---
title: "AI Smarter Than Humans Within Two Years? | Mikko Alasaarela"
titleOriginal: "200IQ Tekoäly ohi ihmisen | Alasaarela | #neuvottelija 307"
episodeNumber: "307"
guest: "Alasaarela"
datePublished: 2025-02-20
duration: "52:15"
youtube: "https://www.youtube.com/watch?v=p4fuMR-znao"
captionsVtt: https://www.neuvottelija.com/podcast/episodes/482-200iq-tekoaly-ohi-ihmisen-alasaarela-neuvottelija-307/captions.en.vtt
captionsSrt: https://www.neuvottelija.com/podcast/episodes/482-200iq-tekoaly-ohi-ihmisen-alasaarela-neuvottelija-307/captions.en.srt
originalLanguage: "fi"
topics: ["ai_enterprise_tech"]
subtitleMethod: "publisher-english-cues-imported-and-qa"
provenance: "neuvottelija.fi Supabase episode-markdown API (platform captions)"
fiCanonical: "https://www.neuvottelija.fi/fi/episodes/482-200iq-tekoaly-ohi-ihmisen-alasaarela-neuvottelija-307"
canonical: https://www.neuvottelija.com/podcast/episodes/482-200iq-tekoaly-ohi-ihmisen-alasaarela-neuvottelija-307/
---
# AI Smarter Than Humans Within Two Years? | Mikko Alasaarela

## Summary

AI development will overtake human thinking capacity within two years at the latest, according to Anthropic's Dario Amodei (Claude) and OpenAI's Sam Altman (ChatGPT). AI model performance keeps improving through specialization, better reasoning, more efficient processor use (thanks to models like DeepSeek), and expanding memory and context windows. Mikko Alasaarela spent the turn of the year obsessively testing the latest AI models, an experience that convinced him that AI smarter than humans now looks almost inevitable and left him thinking about how that will affect us — one horsepower isn't good for much, so what use is a single human intellect? Sami adds prompting AI to his 3D model (delete, delegate, do) as a critical skill alongside rhetoric, leadership and negotiation. The members-only segment discusses human-AI collaboration, including from young people's perspective.

## Transcript

*English transcript derived from validated English subtitles ([WebVTT](https://www.neuvottelija.com/podcast/episodes/482-200iq-tekoaly-ohi-ihmisen-alasaarela-neuvottelija-307/captions.en.vtt) · [SRT](https://www.neuvottelija.com/podcast/episodes/482-200iq-tekoaly-ohi-ihmisen-alasaarela-neuvottelija-307/captions.en.srt)). Timestamps link to the original video.*

[**00:01**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1s) Horsepower is an old-fashioned unit of power cars produce hundreds of horsepower My guest is Mikko Alasaarela I wonder what the unit for human intelligence could be that's what we have to think about now that AI is approaching that one and you can't do anything with one horsepower! And then these cars have gears that allow for acceleration you can't do anything with a horse but one idea is that this kind of general AI that would reach human thinking power could be one human power unit so we are in a situation where researchers decided to build a test called Humanity's Last Exam humanity's last test We are so close to the point where AI or human capabilities that they came to the conclusion that let's make this kind of final test once this is passed there will be very few arguments left that AI wouldn't be better than humans Where are we at the moment with the officially available models Well o3 which is this ChatGPT's

[**01:20**](https://www.youtube.com/watch?v=p4fuMR-znao&t=80s) newest model and it's at 13% in this test but they already announced that in their internal tests they have already reached 35% So we got from the moment the test was launched to the point where a third has already been solved basically in a month so it probably won't take much longer Yeah some talk about a year that then this general AI would also be in use Yeah so Dario Amodei and Sam Altman have both said now who are the founder of Anthropic and then the founder of OpenAI they both announced that according to the current mathematical modeling the AGI moment is in 2026 which is next year Well then but horses surely have some aesthetic and athletic value so I guess we humans also have something, we need to rethink our role in this world and the universe quite a bit over the next years, one of Elon Musk's, well, sci-fi writer role models has been Iain M. Banks, who has this Culture culture where these AIs are benevolent and we are sort of like their Labrador retrievers and they tolerate us and well if we don't

[**02:43**](https://www.youtube.com/watch?v=p4fuMR-znao&t=163s) get in their faces too much. Yeah, let's hope that we can adapt in some way, our societies and we as humans can adapt to this new reality where this kind of intelligence use in problem solving and such shifts from us to other parties. Actually, I've had this big change over the last year in my own thinking that I no longer try to solve any difficult problems in my head, but I try to build these systems and frameworks, models with which I can solve the problems I want to solve better than I could myself with AI and this is a really new way of thinking for me and this has been really fascinating in some way an addictive way to use AI. I was listening last night to Osmo Soininvaara's memoirs, since he is this statistician and he was a six-time 'laudatur' recipient and he tells there something about himself that feels a bit like me, that whenever he studied mathematics, he wanted to get that kind of glimmer of certainty that he understands this comprehensively and intuitively correctly and I've thought sometimes that I used to do that quite a lot, so now I'm a bit like, if I don't really understand

[**04:06**](https://www.youtube.com/watch?v=p4fuMR-znao&t=246s) something, then I just prompt the AI more to, like, explain better so that I get an intuitive flash that this is now true. Yeah, but we're also heading toward a point where we can solve bigger problems, faster and on a much larger scale than before, so I've also given up on the idea that I should even understand every single detail of the problem; instead, I need to understand the whole problem in the big picture well enough so that I solve it the right way. But a good example of this was, for instance, my Christmas exercise when I tried putting all possible AI tools into coding assistance to see where I can get right now. And I realized quite quickly that the human role in that coding side is shifting at a very fast pace from actual coding to architectures and this kind of higher-level problem-solving, since when I produce hundreds of thousands of lines of code alone, there's no chance for me to remember every single line and every statement and every command and every module and so on in that whole entity. So it inevitably leads to me having to understand the big picture and I must have some kind of mechanism to measure what it's doing so that I can analyze whether it's doing what I wanted it to do. And from here we get to the kind of thinking that, for

[**05:31**](https://www.youtube.com/watch?v=p4fuMR-znao&t=331s) example, these front-line AI hackers use this term called pseudocode, and pseudocode basically means that you first build a model of how the code should work, if you think about what the inputs are, what the outputs are and what the logic is in between, and then you build that model of the whole architecture and then it's implemented as code, but you can give return to that pseudocode like with those tests and all the others about how like how this works and if it works and implement that pseudocode then you can then regenerate it like to match the one used by OpenAI and other common language models in a way we throw these Floating Point operator floating point operations with the best possible circuits so much that we build better and better and bigger and bigger well these neural network entities so this kind of flipped when there the Chinese had this FLOPS power removed through processor exports so then they had to develop in my opinion kind of maybe a new angle meaning meaning they brought this like reasoning what what is this term I mean reasoning reasoning into it so isn't isn't here like the idea that we rise from there like the neural network level kind of like to a higher level well what DeepSeek did they practically distilled meaning they like used these big models that for example OpenAI had

[**06:56**](https://www.youtube.com/watch?v=p4fuMR-znao&t=416s) developed to build these quantizations meaning they like had with smaller bit counts you can think of it like if you remember like a computer screen so you have now some kind of 4K screen where you don't see those pixels but some years ago when we played with Sami on our own our own gaming computers as kids they had some maybe 320 x 200 resolution oh awful already but it was like it felt cool that there were some eight bit colors yeah exactly but like in a way we we we still see what is happening there and we understand what is happening there so quantization is a bit the same thing that we go a bit less precisely but still sufficiently like we can like still understand what is happening we get the model into a much smaller size and this quantization is like surprisingly little It lowers, like, the quality of the result in relation to how much the amount of data that needs to be processed shrinks, and now we also at the same time have events like what DeepSeek did; besides this quantization, they made it so that you practically have, like, expert models; this is called Mixture of Experts, and, well, division of labor, so instead of trying to build this kind of world's smartest single generalist model for all fields that always fetches all answers from the entire data mass, we say, 'Hey, I just have this model here that receives questions and directs them to the right expert, and then that expert provides the

[**08:23**](https://www.youtube.com/watch?v=p4fuMR-znao&t=503s) answer; so we suddenly get the same thing handled much more efficiently than before, and we achieved really big leaps in efficiency with these kinds of approaches, where first the resolution was dropped a bit in this case and then it was directed to experts, and because the results were practically, thanks to this chain of thought, so close to what was being achieved with those large models, this woke the world up to the fact that even smaller players can still catch up to these big ones when they utilize these kinds of similar approaches, and that became this like, big shock to the whole market. Yeah, but you can still use that sort of horsepower, meaning the floating-point operation capacity, to run with those better processors, so you just run the model a bit harder, meaning those better Nvidia processors are still valuable, so you can run many more of those models there, and because as the cost of using this AI decreases, so does its utility in various problems increases and along with it, its use is growing exponentially; we aren't yet at the point where every person in the world would use AI daily, but that day is coming, like, quite quickly because it's just so affordable now to adopt and utilize in all situations, so it naturally spreads through the whole market. You can think about this— well, Kahneman has this dual model, which now

[**09:49**](https://www.youtube.com/watch?v=p4fuMR-znao&t=589s) are always, like, distortions of reality, but in the human mind too there's this intuition, a fast process, and then this slow, reflective process, so well, you could think that intuition is some kind of general, like, encoded model from which that vibration comes— 'this is true' and a fast answer—whereas then the reflection vector is, with our processing power, ruthlessly slow digging, but from there comes the more correct answer; so then AI— well, since it can do this— especially this reflection channel—infinitely much faster than us, and since our intuition is quite, quite crappy—well, a poorly encoded, well, general language model anyway—so I guess we have lost this game; so yes, simulating human thinking—we now see it in practice as these models show us how the AI model thinks, like, in different steps in that reasoning chain—so the single biggest deficiency we've had regarding AI being able to, like, start autonomously doing things in the long term is this context window, and the context window in these models has been limited since it originally started back then with, like, the first GPTs, so it might have been just some, well, a few hundred characters, like, what you can put in, something really small, but then it has been growing constantly; currently in these large models, the typical size is 8,000 tokens, which is in practice

[**11:12**](https://www.youtube.com/watch?v=p4fuMR-znao&t=672s) about maybe 100,000 words and and it Well it's already quite long You can do quite many things with it already but it's still a bit like if you had a grandmother like this who has, well, a poor poor memory and she remembers like your conversation for three minutes or five minutes or maybe ten minutes then after that everything that's been said starts to fall out the other end and you have to start from scratch and since we now have these chat windows remaining like this short, then the AI can't make any long-term strategies there because like it forgets everything as the conversation goes along Well in Google's Gemini advanced at the moment which has now also been announced for public use so it has two million tokens and and the biggest model that exists right now is the Chinese MiniMax it has a ten million token like window and now what like this year everyone is trying to solve is how we build such a like ability to take the best and most important pieces of those conversations you have and and save it there for the AI's long-term memory so that it gets the ability to think long term strategically Yeah and a human's working memory is at least well only a few bits if you think about it like we process it visually for example quite large sets of bits and they are tossed around inside the head but so you could also think about this one human intelligence if we go back

[**12:35**](https://www.youtube.com/watch?v=p4fuMR-znao&t=755s) to this horsepower comparison then well well since you mentioned these well low resolution games so well Doom so well it runs just about anywhere nowadays that I don't really know that joke that you can run it on absolutely any device nowadays so so well that at some point this like one human intelligence so so can be packed probably into quite quite like depressingly small devices so Well exactly That Doom is running on a USB cable right now. It feels quite funny indeed as an idea, but on the other hand, that same, well, processor power with which Doom runs was also the processor with which humans were sent to the moon in the 60s, so well, you can make quite incredible weapons in a small kit once these are available for better or worse, maybe in a year's time already these human intelligences. So then there is still some kind of salvation, I guess, in that these first language models were trained on so-called quite clean data, humanity's like bona fide or like scientific knowledge created without hidden agendas, and of course it always had some fake stuff and like propaganda on top of it. But aren't we starting to walk in the mud now? Yeah, regarding that training data on the internet

[**13:57**](https://www.youtube.com/watch?v=p4fuMR-znao&t=837s) so we are at a point where like about two-thirds of the internet's content and traffic is like bots, so you have like AI-produced content as the majority of the internet's content currently, and the majority of the traffic moving there is also bots and the forecast currently is that the level would be about 90% by the end of this year, and it's easy to see why it's going that way, that currently a huge part of all content on the internet is already produced by AI. And and this dev- isn't exactly beneficial for humans in the sense that think how much easier it is to produce content there for an AI than for human labor and and this like forces us to rethink the role of the entire internet, like previously when you went to the internet, websites were built there, for example, where the purpose was to guide humans to behave in a certain way following some specific information hierarchy, like if you go to Amazon, for instance online store you're clicking some buttons there and then some recommendations come and you're like influenced by them or on social media for example you look at the feed and then it's optimized just for you a personal feed that produces those emotional experiences that keep you hooked on social media so think about it like if now at the end of the year if we follow this trend that's already been predicted that

[**15:22**](https://www.youtube.com/watch?v=p4fuMR-znao&t=922s) 90 percent of it is like bots then it won't matter anymore like the whole internet designed for humans is basically useless and we could start thinking at this stage about how this internet should work so that it would work as well as possible with other bots and it's a very different-looking internet than this which is designed for humans that's right and the level of sludge that's there if you go there with human power with your own little eyes then you won't see anything there you need like an AI partner who sees the real data through that sludge so this game is lost already a bit even in terms of sensory channels here yes indeed and this is one of those new situations where human behavior shifts from searching for some website on Google and then scrolling to see which of these thousand options is actually my page m-hm so we go much more easily now to that bot to ask like hey tell me a summary of this thing I was interested in because we get an answer to our question directly then and all these kinds of internet commercial models and such will change as a result of this like very strongly um what is like the challenge in today's AI development is of course exactly what you're saying that if you now you go and scrape that internet like for training data. And it is actually full of content from older AI models, trained data that is full of all sorts of hallucinations and other

[**16:48**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1008s) problems, so if you use that as training data for your new model, it is worse than the old one, practically speaking, and now we should think about how we can move forward from here. Yeah, well, they are trained with something, so we can take this philosophy of science or philosophical angle that back in the day we had Descartes, who said, like, the biggest difference of rationalism was that 'I think, therefore I am,' and then came these empiricists, and in a way, if you can no longer trust your senses empirically, then I guess you should go towards rationality, so maybe one solution is these AIs just start making their own logic and thinking a bit mathematically and don't care about the data created by our human senses. Well, that's already visible currently in this AI training model, that most of the AI training material is produced by other AIs, and and that involves certain filtering processes and others by which the quality is then measured and checked too, and human thinking is also used there, meaning that now lots of people around the world have been hired, like, to analyze the results of AI outputs and such, so that human insight is also

[**17:52**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1072s) included. But, but it's quite clear that future AIs will be the kind that learn from their own operations, meaning this kind of ability to learn from the environment and the ability to store that learning into their own operations has grown. I just read a scientific article like that, which came out yesterday or the day before... very fresh so so Chinese researchers had succeeded in making a model where they could with 817 question-answer pairs, so well, fine-tune, meaning like in a way improve the existing base AI model to a level it can create these chains of reasoning, so they can add with a very, very small adapter like this uh, the AI's ability to do chains of reasoning and I think this is a good example of how we are constantly inventing more efficient better ways to train these models too, and this basically means that in the future well I am quite sure that most AI will be able to train from all that conversation you have with it, an adapter that

[**19:11**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1151s) acts as a lens for the data that you use with that AI. It can like in practice put itself in your position better in every situation and this will be such that every person has like their own, own, well, like these kinds of behavioral memories with these AIs and it makes them, of course, then we can build empathy, like, for your situation, taking things into account and other things for the AI and other such things that make them certainly more interesting conversation partners than other people. Yes, but then if you think, if you again remove these senses, then well, for example, how a bat thinks or how it like senses the world with sonar, but how can we understand when we don't have sonar, how a bat's world is formed, of course we both live in the same place as the bats and our too then so your experience, like, of the color of these curtains might be encoded differently in your brain than mine, but, I mean, in a way, does it then sort of start forming for these AIs that are more intelligent than humans some like, completely their own way, like, a bit of a totally alien way of thinking. Like, a bit like that bat, the bat there, they have some sonar with which they sense and then communicate with each other, for instance. Well, I mean, an AI neural network is hardly a direct copy of the human brain, so fundamentally it already works differently. Of course, it's true that because it's currently trained

[**20:39**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1239s) mainly on thinking produced by humans, so the symbolism and such currently stem from humans. But we've also seen these models where we've been able to build entirely new pseudo-languages and ways to communicate that are even better and more efficient than these current human sym- bols for various needs, so we don't have any kind of certain path or such a known model here with which we could ensure that AI would remain a human-like thinker as it develops to be better than us, and in fact, if we want to build real super-intelligences that are, say, a thousand times better than humans, then it's very, very likely that they work very differently than the human brain by default. Yeah, about this, in a way, it wasn't really augmented reality, but sort of on a different band, a bit like light, infrared light or sonar, or like they're completely different spectra and we sense in different ways. So, here I interviewed for my book, 'New Negotiating Power', this guy Honkela, 'The AI Researcher's Testament,' and he had this theory that, well, like, these AIs converse well, um, like, they change people's, well, like, needs and meanings and

[**22:04**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1324s) like, shaking hands with bricks, aggregating people's, well, needs and demands and sort of like a bit of a separate transaction layer there in the background, so could we humans be, like, involved in the development of that, so, I mean, it would be great if we succeeded in using AI to create world peace, for example, as an anecdote, one Canadian acquaintance of mine just yesterday did a test with that, well, Deep Research, which is like this ChatGPT's latest thing and the one included in that more expensive subscription, like, the deep research model, and he asked it to build a world peace concept and shared its results, which were quite interesting, like fundamentally, yes, here people are probably at least trying to find these kinds of paths in this regard, though we are now seeing in practice the fact that the scary aspects of AI development are now very much on the surface because our pace is so fast that no society or, like, we humans—I personally have always been at the forefront of technologies and such, and even at this moment I'm going, I'm involved in international AI hacker circles for the very reason that I want to be right there at the forefront of the forefront, and I try

[**23:23**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1403s) everything as soon as they come out, and in this position, the biggest challenge I have now, compared to maybe before, is that I might get many ideas a day about what new problems could be solved with AI or couldn't be solved before, then the next day even that is already disrupted when yet another even better idea has come along that works even better, and this, like, this pace is so incredibly fast at this at the moment it's hard work to keep yourself like on track with everything that is happening and even harder work is like to come up with anything that wouldn't be outdated by the day after tomorrow and that's the pace we're going at there on the front line but then among the common people we still know that you can chat with ChatGPT with AI so it's like a quite different reality but when I think about how quickly it moves, that reality from the front line into practice and if we think about this, uh, predicted timeline that this kind of general AI is practically like probably like a mathematically provable thing in 2026, then what does it mean for our societies and what does it mean for how we should organize ourselves in the future as humanity, then it's the kind of thing we haven't prepared for, we don't even have the ability to prepare on this schedule, the point where we will surpass like human beings in AI capability like before we have any clue what we what we'll do after that and and some exciting times will come

[**24:46**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1486s) inevitably because of that yeah so that difference between cutting and bleeding edge that well there at cutting edge you just like cut the wind but then at bleeding edge blood is there and body parts are flying and this is now well if I take this Bostrom's urn analogy for humanity, well, the collective if there are like jars of always white well marble balls and then they are picked out these new technologies and developments and then there are a few black balls which are that kind of existential like destruction and and well here then well hop- dodging those balls here yeah and it's true that like OpenAI's original mission was was well that the good guys would cross the line first in this well AI's turning into superintelligence and then they themselves turned practically into just a purely commercial business where the 'good guy' element was completely left out. I don't know if anyone is currently pushing for it anymore, like, is this just a case where we can't do anything about it, that the market economy works how it works. And when you need enough money to do something, then ultimately that money is power, and if we look at the US like 500 billion investment they plan to make in AI processing over the next, was it

[**26:10**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1570s) three years, then those are just massive investments. Then at the same time, it was proudly announced that Europe has now made this kind of AI program to respond to the American challenge and its budget was 56 million. Then I thought, 500 billion is 10,000 times more money than what we Europeans are putting into this, so I don't really see any chance for Europe in this race. That's how it is. As a teen I read Isaac Asimov's, well, those Foundations. But he had this Invincible book where this kind of, like, the best killer of space technology came to Earth and then it lost to these, in a way, nanobots. So this was the story then, that maybe here is this DeepSeek story possibility, that those American, well, 'invincibles' will fall to these kind of DeepSeek-level things. And then, as an existential crisis, I mean someone could decide to get rid of those Russians again with these small nanobot armies, and let's hope they don't then come up with other ideas.

[**27:13**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1633s) So yeah, we can't prevent that either, that as long as wars are fought, war-related technologies are also developed. And the level of autonomy there has already risen and, well, it's a scary scenario. I still see this societal disruption with this at the moment, like, in the short term, as a much bigger risk than some drone nanobot army attacking and destroying us all, so we have that bigger risk right now that we as societies can't respond to this change and then we drift into chaos as a result, if you look at the ongoing chaos in the States, well, on the other hand, it will be everyday life in Europe next year if we don't do anything about it, yeah and they have the resources, however, with which they can burn a few, you know, inefficient hours there, at least for some people, we don't have that condensed time, meaning money, very much here, especially in Finland, but maybe that transformation of work, so in a way probably some kind of AI butler should be trained now pretty quickly by everyone, so that, you know, you can somehow lift your head in that AI sludge and I'll put one of my Negotiator Channel thought loops here, that when these new models are trained, well, they can't learn anything from that AI-internet sludge, but when there's this kind of curated Negotiator Channel that's visible on all six platforms with hundreds of hours of only

[**28:40**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1720s) quality guests, well, that's where they'll quickly look, this data that looks right shines through on all platforms, let's take some training material from here, so you business leaders who think that some Google Ad search or your websites matter, just come here for cooperation, did we get a deal, so from this we get to the fact that in those AI hacker circles they talk a lot about human prompting, and yeah, well humans can be prompted too, and actually if we think about, say, social media as a platform, they are constantly creating these kinds of triggers like behavioral triggers such as that creating this state of mind where you go and like something or comment on something and and and this algorithmic and emotional triggering has been one way like using AI to guide human behavior and at the moment of course in those hacker circles then they are looking at where this like this new AI world is leading here that if we think about for example DeepSeek then DeepSeek also has other agendas than just like in a way destroying the argument from Americans that this game is won with money this has been like one part of that argumentation but but like behind it also the Communist Party's propaganda which is like baked into that model and then when at the same time they put the cheapest servers to run that DeepSeek

[**30:05**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1805s) in China like where where it is and all the data goes there now so they get all that data from people's DeepSeek usage for China's use as training material which is like a huge benefit for them it is also funny that they say quite explicitly that this data from the West doesn't go to China and then someone had just put a dongle in between to see if it goes there to China of course it went there you shouldn't like even think that it wouldn't go of course it goes when you have these actors like Perplexity as a good example which started using DeepSeek which has in a way removed the propaganda layer from it then like on top of DeepSeek and put it on US servers so in these use cases it is safe yeah because it is an open-source model so it can be modified and you can host it yourself so that it doesn't go that data to China and and of course in this case we get the benefits to use these kinds of new breakthroughs without us having to give up, but the situation is still that most of the DeepSeek usage is on those Chinese servers because they are the cheapest and easiest for people to use, so this is kind of like the TikTok of the AI era, so we got that layer in there, but I'm thinking how clever this model is, that DeepSeek immediately became the most affordable way to code, and now you have millions of coders who use AI-assisted coding tools, and now in those

[**31:29**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1889s) AI tools, they have, for example, environment files that they are using and such, which those AI models get direct access to, including all API keys and passwords and everything else right from the developers' desks, which isn't visible to a normal user anywhere, of course, but for the developers, they are visible there, and they're only visible on their own machines, but now that AI can read them and send them to China, so I'm thinking that we might still see all kinds of crazy things because people happened to use it in some coding job or didn't think it through that, wait a minute, here was a model that just read my passwords from here and threw them to China. Yeah, the risks are high, like there might be some crypto attacks. Isn't this also, in a way, cybersecurity is now one battleground for language models and generative AI, so that there well, the Mikko Hyppönens are now sorting through those viruses still, but you can't really react in time anymore when such an AI-coordinated attack happens, so you need to have your own AI defending. Yeah, it's an interesting new world and there's definitely work to be found.

[**32:38**](https://www.youtube.com/watch?v=p4fuMR-znao&t=1958s) also for these cybersecurity people here for thinking about those things and and surely they have to use their own models that like respond to this new new well like approach which here Yeah in this is in my opinion a bit of this Butlerian Jihad term so from that Dune so the AIs were kicked out of it like from this Frank Herbert's wor- ld because because they drove wars and such so that let's go back to the caves So didn't your Christmas holiday also go into you running quite like on a powerful Apple machine like there like on the RAM level running these AI models Yeah so I tried everything that I I like I I gave myself like a relatively unlimited budget like anyway I burned like probably about 700 euros there during the holiday like AI credits for different different models and otherwise and such and then well I took all possible paid models and others into use and and indeed upgraded my machine most recently to like a MacBook Pro and and and so on and the idea here was just that I looked at what what's going on like to get a bit of understanding of this that what is this state of the art and what came as a surprise to me I mean I haven't been like a rank-and-file coder in my life ever but I have always coded here and there and I understand like quite well how how like certain kinds of codebases work and as an architect I've been in quite many many different projects and then like in a way taken the CTO hat quite many

[**34:03**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2043s) times on my head when projects were done so so I did it with like still such an approach that I thought that this needs to be designed well and because I'm not going to understand nearly all that code which comes from there and because of that it's such a huge amount so quickly that that not nobody even has time to read it like here and and well it went like this that I first created hundreds of pages of documentation meaning I had conversations with different top models like o1 and Gemini advanced and these like whatever were on the edge these best reasoning models at that moment so I had long conversations with them about what would be the most optimal and best architecture to solve these types of problems I now took on a task that seemed impossible at that moment this is also funny by the way that I took on the task of building like an agent architecture that builds this kind of evolving graph intelligence by coding itself so that it would sort of be able to iteratively become smarter just by coding new features for itself meaning this was like an agent architecture and here is now the

[**35:25**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2125s) idea of this kind of recursively self-improving model that eventually ends up at AGI and that's like basically a totally nuts project there's no sense in one guy starting to do that like getting it done during some Christmas break but that was exactly the point like when I started to look let's see what can be done and then I like throw in all all the skills I have plus all the AI skills that I can and know how to use and I quite quickly learned how to like use them in sensible ways and especially when joining hacker groups like being able to learn quickly from the best people on the front line how this should be done and I learned for example this pseudo-coding mindset meaning that if you want to build big architectures then you first have to build this kind of pseudo-code for it about architecture, so well first you have to build goals for it for the architecture and iterate as many times as needed on that model level through the process so that you get the ideal result or at least you probably won't get anything final, some kind of local maximum in this case, and then when you find the best approach and you choose, for example, the programming languages based on that. So now when you go into AI coding, you don't necessarily have to use the one that is most familiar to you, you can use something else too, and well, now that we're getting into it, then we build, like, in a way, the architectural description, and then that architectural description is vetted with another model to ensure that this went well, and then build pseudo-code from it, and then just run tests for that pseudo-code and then the code is built, and practically with that model you get this kind of high enterprise-level code from one

[**36:50**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2210s) guy out on a very, very, very fast schedule, in absolutely huge amounts, and then you can even run it at the machine level so API calls don't cost anything, so, well, probably 700 euros doesn't sound too bad, so, well, in my opinion, it was quite a fun fun thing to spend on a nice hobby like this, but it is more because I tried everything, and I just, like, I just said I don't care, I'll just test it now, and I ran, like, models; a big part of it went to OpenRouter, where I used, like, all different AI models through the API, like, in that system, and there was a huge amount of this kind of iterative reasoning through those API calls. Yeah, somehow it feels like a kind of return from that sludge the Cloud has been stained by to a bit of a hardware world like this, or that's where I kind of agree with this Apple philosophy that AI should be brought to these terminal devices, and for that reason I resent the EU here too, that they try to prevent it, because I think, like, the closer it is to you physically, it's sort of like your trusted butler or stew...

[**37:58**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2278s) the better, and precisely all these codes, biometric, like, logins and such, well, it somehow feels like an intuitively good approach to me Yeah, and it's the same, I mean, that was the Mixture of Experts thinking from the start, that like, you don't need to try and build some model that tries to know everything about everything, you can just put a bunch of models some model can be in the phone, it just interprets your speech into text or text to speech, or just recognizes your, like, facial features or like your fingerprint, and I think this is just like an obvious thing that it's worth doing right there on the spot, then you have things where you might need to, say, fetch information from the internet, then of course you go and get that information from the internet in that case but I mean, in a way, I was just trying to see there that there are some basic building blocks in the library, I found that for replicating PowerPoint text elements into another PowerPoint, but then somehow, like, these algorithms for generating graphical icons, at least my Perplexity and other searches didn't find at all, like, in a way it's quite a basic thing that, like, you see a beautiful slide in image form and then you turn it into your PowerPoint material, so something like this didn't work, at least in my...

[**39:19**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2359s) capability, and not even search capability, so there's probably that approach will be a bit different then, when giving these like mind maps and others for AI to do, I think they are coded more than produced as an image, and this is a much more efficient and easier way to build those versatile ones and the elements appear as elements, but there's still a way to go of course, until we reach that final quality, and I've noticed when I've been coding those that if you prompt the AI to build me a world-class, like one of the world's best user experiences for this website that you're making, then they're quite well able to take the most sensible libraries into use for elements and others and to make all sorts of like for example, I built one glassmorphic user interface just with the idea of seeing how it can do all kinds of animation effects and others, it works surprisingly well, so yeah it requires a bit of understanding of what you're looking for, but if there's enough understanding of what you're looking for, you get quite beautiful user interfaces made with AI today. Yeah, so that UI and service design was a bit of a parallel; for the first time in a while I sat here in the studio for a few hours with Juhana Torkki. He visited a

[**40:39**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2439s) significant Finnish client for negotiation training, so we brainstormed a bit that it's no longer enough to just teach a bit of negotiation roles and the flow of negotiation and some of these general gambits and tricks and such, that now you have to sort of tailor it to the customer and the negotiation situation. like service-designing that negotiation process, in a way, the moments of value creation and the interaction points, like, through which you then win those contracts, for example, in a form that's good for you. So, in that one could still find quite a lot of work to do. So, in a way, this kind of AI-designed, well, process. And its manifestation isn't just a homepage, but it's like the way you conduct those transactions, and in its service design one could still find jobs for people, but when you think functionally about something like communication or marketing or sales, then with all due respect, it's already quite low value-add work. Well, if we go back to the human's role in this new era, then if we

[**41:54**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2514s) accept the fact that AI can simply process more information faster than fits in your head and what you can process yourself, then isn't it the most sensible strategy for you to take the role where you build those processes and frameworks and models through which you can use the artificial to get more out of it than your own brain would allow, and this is, in my opinion, the new mindset we are in. That is, well, let's say that from personal experience now that I've been doing this, I can't even go back to a situation where I would develop some idea, like, without the help of AI because I see that AI brings so much new perspective, things I might not have considered myself in this situation in my own head, and then I can still very easily and quickly, like, test those ideas by saying, 'Hey, let's test these against the world's toughest benchmarks' to see if this is actually a good idea, considering What other options are currently available for this and then you get good deep research analysis from there about how like actually it was already invented three years ago here and there and there and then here it's been done this way this is actually even better than what you had now and and and this way you can iterate and find like those kinds of approaches that you haven't previously maybe thought of and absolutely insane productivity leaps will come like like for anyone once they just learn to use models like that Yeah personally I've

[**43:17**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2597s) had this well it's based on some real model well I just came up with it myself well like from a slightly funny different angle so this 3D model like delegate delete do So now well you have to get the AI like onto the level of all of these well like floating in a way into that chain of command so so so here well then this model's like mentoring or teaching to others since we all probably have to live as a kind of AI-human-entrepreneur combo like in the near future to get that added value scraped from quite complex service designs and we do it for other similar units so so well those interfaces and teaching them still happens from human to human and somehow that trust and purchase signal to make the transaction is still between humans at least for a while so we don't yet know that just think if all the big corporations now end up end up with a solution like let's get an AI here to screen all our procurements then it might quite quickly go to like bot-to-bot negotiations these like these big pictures that we don't know it yet but I mean it's like a completely possible like like development path as well which is like ahead, but one thing I have to say to all the viewers is that

[**44:40**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2680s) I have acquaintances who are CEOs and, well, you know, in various leadership roles in their own companies, so we have discussions in these private channels and such with them, and then they mentioned that I was looking for, say, a CTO or someone, and then there were these candidates, and only one of them was an AI native. You can guess who I continued the conversation with. So yeah, that's like the, you know, kind of the situation we're in, so, yeah. In my opinion, it's a pretty bad idea not to be on that train right now if you want to be relevant in any way in the current job market. Yeah, and I mean, back in the day I tried to illustrate the importance of negotiation in Finland, that we have two skills that somehow aren't taught—well, actually three—like, rhetorical skills were for a long time like, no, one had them in Finland, everyone was just quiet and sticking to their points, and well, introverts, and then negotiation and then leadership, and these somehow all get learned from your immediate bosses in a really random way, somehow getting shaped by it, it can be anything at all, so in a way this delegating to AI—now I would raise this as the fourth skill there, definitely. So, we should now get this, like, national AI delegation thing—since we once had that great AI course from

[**46:04**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2764s) the University of Helsinki, well, make that legendary one-credit course about this, delegating to AI. Well, still, while we are talking about this, as we're on the way toward this AGI and, you know, the level of AGI will likely exceed human capabilities quite significantly once it gets there, but in this transitional phase when we're still on our way there so so uh it's completely inexcusable like if you don't use AI as an advisor in all your decision-making so Maki has built these like separately prompted advisors for all situations so it's told to this advisor AI first like I usually generate these prompts with these big ones like some o1 or well Gemini Advanced so I build through conversation this kind of Advisor profile so what this needs to know this advisor of mine in this matter it can advise me best in this and then such a prompt is created for it to launch that expert and now then it has like an expert who has material plus plus well then also such instructions on how from what angle it should advise you so then you can have with it like a conversation with that expert about like Hey I should now solve this problem so now this expert gives me this this perspective's advice and another gives another so you have like all the time like an advisor team around you that helps you in all those tasks that you do and like I wouldn't be able to im-

[**47:27**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2847s) I'd do anything at all without that advisory board that's there around do you remember that book like Napoleon Hill's Think and Grow Rich I have heard of it yeah I mean it's like quite quite it was this great well a bit like Dale Carnegie's How to Win Friends and Influence People and is one of the most formative books I've read so but it had at the end such a thing that felt quite woo-woo back then but he had learned from these like rich people of their time well that well they have this like hallucinated well advisor group and then it like recommended in it that you can have for example Thomas Edison who gives you like tips like scientific research and then you can have like Jesus who gives you like moral guidance and then like some Gandhi or whatever you take, I mean those kinds of like, in a way exactly we come to this like where we started, these, what were they, dedicated like like AI processes, so in a way this idea too was hallucinated some 100 years ago quite funnily, well now that you mentioned Jesus there mentioned, I was actually just a little while ago in the Archbishop's kind of AI roundtable involved in discussing it, in fact there was just a story about it today, I was involved in this particular discussion, it was a very interesting discussion when when talking about this kind of like humanity and ethics like and and AI, like the Vatican had done

[**48:49**](https://www.youtube.com/watch?v=p4fuMR-znao&t=2929s) some white paper on this, so yes like yes it is quite interesting reflection like like from this perspective and I think it was very high quality discussion and it was really surprising that the Archbishop asks to talk about AI, yeah maybe we can put the Archbishop there in our like pantheon to give us like tips then from this field, but yeah so this kind of certain kind of like pantheon so like different different like Greek gods then like or Roman if you happen to like them so to give like like diversified advice to you, there could be one, though you can then for some Zeus like delegate all those sub-tasks then, like just explain to me better and delegate these onwards from here and then if Elon Musk gets Neuralink to work, then it's directly like no no we won't go like at bit rates of a few bits then like in, but straight into the corpus callosum, these human brains are in the end quite quite like a small processor power computer and I always told my own children when they were young that think of your head like a computer, so the more you throw data in there, the bigger the memory it has, and the more you use that head, the higher the processing power it has, so now it's worth investing because in the future it defines how good of a computer you have in your head when you have to do those things, and well, that's

[**50:13**](https://www.youtube.com/watch?v=p4fuMR-znao&t=3013s) like, in my opinion, quite a fun metaphor with that idea, but our communication channels are very slow, here we speak a few words per minute or at that pace, and communicate to each other, think about the rate at which data moves on the internet, like we don't stand a chance with our pace of communication, that's how it is, I mean, I also with that paid 4.0, when I watch some slightly boring US discussion for two hours, well luckily there in the top right corner is that, like, Chat GPT-4o summary of this, it's there on the video side, you can use it for this too if you want, because that's where my curated data goes, as training material for the future LLMs, that's how it is, yeah Yeah, but that's a fascinating world, earlier you were here as a guest regarding Jonathan Haidt's The Anxious Generation and sort of just about this kids' UX and, well, model development, we could maybe go to the Insider side to think about it because, well, I said like negotiation skills, communication skills, leadership skills, and AI, in a way, delegation skills are important, so I think these, like, these kids of ours have lost that, like, maybe they'll learn that, like AI delegation, but they won't know how to lead, so I think this, like

[**51:36**](https://www.youtube.com/watch?v=p4fuMR-znao&t=3096s) nomad, Nomad world that this becomes, where you and, well, one man and destroyed, so like, will we become some kind of, like, islands that have that AI and then nobody really knows how to lead an organization anymore so let's go inside and reflect on this thought, for example. I've had Mikko Alasaarela as a guest and we've been well, declaring the arrival of general AI. Are you ready? Thanks for visiting. If you've watched or listened this far, please subscribe to the channel and see you there in the Insider area, thank you

---

Cite as: Sami Miettinen, Neuvottelija — AI Smarter Than Humans Within Two Years? | Mikko Alasaarela, https://www.neuvottelija.com/podcast/episodes/482-200iq-tekoaly-ohi-ihmisen-alasaarela-neuvottelija-307/, 2025-02-20. For quotes include episode 307 and timestamp.