude — sort of the American Legal

[**31:16**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1876s) Skills, adapted for Finland. This Claude legal skills pack — Hakunikkola's GitHub has it, whether the skills — the skills pack — is something you can install with little effort into Claude, or as far as I know, into other systems too. [chuckle] Yeah. But maybe on skills, by the way — my own observation — for a long time I did them like this, I'd start building sort of alongside, like: here's a task starting, watch me while I do this task, and oh, stuff like that came up, and then we jumped around, and I realized this was done completely backwards — that you should first do the thing all the way to the finish line, and only then say let's turn this into a skill — once I've slogged through the whole obstacle course to the end, that's actually a better way to make skills, at least at this point in time. Absolutely agree. Yeah, yeah, but nobody really teaches that either — I mean, in that sense or, I don't know, there just isn't anymore that kind of Stack Overflow that everyone could go check to see how this stuff gets done — everyone's kind of learning it at random, I think, but it's kind of fun in a way. It's insanely fun, [laughter]

[**32:19**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=1939s) but the unfortunate part is it's a pretty small group who are excited about learning this kind of stuff. Yeah. Maybe I have to give Legora credit here too since, well, those Legora guys built that too. So at first I installed it, of course, thinking, let's just switch it on and, since I'm the DIY type, run it on this Gemma 4 26B model on my own hardware — and it was total garbage. Then I put Qwen 3.7, the seven, running there, which I already run, and that was a lot better, but then there was this problem where I didn't get their philosophy — like their Harvey-style model, where the idea is you have two different language models, which are, say, different — Opus and then Sonnet, or, well, the best Sonnet, which right now is five, and Opus 4.8, eight, since Fable got taken away from us — the way they, sort of, argue with each other, that sort of produces intelligence that doesn't happen

[**33:22**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2002s) with a single local model. If it's the same model going up against another copy of itself, then there's no real cross-battle happening, it's just — at first I thought this was total garbage when I ran it with that local model. It wasn't getting anything done. But then when I ran it with, like, the best Claude models, it actually produced a pretty okay base. But then when I ran it through your Laki.ai and then through these skills, it turned out even better. So I thought this was a pretty solid workflow. So like you make some kind of base draft that doesn't need to be all that polished, that can be just your own best rough sketch, and then you refine it — I think that's where pretty nice stuff comes out in the end. Yeah, absolutely. [laughter] Yeah. And maybe let's open this up a bit more for the listeners too — you just said Legora, but you actually meant Lovable for sure. Yeah, yeah, yeah, got them mixed up. So Lovable is just this kind of free program too, that you can install. It eats up some tokens once you install it. Though this is — yeah, Legora and Harvey, I've actually never used either, I have to say those are total black

[**34:24**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2064s) boxes to me. I just watch and see that, once again, the Swedes and the Yanks know how to cash in on that too. Yeah, I haven't used Harvey myself either, though I have watched YouTube videos of it being used. Legora I have used, and it hasn't convinced yours truly. Yeah. Lovable is actually kind of fun from a Swedish-coding-scene angle, because they sort of solve that whole minimum viable product thing within a safe authenticated setup, pretty neatly and with little effort, instead of you having to go and set up your own servers, or cloud installations. I still like that — I still occasionally do some little things with it myself, yeah, that's also interesting to watch — I think Lovable actually fits pretty well onto that sort of Pieter Levels path, whether it turns out that, as we go a bit further, I — I'll sign off on that. As a user, Lovable is still pretty handy, because

[**35:26**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2126s) it sort of abstracts away certain pretty laborious steps, like this whole cloud infrastructure provisioning and all this configuring, but how long is it before you're just sitting in Claude's or ChatGPT's interface, saying, hey, just whip me up something like this. [laughter] Yeah, that's how it is. Yeah, mainly for me it's that I at least know what the tech stack is, sort of. They've gone with a super simple one — you've got React and Node in there, and then TypeScript too, so I know that those are good building blocks — and Supabase and so on — and they work, and they don't really leak, as a rule. Whereas if I go and build it myself, then you can end up with whatever happens to come out — like, let's just build this on Firebase and do some of my own messing around, and oh, this doesn't even need this kind of authentication and that kind of API — and then, I can't be bothered, I'll just grab it off the shelf instead. [chuckle] So, so it's sort of like that

[**36:28**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2188s) noob-level-plus thing, that's what it is. And there's still stuff like demo sites or some small group things you need to log into, so it's not just totally wide open — and that's incredibly handy. And then it still keeps improving a bit, and they clearly run all those language models a bit crosswise under the hood, so every now and then you get a proper bonus dose of intelligence, when they happen to run it on some better model instead of dropping down to some cheap knock-off model. But what you said about Cursor a moment ago, now the great and mighty Elon Musk took it into his empire, so as we're recording this, they've released an X MCP, and then they run that into Cursor, so it's really interesting what my friend Musk pulls off there with that combination of Cursor and X and Twitter. Yeah, absolutely. I do have to pick up a bit more on this Lovable thing, because right now Laki.ai's development has definitely been

[**37:28**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2248s) mainly done by Juuso from Vesanto, who, uh, a few years ago — before Lovable even existed — showed me this kind of business case, where he described what Lovable has done in practice. And whether we should maybe build that. [laughter] We ended up deciding together, as a group, well, maybe we won't build this. Yeah. There's Replit and others out there, sure, that's eating into that market. But I mean, even Cursor was bought for 60 billion dollars, which is 30 times revenue — and when X was acquired, at first they didn't even have to pay with hard cash, it was just a 3% dilution into SpaceX, which is just an insane amount of paper money right there. So surely something like that will turn up for Lovable too, out of this, as long as it doesn't take too long to pay off. Yeah, right. [laughter] But yeah, but that's exactly the thing — if you think about what Lovable has done

[**38:28**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2308s) smartly, it's specifically those smart default choices. So they have good default settings. Whereas — I don't know, I haven't used Replit myself in a long time. I don't know what its status is these days, but at least earlier on, it used to leave the user with quite a lot of decision-making power there, which was maybe exactly the kind of thing users didn't want to do — they didn't want to think about what kind of environment this is being run in, they just want good defaults and then exactly that, and they've done that really well. Yeah. And so far it hasn't leaked badly. So just that — the authentication stuff and the Supabase instances — and then they've held up to scaling pretty well, they haven't blown up so badly that you'd have had to pay a ton of money for it when suddenly 10 million users are hitting some database — somehow they've managed to handle that too. And then I think they've also, in a way, in that harness layer — that dumb chat text box — they don't

[**39:32**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2372s) tell you what model they're running there, but they clearly always make that kind of trade-off between cost savings and then, on the other hand, whether you get quality out of it — and I think they've handled that too pretty well. As a rule, when I put in a smart prompt, what usually comes out is pretty sensible stuff, not garbage. Whereas if you put, say, Sonnet in there as the Claude model, you might actually get some surprisingly bad output — like [chuckle] — if it's just a bad token day and it doesn't feel like running on that 4.8 Pro, or whatever it's called, that ultra-ultra-effort setting. Yeah [laughter] yeah. On that note, actually, it's an interesting day today — yesterday evening actually, when I was reading the news, Anthropic announced that Fable is coming back today, just like Sonnet 5 came too — yeah, Sonnet 5 came out yesterday, and Fable is coming back here. Ah, has Trump had a good day, then, so that apparently — ah, okay. Well then, full speed ahead,

[**40:34**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2434s) because I was still in that post-Arctic15 hangover haze, so my last Fable run was mainly YouTube subtitling for this channel, so, honestly, that's about where things stand. Yeah, I have to say, luckily, when Fable was released — if I remember right it was available for three days — I hammered away at it pretty much morning to night, and I have to say it was truly a significant — based on my own testing and experience — a significant improvement over anything before it. Right, and you had that slash-make-me-rich-make-no- mistakes thing. It made money land in the account from a single prompt. Yeah, and there's — this ties into [laughter] — I think there's a kind of interesting story here, like what, or

[**41:35**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2495s) this kind of objection I actually run into pretty often from software developers. A lot of them say this model is already good enough for me now, so I don't need a better model, this is enough for me. To which my question is, well, are you at the point where you just give it instructions in the morning, and then check the next day's standup to see what's been done — and then they're a bit confused, like, well no, of course not, I write functions with it. Yeah, yeah. Well, I think sometimes you just hit enter too [laughter] as in, like, 'dangerously appro— ve.' The thing is, I personally see an insane amount of value specifically in the model's intelligence improving, because it enables — and now the point is, how you should think about it, I think, is just that we're still not at the point where I could treat a model like a good software developer. And that's exactly

[**42:35**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2555s) what I want. Yeah, I want to tell it, like, in a weekly meeting, hey, here's this week's tasks, and then we've got — whether it's a daily, you know, a 10 or 15 minute check-in where we look at how it's gone. Mm. So then the thing is, that these current ones are good enough — well, no, definitely not, because I want Fable back quickly, and luckily it's coming today, and then we've also got GPT-5.6 coming, which is interesting too — some people already have it on the approved list there. Exactly. Yeah, some already have it, but hopefully it's coming for us too, yeah, us regular mortals, probably within a week or two from around now. So we'll get to that shortly, but I have to say, when I started down this rabbit hole back in January, I of course picked up Claude Code and Cowork, and even switched to Mac because of it, since it didn't run on PC, so all that — I liked Claude's philosophy compared to that. Codex, well,

[**43:35**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2615s) already one-shotted things back then, but I liked it when I sort of learned that you have to actually watch Claude a bit as it fumbles around and keeps looking things over and making choices, then hitting enter, and sometimes going into the shell [chuckle] and having a look at what you need to press yourself, and oh no, that's not working — you take screenshots or paste this text back in, there was an error and so on. And that's kind of, for me, still a basic principle. And then, back when Claude ages ago switched to this token-based pricing, I decided to just cut OpenAI out of the picture, because it started costing real money — the first day alone was like 90 euros, so I thought, just drop it completely, and go with OpenAI's monthly plan for Codex as my daily driver. But at the same time, as I started using Codex, [clears throat] it really is kind of scarily good at one-shotting things — you give it the task, hit enter, and off it goes, and pretty solid stuff comes out at the end, so I'm not totally sure I like that. I think I'd still rather be the artist

[**44:37**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2677s) who steers the ship every now and then. I don't know if you have the same experiences. Well, yeah, I do. [laughter] I mean, as personalities, Claude — or if you think about the Claude models versus these GPT models, they're really different. And I do share that experience, that working with Claude is much more pleasant, so it's just a lot more human and more considerate, and it actually explains things — and that's maybe hard to explain. But there's actually a pretty big difference in how it feels to work with, whereas then, the GPT models are a bit blunt, or kind of arrogant — it's like you've got some kind of [laughter] really senior type, like some professor who's a bit high-strung, like, you're not really — you're a bit on edge when you go to tell it something, like hey, could you do this now? Yeah. This actually, by the way, Claude —

[**45:40**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2740s) when it went from Opus 4.6 to 4.7, it was a shock to me when this HR lady showed up, and then this security guy showed up too — or the other way round, HR and security lady — it was like, well Sami, it's already 8 o'clock, haven't we already done quite a lot of work here, maybe continue tomorrow — and I was like, what the hell, it's none of your business how long and how hard I work here. And then the other thing, this security business, where it flashed up something about an API key, like you basically need to delete your hard drive and go get a new key again, like this isn't going to fly, this is just a terrible security risk — and honestly, with both of these, I was like, I don't want any of this at all, and that was honestly part of the reason I jumped over to the GPT world, because they don't really have this stuff there, so — Mm. Yeah. [laughter] But that — there's actually been a really interesting blog post from OpenAI just a few days ago, they published. They'd looked internally into how many tokens from all their employees' AI usage

[**46:42**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2802s) come from Codex versus ChatGPT. Yeah. And there had been a really dramatic shift there — it was something like over 90% of all the tokens, basically all the output that these AIs produce inside the OpenAI organization, comes from Codex, even outside of software development. And this is something that I don't think has really in Finnish organizations, OpenAI - look, they go through Codex so no one's really caught on to that yet. [laughter] But the thing is, I think that's how it is - or actually, I do know that's how it is, that people just use the chat. Mm. Like if we look at it this way, inside OpenAI everyone uses Codex. Nobody really uses the chat - chat usage is minimal. Whereas in Finnish companies, and maybe in general, aside from these kinds of tech startups or whoever, who are actually in the AI world, AI adoption is completely

[**47:47**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2867s) still in its infancy. so what's needed now is a shift away from doing just little one-off things in chat. Yeah. Toward actually being well, if you're in the OpenAI ecosystem, having the work actually happen in Codex itself. Yeah, that's a good point. And actually now we get to that harness thing, or the OpenAI setup, or actual AI workers - not just these kind of dumb, in-between, whatever agents. So maybe two themes now - if these are the current front-runners - let's leave Google aside for now, even though they're trying to get into this game with Antigravity. So you've got Claude - Anthropic started that game already back in early this year. So we got that Claude Cowork, and along with it came the desktop app, so you've got the chat, then you've got Claude Cowork, which gets you into those office folders to work in, and then there's Claude Code. And at first for me it was like, I went from the chat into Code, and then I went into Cowork. Now I don't

[**48:49**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2929s) really know where I even want to be. It's almost the same whether I'm in chat. Actually I pretty rarely go into Code, because if I'm running code, then I go to the command line interface, which is the text-based thing. And then OpenAI's side was kind of whatever, before Codex came along - there was just the chat and the command line, but I never really used that. But now it's kind of like this: I actually run quite a lot of coding straight from chat, but then I go into Codex if I want to one-shot something. But then I run things with my own Samantha on OpenClow. So OpenClow is this kind of open- source project that was built by this guy, Peter Steinberger, an Austrian genius. So I picked it up like a week after it was released, and it runs on my Mac - this Samantha, which runs on its own physical machine, but it's clearly just hammering away on Codex the whole time because it codes and uses that as its main brain. So, um, I still think of it as the harness - you've got this smart agent, OpenClow, that's still a really big deal for me. But are you following

[**49:53**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=2993s) this idea of the desktop, these three pipelines, and then this kind of agentic building block - how do you actually see access to that intelligence. Well, especially if we go into this context of an office worker or knowledge worker, who works at some corporation, then I think the shift there - the first thing that needs to happen - is getting rid of Copilot. Yeah, it's terrible. And it's funny because I've done a fair number of these trainings over in Saudi Arabia, and I don't run into this problem there. It being a religious country, yeah, over there it's either companies operating in the OpenAI or Anthropic ecosystem, but I've never once run into this thing I keep running into in Finland, constantly. And the worst part of all of this is

[**50:54**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3054s) that we have companies that say, oh yeah, we're right there on the AI front line, and then it turns out, yeah, we've got Copilot here. [laughter] And Copilot has nothing to do with actually leveraging AI in any real sense. So that'd be my first piece of advice - [clears throat] you have to get rid of Copilot. And then as for what good alternatives there are right now, I think the best options are, like you said earlier, that Google has kind of dropped out of the game, so you've got either the OpenAI ecosystem or Anthropic. So you pick one of the two, and once you've picked one, you should try to make sure the work actually happens either in Codex or in Claude's Cowork - not in the chat, because this is exactly the point from that OpenAI blog post - yeah, that the chat is kind of like, if you have a question you want an answer to, or some small thing like that, then

[**51:56**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3116s) chat's fine. But in other cases, if you're actually going to do knowledge work, mm, and you're not a software developer - for developers, of course, it's Claude Code or Codex, but now if you're say, someone doing marketing or sales, that kind of person, you'd notice, okay, my usage - like maybe 90% of it - if you're in the OpenAI ecosystem, then you look at your monthly usage and you realize, okay, 90% of my usage is in Codex. Mm. Then you can say, hey, that's going pretty well. Mm. So that's maybe as an intro, that, well, then, um, these - I wouldn't necessarily dare recommend to companies these OpenAI solutions yet, at this stage - although now Microsoft - I haven't followed this super closely, but Microsoft is now, on some level, offering companies specifically this OpenClow

[**52:57**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3177s) Yeah, exactly, and actually Peter Steinberger also released it for mobile too. So yeah, it's coming - look, since Steinberger is at OpenAI now, he's not just some renegade kid anymore, cranking out a bit of sloppy code on the side - no, he's properly, like, right there inside the OpenAI empire. But maybe, in a way, the desktop thing is kind of interesting to me, because it might be that browsers basically die out entirely, that it's Anthropic's and OpenAI's and Google's own apps, once they get things sorted out - those become the portals you go in through and yeah, and that's actually a good point to raise - when I talk here about Codex and Claude Cowork, these, um - you can actually use Codex through the browser too. What I'm specifically talking about here are these desktop applications,

[**53:57**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3237s) meaning the desktop apps you download onto your machine and run things in. And that's exactly it - it's really interesting to think about why you'd actually want to go yourself to some website - like, does that need go away? It probably won't disappear entirely, but it'll shrink really, really significantly. What's interesting here too is that at least I have a really hard time finding tasks anymore that I'd actually go do myself. Like, what would be a thing where I'd actually be better than Codex? [laughter] taste, and long-term goals - 'make me rich' or something like that, [laughter] you need some kind of long-term vision, so that essentially the task-by-task work versus, say, what happens a year from now - the gap between those is still significant, definitely. Yeah, and yes, exactly, so even if you get there a year from now one task at a time, and

[**54:59**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3299s) even if every single task were done entirely by AI, you still need to first have some kind of direction. Yes. Yes, absolutely. Yeah. Well, that's the value a human creates, but the thing is, that you yourself doing something routine - like, this task-doing is maybe exactly the thing where nobody should really be doing any tasks themselves anymore - they should really be left to AI. And I mean right now we're doing a huge amount of all kinds of tasks which is honestly kind of crazy. Yeah, I think, if you think about what Musk did back with Twitter, or X - he threw out like 70 percent of the staff back in a time when we didn't even have these tools yet, so I think it's genuinely hard to find organizations where you couldn't throw out something like 70 percent of the staff and still have things carry on pretty much as before. Yeah, that's brutal. And the other hope people voice is that everyone would adopt these tools, and then everyone could basically keep

[**56:01**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3361s) their jobs and just scale things up tenfold instead. But the more likely outcome is exactly that - scale does grow among the ones who remain and who push hard, but then, in a way, people fall away too - and I think it's worth summing up like this: if your role at work is that some colleague or boss or client comes and says 'do this for me', hands you the task, and then you do it and hand it back a bit later if you remember, then that's not a great position to be in, because then you're basically pretty easily replaceable by AI inference. Yeah, that's [laughter] exactly right. Do you remember, by the way, there was a big fuss about that back when Mikko Alasaarela said on some podcast that he had calculated how many employees like that Kela would actually need, and it turned out to be some incomprehensibly small number. Like it was 18 or something. [laughter] Yeah, and 8,000 was the wrong answer, whatever the real number is, but yeah, they really should have

[**57:03**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3423s) ordered some 500-million Salesforce license, so right away, right away they went about the organization's task management the wrong way. But yeah, Mikko Alasaarela has been on the show many, many times, definitely. Although he's mentioned getting this AI brain fog - like, I haven't found it as fun yet as, in a way, this being a return to childhood, where I always get excited getting into it, having everything set up and available, and it's genuinely a joy prompting in some cool tasks. I think that's great, but yeah, I'm still working it out too. Yeah. And, [laughter] I really do have to give credit, even though, um, I'm generally maybe a bit of a well, or I do enjoy building these agents myself, but the thing I'm a bit skeptical about, and what I don't really like, is that a lot of these AI influencers give the impression that in order to actually get value out of these tools, you need to be, like, some kind of guru and build some agent farm

[**58:07**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3487s) and stuff, that it's really difficult. Whereas I'd argue that in most cases, or even almost every case, you can just take something like Codex or Claude Cowork and get pretty much the same results, if you just learn to use these off-the-shelf products well, right out of the box. But, um, back to the point - Samantha, which is this kind of divine, legendary being you built, that can do human-like magic - it's exactly that kind of magic box. You write whatever you want, and out comes the best answer in the universe, right? Yeah. Which [laughter] works. Yeah, which works in WhatsApp, specifically. And I think that's, again, a genuinely clever way to use agents, because the added value, in my opinion, isn't

[**59:08**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3548s) that if you ask it questions, I haven't seen that the answers are meaningfully better than what you'd get from, say, Claude or Codex. But where a lot of the value comes from is that it's now part of that community. Mm. Because something like that isn't possible anywhere else - it requires something like WhatsApp specifically - I think the WhatsApp channel is a great place where people and an AI like that can chat together, and Musk has actually kind of copied you over on X, in that Grok is there on X- slash-Twitter, in the discussions, and I think it works really well there, being part of that kind of conversation, so it can even correct Elon Musk's own statements - you can call it in to comment based on Grokipedia or whatever else, on whether what's being said actually makes sense. Um, but yeah, this Samantha thing, um,

[**60:11**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3611s) by the way, it's this - [laughter] - shameless plug - when you subscribe to Neuvottelija Insider and the bonus content, you can then get in touch with me, and once I check that you're a real person and not, um, a bot or some malicious infiltrator, then I'll invite you into this closed community of over 200 people, the Neuvottelija Insider community, where you can tag @Samantha, and from there you can ask literally anything and she answers so wonderfully - and by the way, everyone's been trying to mess with her for months, trying to crack her, to get her to reveal Sami's, uh, salary, or [chuckle] some file, or, like, pay, or whatever - but no, it's a beautifully built data structure, [laughter] but just a little plug there. But maybe, maybe a bit of skepticism - I appreciate that. It earlier, when we talked about this over lunch - a lot of people build it kind of however it happens to come out, so it doesn't add any real value, or it just ends up being another new communication [chuckle] layer, and at worst it's a security risk between you and the actual, real AI, i.e.

[**61:14**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3674s) say, Anthropic's best model - and in that case there's no point to it at all. So that's when it's just a dumb wrapper. But then if it has memory - so I was about to mention Andrej Karpathy's brilliantly envisioned wiki memory - it's able to itself pull from file structures that are semantically organized, curated, and nuggets of information I've burned into it collaboratively, from notes. Then there's this extended context memory - meaning even if there's a WhatsApp channel with a lot of text in it, usually the tokens [clears throat] run out because it can't hold more than a few messages in that short context window, so it extends that up to a file-level, kind of lossless context memory. So that's already a huge deal - it can see the whole conversation space, not just the last couple of messages. And then there's the secret ingredient, which is this Memory Lapse graph memory, which is exclusively between me and

[**62:15**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3735s) it, and it stores things semantically without any of that RAG garbage - just clean data structures. And then there's the five-layer security system that keeps malicious attackers out. So, yeah, it's lovely. I've been tinkering with this for the last four months, and it's pure joy just playing around with it, so I do like it, even though it's a bit silly, so but it's a great hobby. It's great, and an extremely useful hobby too - but the thing is, my point here, maybe, or my main point about why I'm a bit critical of these personal agent setups is just that some people are left with the impression that to use AI effectively you need to spend like 10 hours a day tuning these things. Yeah. Which just isn't true anymore - and maybe it was true at some point, back when yeah, those crash-prone setups were really fragile - but now it's good enough that, yeah, you can get really far with just off-the-shelf products, and these days they get you

[**63:19**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3799s) a whole agent farm to work with. I mean, with just a basic Claude license and Claude Cowork, you fire off some genuinely big task in there. For example Cowork just made me, like, over 150 of them, of which quite a few turned out genuinely usable - so it fired off an agent army - it wasn't just one Claude doing it, there was a whole agent army, working away on our slides without me having to do anything, like I didn't have to build any kind of agent system myself. It's just that you get it all, like, if you want an agent army at your disposal today, all you need is a Claude subscription and off it goes, and yeah, there it is. Yeah, so, I guess I started this back in December, I built a kind of second brain - a better notebook - and put it, at first, into this free tool called Obsidian, an MD-file-

[**64:20**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3860s) based journal system, for my precious observations about the world, and then I started building these skills into it, and then I noticed that both of these were kind of dumb - like, my terribly brilliant insights about the world that got curated into that journal, well, maybe they weren't actually that brilliant, and they went stale pretty fast, and they just ate up the agent's memory, since it had to figure out what Sami had thought about on some other day back then, i.e. in December 2025, and then those skills got even worse, because they basically just got in the way of the agent's much better general intelligence, so I just deleted all the skills, completely cold turkey. And then I also deleted this, like, my notebook - but I kept it around for the assistant, in a way, so that now if genuinely curated nuggets of information or important stuff comes up, then it gets burned into the Karpathy-style memory. And I think this has actually worked pretty well. Yeah. So even though general intelligence keeps improving all the time, there's still a role for a 'Samantha', one that you still can't quite get off the shelf. Yeah. Well, let's put it this way - maybe not quite. So now there's a good question,

[**65:24**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3924s) which is how far we are from that point, because right now actually - OpenAI, I haven't looked into it that deeply myself, but they announced recently that they've overhauled their memory system. It was previously actually implemented pretty poorly, and I didn't even keep it turned on myself. But now there is, yeah, um, at least according to their own marketing, they have a considerably more advanced memory now, where, and then you can have separate memories for different projects, or you can use the global memory - so they really seem to be coming on strong now in this area, and the thing is, it might get hard to compete with. I mean right now it's the case that at OpenAI, and at Anthropic too, they pay pretty solid salaries, which attracts quite a bit - well, Andrej Karpathy went there, and Peter Steinberg went over to OpenAI, so basically these memory-world gurus are on both sides now. Yeah, and actually

[**66:27**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=3987s) OpenAI - the whole 'Open Glow' project can be seen as OpenAI's public development platform for their memory architecture, so... Yeah, yeah, exactly, [laughter] I mean, it can be, it can be challenging, but still, I don't think it's a lost cause - because even if - well, even if it turns out that these so-called general-purpose, off-the-shelf products end up being, and probably soon will be, better than what you're able to build yourself, I still think that's really fun. Yeah. And a really good learning experience, because in any case, even if those off-the-shelf products turn out better, you've at least learned how to use them. You get to use those off-the-shelf products effectively, which gives you a huge competitive edge in this day and age, where knowing how to effectively use AI is honestly an insane competitive advantage right now. Yeah. And there's still some of that anthropomorphizing magic dust to it too, in that

[**67:28**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4048s) more and more people know what a 'Samantha' is. In a way, that's, in a certain sense, a skill in itself too, that I can - I've swapped out the brain from Anthropic for OpenAI's, and just for fun I set up DeepSeek too as a backup model, and then say, whichever's the era's best video generator models - the point being that the harness isn't model-locked - or rather, I can swap it with minimal hassle from one camp to another, or one API to another. And that, to me, is a genuinely strategic defensive capability. And then I could go full bunker mode. I've actually been thinking about buying, like, a 256 gigabyte unified-memory Spark, if I go the Nvidia route, or a Mac Studio - once I've got one running, bunker's ready, internet cut off, and everything runs 100% on local inference. So basically the thing is, since I don't fully trust that if I go with, say, Claude, they're going to try to build a moat, so that you just can't

[**68:29**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4109s) switch over to OpenAI - let alone Gemini - or Chinese models, not to mention. So self-sufficiency is, in my view, also a strategic competitive advantage. Yeah. Yeah, absolutely. Yeah. And nothing is as fun as tinkering with local models. Yeah, though it can be a bit of a letdown too. I was super excited - like, when Google put out Gemma 4, I thought, now we can do anything with this. [chuckle] Turns out you couldn't really do much of anything with it at all, [laughter]. Yeah, it's - I'm at the same time disappointed and happy about open source development. Like, disappointed right now because it's a fact that [clears throat] the gap is pretty big. If you think about, say, Fable or Mytos, and then something like GLM 5.2, which is maybe currently the most advanced open source model, unfortunately there's still something like a 16-to-12- month gap. Mm. And that makes me sad. But

[**69:31**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4171s) then in the same breath, now that we've seen this US policy, where, where we're being blocked from access to these best models, I'm honestly insanely happy that we do have a reasonably strong open source ecosystem and movement. Because if it turns out that these, for example, end up being available exclusively to US citizens, then we're not going to be left dropping into thin air here, so yeah, that's such a juicy topic that I'd suggest Aleksi Paavola and I head over to the members-only side. I'll put the Laki.ai link in the description - I'd recommend it, maybe with one small security caveat: when you upload confidential documents into systems like this, it's worth pseudonymizing or anonymizing those documents. If there are client or people's names in there, then change them to, say, 'Person A1', and

[**70:36**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4236s) then convert 'Person A1' back to Sami Miettinen afterward, once the run is done, if it's genuinely, truly a security- sensitive matter, like insider data on publicly listed companies. But otherwise it's a really good tool, and reliable, and this is genuinely amazing work - that in the spirit of open source, the end result doesn't end up being monetized by, say, some Alma Media - you've gone and opened up that legal interface for us Finns. Yeah, this really is a good, good cause, and it's genuinely such a sweet thing that now we're able to offer a fully competitive, or even better product, for free, to consumers, when the equivalent bought from Alma Media would run several thousand euros. It's just not reasonable or realistic in any way for some individual consumer who

[**71:41**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4301s) has some legal hassle to have to pay several thousand for something like that. And then these Legora and Harvey-type tools, which might not know the first thing about Finnish law, well then that's still another option for the consumer too, so for this consumer it's - and then, for companies, I think it's worth spending a bit of money on the use case, so that you build proper encryption chains. Actually, one thing that [chuckle] I could do myself too - this anonymization wrapper, I think, would be something worth building, maybe as open source, because it's the kind of thing that makes me a little worried about these MCP servers, if people go and upload documents that are too sensitive, then there's still, at least in theory, a small security risk. Well yeah, that's exactly right, that's how it is - what we're aiming for now is that the limiting factor would be your Claude or ChatGPT license. So if you had, say, a Business or Enterprise plan

[**72:42**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4362s) from Claude or ChatGPT, you could put all the same things into it that you'd already dare put into email or OneDrive anyway. And that's where we're going to be with Laki.ai pretty soon too. So then you'll be able to put in those, kind of, sensitive things, the same as you can now put into public inter- or, well, kind of into systems connected to the public internet. Like, for example, email - you can put them in there. But of course, you still have to trust whichever it is, Anthropic or OpenAI - so that's maybe something everyone has to personally weigh, what they actually want to put in there. And then, I don't want to feed the security-doomsayer camp too much here, but then, if it's a public use case that requires an EU server, then yeah, Anthropic's cloud might well be over in the US. So in these

[**73:44**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4424s) special situations - and you users out there know who you are - it's worth actually reading the fine print of these systems carefully. [laughter] Exactly. Exactly. EU legislation is going to be in Laki.ai pretty soon too, so and any day now the AI Act's final implementation is coming. Luckily the Omnibus has kicked some of the worst messes further down the road. But you can save on legal fees by hooking up the Laki.ai MCP and yeah, getting familiar with things yourself with AI. Great. But, thanks for this, Aleksi Paavola. This turned out to be way more interesting stuff than I expected. It's always great when I haven't googled the background too much beforehand - you always get pleasantly surprised that you're right, like, at the very core of the scene. This is always such a great thing. So, thanks for coming on. Thank you so much. This was a lot of fun, and, and if you've watched or listened this far, go ahead and subscribe to the channel, and let's head over to the members-only side and think about this data

[**74:44**](https://www.youtube.com/watch?v=ain1HvF4o7Y&t=4484s) sovereignty topic - I thought that was so much fun, so on the members-only side we'll spend a bit more time thinking about whether Europe still has any say in this at all, and what happens if, say, some future Trump cuts off our access to the better inference - so, see you on the members-only side then. Thanks. Thank you. [chuckle]

---

Cite as: Sami Miettinen, Neuvottelija — How the Laki.ai MCP Brings Finnish Law Into Claude | Aleksi Paavola, https://www.neuvottelija.com/podcast/episodes/737-finlex-suoraan-claudeen-nain-lakiai-mcp-toimii-aleksi-paavol/, 2026-07-18. For quotes include episode 396 and timestamp.