It's Just A Model

013 - It's Just A Model - Church of Mode

• Peet Sneekes, Ron Kersic • Season 1 • Episode 13

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 1:04:17

It's Just a Model is a biweekly conversation with Ron and Peet on the interstitium of innovation, tech, design, philosophy and behaviour.

Your hosts are Ron Kersic and Peet Sneekes.

In this podcast, we’ll be talking about the good, the bad and the interesting, all wrapped into an informal, unscripted conversation.

CHAPTERS

  • 00:00 — Intro
  • 01:20 — The Bitter Lessons
  • 14:00 — "The December Event" in Coding
  • 28:35 — Dominant Designs
  • 33:30 — Actors and AI
  • 40:14 — What is Old is New Again
  • 45:45 — Wander/Wonder Machines
  • 52:25 — Fashion!

LINKS

  • The Bitter Lesson — http://www.incompleteideas.net/IncIdeas/BitterLesson.html
  • Deep Blue — https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)
  • Meta classes — https://www.cs.sjsu.edu/~pearce/modules/lectures/uml2/details/index2.htm
  • Roadside Picnic — https://en.wikipedia.org/wiki/Roadside_Picnic
  • Project Glasswing — https://www.anthropic.com/glasswing
  • Zero-Day Vulnerability — https://en.wikipedia.org/wiki/Zero-day_vulnerability
  • John McAfee — https://en.wikipedia.org/wiki/John_McAfee
  • Dominant Design — https://en.wikipedia.org/wiki/Dominant_design
  • FAFO — https://www.youtube.com/watch?v=QJf6iFA2xCY
  • MemPalace — https://millaj.com/mempalace
  • Ben Affleck's InterPositive — https://www.inc.com/leila-sheridan/ben-affleck-just-sold-his-stealth-ai-startup-to-netflix-for-600-million-heres-what-it-actually-does/91315812
  • The Accountant — https://en.wikipedia.org/wiki/The_Accountant_(2016_film)
  • Zettelkasten — https://en.wikipedia.org/wiki/Zettelkasten
  • Agents & Wiki — https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
  • RAG — https://en.wikipedia.org/wiki/Retrieval-augmented_generation
  • Golfen Gate Claude — https://www.anthropic.com/news/golden-gate-claude
  • Wrong on the Internet — https://www.explainxkcd.com/wiki/index.php/386:_Duty_Calls
  • Lidewij Edelkoort — https://en.wikipedia.org/wiki/Lidewij_Edelkoort
  • Devils Wears Prads — https://weareactors.com/wp-content/uploads/2024/08/The-Devil-Wears-Prada-Monologue.pdf
  • An Engine, Not a Camera — https://mitpress.mit.edu/9780262633673/an-engine-not-a-camera/
  • Ana Andjelic — https://www.linkedin.com/in/anaandjelic/

ABOUT US

SUBSCRIBE

SPEAKER_02

Are you ready, Ron? Sort of, yes. Alright, because this is It's Just a Model, a lovely conversation with Ron Kersig and myself, Pacers, and uh we're talking we talk about design, philosophy, technology. Uh, are there E's? I forgot. Probably innovation is a big topic. Uh, we uh haphazardly also stumble upon uh AI sometimes, just because of your uh profession and my fascinations, and uh culture, culture might also be uh a thing. I think that might be uh relevant for this episode. So and you kind of like ah so I I did a big prep, right? I said, Man, we need I need to have something for this episode that kind of sets it off to what to other episodes, and it just hurts you also did the same thing. So I'm really really keen to find out what what your thing is.

SPEAKER_00

Yeah, so I was ununiquely prepared for this broadcast. I have three bullet points, yeah, yeah. It pales into insignificant about what what you collected. No, I really would like to just to spend a few minutes on the uh the bitter lesson of the month.

SPEAKER_02

The bitter lesson of the month.

SPEAKER_00

All right, that's so that sounds lovely. The bitter lesson by now is in um quite a legendary um blog post by Richard Sutton on the state of art in artificial intelligence. I'm sorry, it's d uh AI again. Yeah, it's from 2019. That's really early for AI, yeah, yeah, or at least for the yeah. So he observed the fact that it's the uh you can make uh artificial intelligence systems by encoding knowledge and domain wisdom and domain practices in there. Yes. Right? That that's one way of doing it. You can make a chess computer by making the computer by programming the computer on all things to know about chess.

SPEAKER_01

Yeah, kind of the rules, the rules, etc. And the patterns, I guess.

SPEAKER_00

Yeah, yeah. Right, and then the opening library and all that kind of stuff. So it's the uh and and you talk to chess masters and how do they look at a board, and maybe you can encode that one into the system, and and then you have a kind of system that plays chess.

SPEAKER_02

Yeah.

SPEAKER_00

What you can also do is build a really, really, really, really big computer that just does a very uh uh deep and thorough search of all possible positions, and uh uh so far ahead that no chess master can actually compete. Yeah, uh pick the best move, and so you win from Gary Kasparov.

SPEAKER_02

That was kind of yeah, indeed. That was the IBM IBM Watson uh way to do things.

SPEAKER_00

Deep blue. Oh, deep blue blue.

SPEAKER_02

Exactly.

SPEAKER_00

And that's the bitter lesson that uh in the long run, more compute always turns out to be the better solution. Another example uh that he gave also in this uh in 2019 in in his blog post is that uh uh speech synthesis, so getting a computer to generate speech. And if you now look at the current state of art, but if you go to notebook LM and if you would give it a transcript of this podcast, then you get two people discussing about our podcast. And it's oh yeah, and it's really meta, but it's really meta, meta meta, meta, it's really meta because meta is better. But it's really interesting, and you really have two they sound like two people, and they say um ah, and and they laugh and they joke, and it's it's really convincing.

SPEAKER_02

And they play to each other, but it's a script. I I found out. I I you really can tell that it's uh the banter is like repetitive.

SPEAKER_00

Totally, but but the voice is really convincing. It's it's very nice indeed. And once upon a time, people try to do speech synthesis by actually modeling the vocal tract and the vocal chords, and right? So again, this is modeling the domain or simulating the domain or approaching the domain. You can also just do what you spend a whole bunch of computing power and a whole bunch on training power on speech patterns and just have a system that knows nothing about voice and human voices and just mimic things.

SPEAKER_02

Like a parrot.

SPEAKER_00

In a sense, yeah, right. And and I had my bitter lesson myself when ChatGPT happened, because I spent more than a decade on building uh code generators or autocoders or 4GLs or whatever you want you want to call them, right? And I got them to a really, I like to think, pretty nifty state of art. But these things were really complex software systems, right? And had abstract syntax trees, and it was modeling knowledge, and at four layers of meta models, M0 to M3, right? So you had objects, classes, meta classes, meta meta classes, and with the meta meta class, you can also do a meta meta meta class, so it's yeah, right, it's it's turtles all the way down. Yeah, yeah, so it's fantastic, and then this thing comes along that doesn't know anything about syntax or semantics for that for that matter, yeah, just saw pretty much any piece of code on the World Wide Web, was given a whole lot of GPU power, compute power, yeah, to figure out patterns in there, and then given a lot of compute power to actually generate plausible fragments. Yeah, and it turns out if you train long enough and hard enough, plausible means correct. So now you have this thing that does code to an impressive level, yeah, almost unbelievably good level, to the degree that I see software engineers not software engineering anymore, in a sense that they make code, because the code is better, yeah, it really is, and it doesn't know anything about the programming language. That's the bitter lesson.

SPEAKER_02

So uh, are you saying that your efforts before LLMs were futile? That that you that it was uh it was for nothing, for naught?

SPEAKER_00

No, but I mean I think I mean it was uh hugely enjoyable.

SPEAKER_02

You got smarter.

SPEAKER_00

I got smarter, and it is something I wanted to build and was still fascinated to this day by compilers and these kind of kind of things, right? But, and here's the bitter lesson if you want a better solution, it's just throw more compute power at it. Or wait until Moore's law is advanced far enough that you can throw more compute power at it. And there is something I don't know, sad about this, right? Because it it doesn't know anything about the domain, it's modeling, right? I mean, and people always keep telling me, yeah, of course, generative AI works for programming language because programming languages have to be correct and otherwise they want them. But that's the thing. It's not trained on the formal system. No, it's not trained on a formal grammar, it's it's trained on the expressions of the formal grammar. Yeah, and apparently, if you train it, if you give it enough examples and enough power to do all these training iterations, you get something that has knows jack shit about the grammar, part of my language, and still makes it fit into the grammar. Yeah, you know, and that that is a really picky grammar, get one comma wrong and it explodes.

SPEAKER_02

But it's all it's very it sounds like it's very frustrating to you that you can be bested by sheer compute power. Uh, and even then the result is if you would ask an LLM like, hey, what does this mean? The thing that you just produced, it will say it will say, Oh. It will not know, it doesn't know why it is correct, why it is no, it gives you a mimicry of correctness, yeah.

SPEAKER_00

And that mimicry, and that's the bitter lesson, but sufficient compute power is so convincing that I mean it's almost like fiction uh is better than reality or surpasses reality, right? Oh, yeah. So that's what you have.

SPEAKER_02

Reality surpasses fiction, I guess. Oh, is it? Yeah, I mean our fiction is boundless. Okay, good. I don't know. But uh but I I've so when I listen to you, and I I well I wanted to say. Um, let me reframe what you say by saying it's not about compute power, but it's about the capacity to learn. But basically, you say, no, no, no. There there's nothing, not nothing learned. Anything, no, it's the it only emphasized patterns.

SPEAKER_00

It's the uh it's the I mean so so what it learns is the uh is the the the derivation of the thing you're about to learn. Yeah, right. So you don't learn about programming and programming language and syntax and semantics and patterns and anti-patterns. You learn from the expression of of that, yeah. And with enough compute, it's so accurate that it looks like it actually learned the semantics and things. So it's the it's it's the one level down that is so convincing that you have to believe it also knows the one level up, if that makes sense.

SPEAKER_02

Yeah, yeah. I think it it I mean we always talk about sycophancy of uh specifically of LLMs. So somewhere in the interface, somebody decided it would say, Okay, let's reward you when you say something to me. It's like, oh, that's a smart ID, let's do that, blah, blah, blah. But in fact, it is 100% sycophancy, so it is trying to mold itself towards your approval, right? Until you say this is a working uh program because I can run it through a uh compiler, yeah. It will say, Okay, uh, the road up until then was well a little bit more correct, a little bit more correct, and then we can run it, and then it improves in quality, I guess. Uh but still it seeks or it's not actively seeks, but it is um at the at the end of the line, it will uh oh man, how do you call it? The quality is uh measured by our measures. So do we recognize it as a real language? Do we recognize it as being compilable?

SPEAKER_00

Yeah, and do we recognize it as a certain type of behavior for which we have a term that we normally use in between people? Exactly. I mean, you would you wouldn't use that uh that that word that I'm trying to not use because I can't pronounce it. But we wouldn't use that that word for uh a cat, for instance.

SPEAKER_01

Superfluous, see so super, yeah, but it's not that. No, no, no, okay.

SPEAKER_00

Sycophancy. I mean it's sick of it. If I say it really quick, you don't uh yeah notice that I can't pronounce it. Yeah. Right? We could that's a that's a thing, that that's a phrase we typically reserve between people. Yes. You don't say that about the the parking machine, right, that I have to use uh to actually do my parking here. You don't use it for your espresso machine.

SPEAKER_03

Nope.

SPEAKER_00

Right? I mean, and and now we have a machine and we use that word for because it has been trained to repro to replicate this behavior to such a degree of fidelity that we say, oh, stop it.

SPEAKER_02

Yeah, but even then the fidelity that we say let's say we have music and we express fidelity as being uh music that's very enjoyable but also has high quality, right? We can hear all the things basically. That's that's how we express fidelity. Uh, but it's also it's always us, it's always us as the measure of something being okay. And I find it very interesting. So let's say this was a child. I mean, let's say a child is is something with uh a certain amount of compute, and it learns from mummy and daddy uh how to talk. Yeah, only a child will understand when it says blah blah blah, it will get milk, or when it says something else, it will uh be comforted, etc. But an LLM doesn't know that at all, it doesn't reflect on what it uh gets in return, it doesn't make correlations in okay, so if I tweak it a little bit like this, then people will understand me better, or then I can some can get something done. And that to me is very fascinating. So even with that, all that learning, with all that training, it's still like an amorphous blob for all intents and purposes, it doesn't have consciousness or it doesn't have purpose for all intents and purposes.

SPEAKER_00

No, I mean and and and shoot it.

SPEAKER_02

No, no, no, it's a tool, but but we we we we say it it has all kinds of properties. You know, it's nice, it's helpful, it's friendly, it's wise even.

SPEAKER_00

Yeah, that's what we say. And again, that that's the bad bitter lesson, right? We created that that machine by just throwing a whole bunch of data and compute at it.

SPEAKER_03

Yeah.

SPEAKER_00

There is no no knowledge graph in there, no ontologies, and it's it's nothing. There is no database of all human knowledge, it's it's a bunch bunch of nodes with a value between zero and one.

SPEAKER_02

So let's say this is a stage that we live in. We are now depressed. We're we're past the hype cycle, we're depressed about what we just reached, or we but let's say that we learned about uh what we just the level that we just reached.

SPEAKER_00

I think it's fascinating because then you ask, I mean, what's the next lesson exactly? And yeah, and and and that's why I said the bitter lesson of the month. Before we get to that one, I mean what I found also fascinating is I mean, we had this bitter lesson at the end of last year, November, December. You had uh Gemini 3 Pro, uh uh uh Clot uh opus 4.5 and ChatGPT 5.2 or Codex 5.2. Yeah, these these coding agents that suddenly turned out to be really well. Yes. That you could give them longer running assignments, tasks that they actually completed and did really well.

SPEAKER_03

Yeah, right?

SPEAKER_00

With these all these point versions, right? 4.5, 5.5.2, not six service pack one or something. No, no, no, it's it's just a it's a point release, right? And at first glance, these things look and feel the same. But it's not like when you were having upgrade an app and suddenly you have an extra button, right? And now you can do this. Yes, no new features, no new tricks. It's it's it looks exactly the same, yeah. But it isn't. And it's this amorphous thing. So what really changed? And you just don't know it, but you feel it, right? And suddenly you see, damn, I can actually do much better and different coding tasks. Yeah, and suddenly it's there, and everybody now talks about what happened in December. It's almost like an alien visit, right? I mean, something happened in December.

SPEAKER_02

Yeah.

SPEAKER_00

Something just locked into place, if you will. Yeah, I'm now uh referencing to a book called uh I should not. Link in the show notes, it's it's it's it's one of my favorite science Roadside Picnic. Okay, I don't know that one. Oh, you should read that one. It's about an alien visit without the aliens actually being in the book, it's fantastic. Oh because the the uh the thesis of the book is if aliens ever visit uh Earth, yeah, they will treat us the way we treat ants during our roadside picnic.

SPEAKER_02

Right.

SPEAKER_00

Yeah, that's that's probably true. We leave and the ants still feasts week after we left with our little leftovers. Yeah, that's what the book is about. I love that.

SPEAKER_01

That's interesting.

SPEAKER_00

Uh, how did they get to that one? Um I have no idea how we got to that one. You were you were going to the next lesson. What might be the next lesson? And of course, now um, so in a couple of the the recent weeks, I saw security researchers and professional pen testers uh start noticing ah, these LLMs are getting really good. We don't know really why or how, but they seem to be really good at finding zero-day exploits in our software. Software we already have been working and investigating for the past 20 years. For decades. For decades. And now it seems to find zero-day exploits.

SPEAKER_02

Which it which is like the it isn't the most basic exploit or basic bug or uh uh security breach, but it is uh it should be like the most obvious one, right? It's it's something to go straight into one system and break the system with without any like difficult paths or whatever.

SPEAKER_00

Well, I mean it it could have difficult paths, but I mean, but it's the uh that your your your response time in fixing this is zero days as and now. Yeah, right? I mean, you have to, I mean, so it's it's it's a very fundamental flaw. Exactly. That's something that's the word I was looking for. And that is the um no, but I was really surprised, again, by how surprised these professional security researchers were about the fact that something happened. Yeah, a roadside picnic, right? Something happened. We don't really know why, but we now see that the uh uh uh that the that the exploits being handed in by these tools are getting really, really, really good. Half a year ago they were still very naive and uh yeah, well, a bit meh. And something happened. Yeah. Right? A roadside picnic. Something happened, and of course, now we have uh mythos.

SPEAKER_02

Oh right. Yeah, I I've I've uh I've actually already already written a blog on mythos. Because to me it's like the so I read the entire takeaway. There's this analysis of uh the people at Anthropic, what mythos why mythos is what it is, but also they they're measuring things, yeah. And it seems especially that there there to me there were two takeaways from their brief or or preview of of the brief about the preview of mythos because it's it's not out in the open yet, because they've they dubbed it um uh a security risk, like a uh or at least a possible security risk. So the first takeaway from me was a thing like okay, you have a certain the roadside uh picnic that happened, and all of a sudden there's more execution execution power, it seems like yeah, it is more capable into doing things, and they found out their imprint that they have for each of every model of uh uh of the entropic models is not sufficient to uh counter all that power or to get to have it's like Spider-Man, right? Uh with great power comes great responsibility, they had more power than before, but they they didn't adjust the amount of responsibility a model would take. It's like couching the security uh blanket, uh if you will. And um that was their biggest insight. It's like, okay, we thought with this imprint that we give to all our LLMs that it would just be more uh be uh more responsible with all the new power it's it's gotten, and the conclusion was no, it's basically unaware of the amount of power it has and the consequences of the things that it does. And to me, that's like wow, that's really weird, right? Because it's it's not an animate object, it it's it's not a living being. How does that work? Um, and that that was my that was the first conclusion that I read about, and the other one was so why what is why do we see this roadside picnic effect? And their conclusion seems to be that rather than um so if you are a security expert, you're expert on most most of the time, you're expert on a very specific domain, let's say a piece of program that is very risk, uh risk uh sensitive, or you are very broad, like a generalist, and you know a little bit about something, right? And as a human, there's this um Venn diagram that you can't overlap all the time, right? An expert knows way more about a specific topic than a generalist would be would ever be, and now we're at a state that this LLM model knows way more than a generalist about everything, and it knows enough to even compete with an expert. And that combination makes it very dangerous because all of a sudden it can't just say okay, these are the specific security targets in this specific piece of program, let's say uh a web server or an operating system, but it knows all those things about all the other uh layers of applications and servers, etc. And by that knowledge, it also can make a horizontal route, right? It can say, I can go through the OS, go through the server software and uh creep into the browser software at the client side and that is a new vector of a security breach and that kind of seems to be like the capacity that makes the the this particular LLM very very dangerous.

SPEAKER_00

Yeah and but that's the best bitter lesson it doesn't know jack jack shit.

SPEAKER_02

No no no it it does it does yes right and that's the uh and apparently anthropic learned that uh the hard way yeah right so they found out yeah I mean but I have to say it like like this because this is a very also I mean it has a very strong whiff of marketing right it is yeah we have this thing oh it's so powerful don't think we can give it to people yeah yeah yeah okay if you would give us two thousand dollars a month what but no no no no I mean it's it's the McAfee uh power play right McAfee the the the the virus uh person the the person that created the first virus scanner basically said well viruses are the new security factor of um um uh for computers so way somewhere in uh the 19 uh and there two things happened one people were looking for uh a way to mitigate that and surprise surprise he had the way to mitigate it with anti-for uh anti software uh anti-virus software but he also pushed like the notion that viruses are a very interesting and lucrative way to um yeah inflict danger or inflict risk into companies so all of a sudden a lot of new virus makers also yeah got their uh I don't know incentive I guess yeah but what I mean so but I listened to the to this podcast and it but was a bit too too hypey and to and and too brief before seeing but a guy from Anthropic said that I mean this the the the new Mythos model yeah great name by the way yeah very mysterious was running internally and a researcher went uh back uh went to lunch and when it came back it had escaped the sandbox oh yeah yeah it escaped right and it's the uh but I think this is one of those bitter lessons things right I mean you have something that is that's upfront you wouldn't know that it would be very good at XY or or or Z but it turns out it is yeah right and that's this this moment again so somehow these models in general and probably the next generation model of anthropic uh why wouldn't it too yeah has this new capability that got all the security researchers that they have their panties in a in in a twist now. Do you think it's like a 3M uh uh like a sticky note moment right they try to make glue to stick two services together and haphazardly they make it like the sticky note glue?

SPEAKER_00

I mean apparently what I read so far and I again I'm not working there but they didn't have the intention to make a model that's very good at pen testing.

SPEAKER_03

Yeah.

SPEAKER_00

But it turns out it is yeah and I'm for sure they they have done something in that direction but not that specific yeah I'm sure that the uh the uh the the November models were trained to be good at coding but I think they were a bit surprised how good they turned out to be at coding.

SPEAKER_02

Do you think do you think they they might like what I just said they might say okay rather than going the single domain route because they kind of like had this capacity problem that they couldn't fit all the domains into one model so they make a router and basically decided okay you're asking about literature let's go to that specific domain. And now they say you know what we're colliding all those domains again together and all of a sudden we see this new power.

SPEAKER_00

Yeah I think I mean the a general research direction is dealing with with large contexts in in an in an in an effective manner. Yeah efficient and effective manner. Yeah right and this seems to be one of those context things because you have the operating system and you have the web server and you have the uh the web browser and how do you correlate all these things so apparently I mean that would make sense that that is a research direction that one would follow and then you get the bitter lesson that it actually does things that really surprises seasoned security researchers. Yeah all right I mean if you look at the the paper that Anthropic published I mean it has Bruce Schneier on uh his name on there I mean he's not your average Lorem Ibsen generator. Yeah generator I mean that's the I mean if somebody has a reputation and deserves one it's him yeah right so uh this is real but this this again is the bitter lesson again.

SPEAKER_02

But do you think he also is kind of surprised about so it's more the surprise that affected his uh I mean I'm I'm almost saying this in a in a good riddance manner right I think they have the same I mean I'm not talking to them but I can assume they have the same surprise like I had when ChatGPT turned out to be a really good coder. Yeah exactly better than my coding systems yeah right and I thought damn how does the stupid thing because it is stupid do this yeah right and and again that's the bitter lesson and the bitter lesson is real which also is sobering and etc I mean and you can be bitter about the bitter lesson yeah but I think this is what we're seeing just enough compute and there's a bunch of compute behind this and then you get this and then you get this who knew yeah exactly and there it is yeah so so there and the bitter lesson would also be let's just put in more compute even more and let's see where we get the next aha a labeling.

SPEAKER_00

So yeah I'm really looking forward to the next bitter lesson yeah I mean what will that be?

SPEAKER_02

Yeah and even if it's bitter I mean you're surprised that it might a certain LM is great at coding the the only reason that it was bitter that it's basically nulled your all your efforts before yeah I mean it's I mean it's it's it's it's so to a lot of uh the new developers it might be like the perfect tool.

SPEAKER_00

Of course of course that's what it is I mean I mean actually no thinking what was I thinking right I mean it's the uh but but again um um here here it is yeah at some at some point the dominant design forms and this is what you get I mean if I'm into cars and also now in in into old cars. Okay and now we in really really really old cars as as the early ones right and you see like the 1900s?

SPEAKER_02

Yeah yeah I mean so I mean so you had accelerator pedals that you need to pull instead of push yeah so not down but up to actually accelerate right I mean so yeah yeah it's it's the I mean brakes that I mean so many attempts at getting the thing done yeah and of course now we uh uh converged on a dominant design right this is an accelerator pedal right down is faster or more whatever you want to call it right I mean it's the uh with with with brakes the same we even have a certain order for for brakes and accelerate I mean it's the and even electric cars adhere to that that principle although they they are also changing that same thing I I so I watch this it's a really weird segue but I watch these videos about a uh a man that uh helps out with people stranded at the side of the road with broken cars so it basically picks them up and puts them at in a safe safe place and so I I see a lot of cars I I don't have a driver's like a driver's license so I don't see that much that many uh cars from the inside and I saw like the recent generations of drive uh cars have a button for the uh the uh the the how do you the the brake the the system brake yeah you know there's like the the use of it be a lever it's like yeah and then the physically the uh the yeah the wheels will be uh braked and but now it's like a button like yeah so he has to switch on the the car yeah and then say oh no no buddy uh you're still uh Bluetooth with your car so could you please disable that and then he could push the button and to me it's like no no so I guess these patterns are also broken right it's as soon as we find a useless I mean we also lost the pedal uh we lost I mean yeah yeah from uh from uh shifting shifting to uh automatic yeah and of course we had automatic gearboxes for quite some time yeah exactly yeah yeah and but with electric we don't have any shifting at all no yeah no so we don't need that okay some cars have but they do it in the in invisibly yeah indeed yeah Porsche has a the a car that has a two a two speed gearbox okay that's for their power uh because they can probably probably while we as kunnen I I just can hear them say that as an associate but that it to so that's one of my core philosophies specifically with unlearning people that that particularly particular remark because some people say why would you do that because because I can yes why would we go to the moon I think I guess that was the first thing uh I uh heard that remark is is because it's because it's difficult yeah because we think we can do it yeah and why why wouldn't we? Yeah it's it's called the moon shot for nothing right I mean it's the uh not for nothing yeah but but even with the small stuff it's it's it's a learning opportunity waiting and I guess that appeals to us.

SPEAKER_00

I mean famous last words right what does this button do?

SPEAKER_02

Yeah yeah yeah yeah fantastic yeah and we found out and and it cost us dearly but we found out and that a lot of great moments in history are are bound to that yeah what if we would not break like this certain uh conviction what if we would not have this cultural change throughout our entire society and just gave it like all the the energy and all the power yeah what would happen well we found out we found out and still finding out yeah bitter lessons yeah what was it it's uh it's also something you would say to kids it's like uh fuck around and find out yeah yeah and to me is it it kind of has a negative connotation I know but also that's exactly what we do there's this great meme on there on of course on on the internet with a with a guy that actually has this little graph one says fuck fuck around and one says find out and he has one uh curve drawn here and he says but the more you're gonna fuck around see the more you're gonna find out brilliant yeah it's a scientific it's a scientific fact yeah okay let's uh let's go uh from from this uh well I I wouldn't say bitter but let's say it's it's definitely a lesson each and every time and certainly we aren't finished learning from this interesting behavior of LLMs uh to uh something I saw and it seems to be a trend I I'm not sure but it seems to be actors or people in the film world uh um trying to do things with AI yeah I mean of course it's very new and uh the first reaction to AI was in particular from filmmakers and actors to hey this is this encroaches to our business I might be uh uh like yeah replaced by a virtual actor I guess that was the that was the biggest threat or to the rightest we might be replaced by LLMs yeah and so there are two actors that basically outed themselves in as in okay we're entrepreneurs in the AI space and the one is I need to read this because I don't know how to pronounce her word but it's uh multipass it's uh Mila Jovovich which were mem palace yeah and so she caught on with a very old model it's the memory palace it's a way to remember things it's very it's it's like old as in the Greek time the Roman time somewhere around that and apparently she tried to she stumbled upon it and with the help of a very uh famous uh LLM or AI expert she she built it yeah and it's really her right it's really her yeah but I guess it's like she's she tumbled into a great idea or an interesting idea and the application seems to be a little bit broader so the bottom line is what if you would cross this mind palace principle with LLM efficiency uh the the efficiency efficient use of memory specifically yeah and uh it seems to in her in their tests it seemed to compress uh memory a little bit yeah more efficient than others yeah and uh but they're also counter uh things but she opened it up so it's somewhere in the GitHub and it's very interesting. So what and after reading like uh the benchmarks aren't that great but they're interesting at the least uh I also found out that Ben Affleck uh of course made another interesting leap he had a startup so basically Ben Affleck is a very broad filmmaker right he uh started out as a as a writer then haphazardly because well money uh he also became a director and an act and an actor at the same time and uh now he also specifically with the AI he also started to think about how can we use AI as a helpful tool that isn't a threat to the creative business. Yeah and to me that's that notion is interesting because he seems to like try to skirt that particular piece of yeah how would you call it Morris into their business right and so how can we make sure people have uh work but aren't busy with busy work and uh so he made things like how can we organize our work yeah uh automatically how can we make sure uh if we failed a certain shot or we don't have budget to reshoot how can we correct the things that we made yeah like the malleability of the art yeah without encroaching on creativity yeah and creative expression and that to me is like well you probably made he made the tool if this is going to be bought and it had it's now bought by Netflix who doesn't have any interest in creativity or creative moris I would say I think the first thing I would do as as Netflix is break all the safeguards and say okay now we have now we can produce create creative works without any boundaries uh with very small budgets with very few people great so you can have more well because Netflix needs more makes better what more yeah exactly more more more yeah so nothing is left about the attention of uh Mr Affleck I have no idea can I say that the accountant is the best movie ever the accountant is very very nice. It's the uh the final scene oh yeah just thinking about it makes me cry again every time I know why yeah I it's so good when you actually learn who showboat is oh yeah yeah yeah but to to it it kind of like uh it uh also brings an insight into people that might be a little bit on the spectrum or a little bit much yeah and to me that to me was an a very interesting thing to it it also is very in the mode uh of today right it's very popular to do that yeah uh but to me it was yeah quite so he basically it seems to say being on the spectrum is not a limitation but it is uh by all intents and purposes a s very specific superpower and it's uh uh if you are smart enough like uh Ben Affleck plays in his uh in this uh film and is in real life I have a feeling and probably also is in real life um it is very it is uh uh a superpower yeah yeah and and and this superpower in this case is very like a very structured life and very structured way of looking a very rational look way of looking at things yeah which is very stereotypical and this this is not for all people in uh on the spectrum there is also just blind chaos and of course emotion of course uh at the way other at the other end yeah but uh but nevertheless a very interesting and very nice uh video so please please watch this yeah please watch this video no but I think the the especially the Ben Effle case is really interesting.

SPEAKER_00

I mean Mila's case is uh I mean but this is also interesting one because one thing I notice is that old stuff yeah becomes relevant again with large language models I think we talked about this I mean and now she comes up with a with a a memory trick process system.

SPEAKER_02

Yeah it's a system yeah that um uh uh uh suddenly becomes very applicable again yeah with these things that that that that that mimic human cognition yes at scale yeah right so it's the uh you see a lot of things I mean uh the same thing with the uh the settle custom um the second brain thing okay okay okay so they they so I have to explain this one yes link in show notes but it's the uh no but it's uh it was a German researcher uh he wrote a ton of books yeah and um every idea every every observation every well every reference he wrote down on a little card and he put that in a big cupboard yes small card is called settl in Germany the cupboard is castin so settl cast basically hyperlinks and flash cards flash cards and and every card refer to another card it's it's that simple really I mean that's nice I like that concept yeah it's the end and and I've been using it for the for almost a decade and a half as in Markdown exactly which I didn't know would be a a medium for LLMs I mean that uh so we talked about it in the previous episode or we kind of began that topic and I found this is a thing it's people writing all the things in Markdown because it's just like this very generic uh uh format yeah uh but also it's structured in some way so it can also be a little bit recognized what's important what's less important yeah and it is fodder for an LLM you can make your own LLM based on all your own work.

SPEAKER_00

Yeah you can make your own agent based based on your own work and that's what I have been doing with my second brain or settle custom yeah and uh what was it was it last week two weeks ago with Andre Capati there he is again yes he had an approach where uh his agent would collect all his little set settles and uh consolidated that one in a wiki I thought yeah that's a nice one intermediate stage never thought about that one and well what's the what's the benefit of a wiki it's like the cross references yeah there's a cross reference and then you can have other agent agents reason about the wiki yeah right basically what we just discussed right operating system web server web browser consolidate that one in one world view if you want one wiki yeah one model yeah what one model and then have all the other agents reason and contribute to and on the wiki yeah so that's what I'm now running as well on my separate Mac Mini. Of course everybody has a separate Mac Mini nowadays right but it's but this is an old idea wikis are old Tetlecast is I think it's 1950s or something link and show notes and probably will be uh will be earlier. So you see all these old techniques and that people use to get up to get by become relevant again. Yeah so when I saw the Mila thing I said yeah of course not surprised what's also interesting and and perhaps even that the LLM is way better to do that right it's it has more capacity to and this is another one I mean what you all there's also a bitter lesson in itself a smaller one is that whenever you have these complicated setups like like like like like RAG right retrieval augmented generation right where you chunk your documents into vectors and store them in a vector database yeah compared to just have a structured repository of markdown files. Yeah you mean the the work the work already has been done well I mean yeah I mean the that and and the work can be done in in in runtime right and and making runtime and making the wiki is a form of RAG. Yes. But just an inspectable one. Exactly right instead I guess instead of a binary one. Yeah and one of the reasons I think that the World Wide Became the World Wide Web is because HTML is text and not binary.

SPEAKER_02

Yeah, and it's also understandable by machines and humans.

SPEAKER_00

That's it. That's that's what I mean. Yeah. I can go into the wiki and read the wiki.

SPEAKER_02

Yeah.

SPEAKER_00

And you can correct it even.

SPEAKER_02

I can correct it. Or enrich it.

SPEAKER_00

I mean, yeah, whatever, right? But and I cannot go into a in a vector database and see the vectors.

SPEAKER_02

Right.

SPEAKER_00

I mean, they binary things. Yeah. And even if I decoded them into readable things, it would be a vector, so a list of thousand twenty-four numbers between zero and one. I mean, I'm weird, but I'm not that weird. That I can read that stuff.

SPEAKER_02

Well, I guess if you if you are a wiki editor, then you might be that weird. Then you would like, for example, make as much cross-references as you should just because it's and uh But again, it's I can read the wiki, I can see the wiki, and I can edit the wiki.

SPEAKER_00

Yes. I cannot edit a vector database by just opening in a binary hex editor and just uh uh uh edit the things.

SPEAKER_02

So it's like one of the best mutual formats that we can find between humans and LLMs.

SPEAKER_00

So I mean, I mean, and and every time we have a complicated setup and we replace it with something more simpler, yeah, either it's better or it will be better with the next generation of LLMs. Yes. Okay, I mean so RAG was a hack for the old ones because it had very small context windows, yeah. And now kind of in this camp. Again, a bit a lesson in the in the small.

SPEAKER_02

Yeah, to me that's uh very uh so one of my I have all I have like this bucket of ideas I really want to make, and one of the ideas is uh what I call uh the wonder machine. And the wonder machine is based on what we like to do is wandering around and oh, that's interesting. It's like being the ultimate distracted person. Every time you see something shiny, you go to it and you click on it, and you go into this new shiny world.

SPEAKER_00

So it's wandering with an A.

SPEAKER_02

Exactly. Oh, fantastic! Yeah, or wandering as in being very uh for a time.

SPEAKER_00

That was my tagline on LinkedIn wandering and wandering, yeah, yeah. And an O and an A.

SPEAKER_02

Yeah.

SPEAKER_00

And people ask me, What do you mean?

SPEAKER_02

So in Dutch is also very nice, it's dwaalmaschine. Dwaalmaschine, and I I'm pretty sure if we do it in German, it might be even better. But uh, so the but the point is making as much references between things, so in the and the start is an art database, yeah. Making tupels of art let's say uh paintings and uh uh sculptures and etc., buildings, even yeah, and connect them through time, medium, uh artist, etc. And to me, uh I wasn't satisfied with that limited uh these limited connections because I saw this one um uh television program, it's called Bilden Storm in Dutch. So it's uh statue storm. It's it's it's one of the things that happened in our history when uh Catholic Catholicism was uh not accepted anymore, and all of a sudden, specifically like the very it's very rich. There are all kinds of statues in the churches, it's very uh uh opulent. And uh the Lutheran people said this is not what we want, as they started breaking all these buildings but also the statues, and that's Bildenstorm. So this and this program talks about okay, so this image or this painting has been made in this time, and that's why it looks why it looks. So, for example, uh, expressionism is a typical reaction to um the liberation from the Second World War or the First World War, I'm not sure. So, I want to make that link. I want to make the link like this particular thing happened or is about to happen in history, and that's why this art exists. Yeah, and to me, that's like as similar to like uh book printing. Book printing can only happen when uh or uh can uh when it happened, it also started like education or on a really really big scale.

SPEAKER_00

Yeah, all these second order effects or the higher order effects.

SPEAKER_02

Yeah, exactly. Yeah, that's interest and I like those interconnected things because I'm pretty sure at the at the end of the line all things are connected that way.

SPEAKER_00

Yeah, yeah. I mean, I I just made an argument against uh uh vector databases, but one of the cool things about them is if you embed things, text, in a vector database, it's the uh they you can um you can you can link them on proximity on based on meaning. That's which also means you can take two things that are not proximated in meaning, yeah, and still try to make a convincing argument. Yeah, right. I mean it is also I think it was also anthropic that they actually modeled it. It's a bit like like like like like Photoshopping. They were actually um manipulating their internal structure of their model to make certain connections that normally would not uh happen.

SPEAKER_03

Right.

SPEAKER_00

So for instance, they manipulate the model in such a way that whatever you asked the model, it would mention the golden gate bridge. Ah. And then it's amazing what kind of uh uh uh leaps of faith you actually get in the in the response just to include the golden gate bridge, yeah, yeah, yeah. Which is really, really interesting. I really want that as a feature, right? I mean, whatever you do, mention this. Whatever you do, don't mention this. And then, of course, I'm gonna ask you that actually makes you mentioning this. So don't I mean don't try to mention the word, John Click, right?

SPEAKER_02

I mean, don't think about uh pink elephant.

SPEAKER_00

And I think, and again, this is the fascinating bits of this technology, yeah, right? I mean, that that that that you have this correlation on meaning or uncorrelation on on meaning and just force it to, yeah, right?

SPEAKER_02

Whatever it is. So we basically need to make uh make an LLM or the machine that makes an LLM make it hypersensitive, basically make it uh a little bit allergic to a certain thing, yeah, so it would react. Yeah, uh reacts abundantly, yeah, yeah.

SPEAKER_00

Yeah. And also when it doesn't need to beyusible in regard with it with us training. Yeah, what we would say is real or this is a sufficient approximation of reality. Awesome, right? No, no, no. I mean, what, how did you get the golden grape agent there? Yeah, I mean how? But it's it's it's interesting I mean, and again, this is the fun bit about this technology, right? It's it's not a search engine, it's not an answering machine, it's a wondering machine. Yeah, exactly. It's literally literally about wondering and wandering with an A and an O. Yeah. Just wonder and then see things and say, ah, yeah. We literally said a unison up. It's the uh but that's the fun thing. And I oh I mean so also yesterday. I mean, somebody complaining again on about LinkedIn, how stupid LLMs are. And if you ask it an an answer, uh a question, you get an you get a generic answer. Yeah, well, don't ask it generic questions, bozo. Ask it silly things, you know. Don't ask it, ask it wander with me with an A. Yeah. And let's see what wonderment you can serve up to me. Yeah, without knowing up front, it's not a bloody search engine. Yeah, I know somebody's wrong on the internet, right? I mean, but it's I typically what this is which is a meme. Link in the show notes.

SPEAKER_02

Well, what is it? It's like this image of a of a of a dark uh silhouette punching on keyboards and his wife in uh in a more closer proximity. Honey, are you coming to bed? And the person says, No, honey, somebody's wrong on the internet, and starts stamping on this keyboard. Which is brilliant, I think. It's yeah. So, Ron, I I have this proposition for you, and I think we should do it because you are now a lecturer in uh in the space of uh AI, and one, I think it might be uh I think it it's uh probably time to broaden that horizon already. And um so do you know Liebei um uh oh Adelkort?

SPEAKER_00

I do not. You do not? I have a feeling that in a short period of time I will say, of course, I do.

SPEAKER_02

Okay, so when I was young, I worked at a company uh it's called um uh Mediumatic. It's very well renowned in uh in Dutch uh internet culture because we saw we were at the front line of the internet. We invented the internet. No, we invented the things on the internet while the internet was built underneath our feet, and to me, it was very wild. And this company has this interesting Venn diagram of technology, design, and art, yeah, and which also implies that you start thinking about this this very interesting intersection. And uh when I started moving in the space of art, I found out that a lot of artists were thinking about technology, possible impact, and how it might steer our um our society. And there there's this one person that came from the direction of fashion, and she knew a lot which is uh Ledewey. And Lidaway knows a lot about fashion, about colours, about all kinds of things. But the way she predicts, and she made like a yearly or a bi-yearly prediction on how the future uh might look like, and she did this for fashion, so the results were textures, uh, clothing, would it be tightly fit, would it be very broad, would it be very texturized, or would it be very smooth, stuff like also what kind of colors? And um she does this like regularly, but when she does certain predictions, she also says, Okay, let's put this in the context why this is, why we, for example, why we uh envelop ourselves in bland colors uh these days, right? The topic of a previous podcast. And she says, Well, this is because of this, it is, it is. Uh, we do we were looking for uh like homogeny, for example, or like all kinds of psychological and societal things that make uh uh shape our future. And I think this is wonderful, and I think this is not happening enough in our world because we have like the McKinsey's might be interesting. McKinsey has like this business overlap of with technology, it kind of uh gleans into what technology might uh help business in the in the near future. We have frog design that does similar things with uh the digital space, digital design and uh societal impact, but also vice versa. What society might need uh from uh uh digital design and digital designers. But there isn't something that is as broad as leaderway education. Like the bridge is to me, from my perspective, is so large going from fashion uh uh to societal impact and vice versa. To me, it's like a huge bridge. And I think it's wonderful. I think it should exist in particular for our current uh world. And I think next to Leaderway, there is like a huge gap of people not doing that, a huge gap of uh technological companies and uh advisory companies and management companies thinking just about the things just in front of their noses in rather than uh uh possible societal impact that it might have or the way we're thinking about things, uh, but also the other way around. It's like, hey, we shifted our attention. Is this good? Yeah, should we change it? And I think this opportunity lies before us as uh people that are a little bit distracted about more things than just the technology, yeah. Right? We we think about how and not per se, so there's also this one company. Uh the the name I I think it's it's on here somewhere. I have this short list of uh names. Uh Deloitte's uh IDC, for example, and Gartner. Yeah, they they have this interesting way of um gatekeeping. It's like everything needs to be sustainable, or everything needs to be like a certain direction, uh, that's more a fashion rather than uh than a uh uh so it's more triggered by hey, this seems to be like the the way we are thinking at the moment, the things that we our society finds interesting, rather than predicting what the next interesting thing might be. And I think this is something that needs to be done, and I think it's it's not done at this moment. I'm not sure. I mean, you you frequently uh quote books that you read and things that you see. Are there like inspiration from outside the space of technology that you've that you think like, hey, this might be an interesting way to guide us through this uh or broaden our horizon, uh change, change our perspective?

SPEAKER_00

Yeah, I thought it's the uh it is an interesting one. It's the um my answer would be no. Yeah, I think for instance, if you look at most futurists, yes, I think it's always always very very bland and it's the uh and and and uh very unexpected, very predictable, even for the unpredictable stuff. It's yeah, no, it it does it it it almost what I always miss is the groundedness of things, right? It's about the opinion and selling as certainty.

SPEAKER_02

Yeah, right?

SPEAKER_00

And I think the fashion bit is a bit more interesting. Well, a bit more it it's it's it's uh it's very volatile. It can be. Yeah, it isn't at the moment. Yeah, I mean, and that of course is but what because I'm I'm even though you s you don't see it the way I dress, but I'm really fascinated by fashion, right? I mean it's the uh like that famous quote in the Devil Weas Prada about the colour blue. You think this is blue, right? And then this whole tirat about it but as a system, right? I mean, fashion is really interesting, and and and the expression of where we are as as people and society, and and and and why do young people dress so incredibly bland and boring, and you should see how I was dressing when I was your age. But okay. Yeah, it's really yeah, but uh but this is a really nice one to explore, actually. I think the uh and it never dawned on me that the uh um the whole fashion because I'm always see that there we go again. Um always surprised how a whole industry moves without I think upfront collusion. Yeah, there's no intent. There is no intent, but still it moves like a system, right? It moves like an organism, and they are always fringes, right? But I mean, if I see the collection from this uh uh designer, and I can see the trend, right? I mean, so but these trends are they're not sitting in a room and say, you know what, the trend for next week is it's gonna be blue, right? Or something like that. No, no, and so it emerges, but from a collective, yeah. And uh what I always don't always what I don't like typically about the tech predictions, like Gardner, is it doesn't emerge from something bigger, right? It's just some someone here when it says this is gonna be this. Exactly. And if you look at the hype cycle, it's gonna be this, right? And all the certainty. Yeah, and the market for for blockchains are gonna be four trillion euros in 2025.

SPEAKER_02

Yeah, it's it's a self-fulfilling prophecy starting. The start of a self-fulfilling prophecy. There is a great book.

SPEAKER_00

Link in show notes, and then I'm gonna stop. Another great book. Yeah, it's called uh A Camera, not an engine. Yes. Or an engine, not a camera, one of those things. It's a book about economic theories. Yeah, and the guy writes about uh we think about theories as cameras, they take a snapshot of things that they're right. Actually, they become self-fulfilling prophecies. Yes, so because people start believing and acting on a certain theory, it actually becomes an engine because it comes true. And with all these hype cycle things, right? It because with self-fulfilling, oh Gardner says XYZ, and we're gonna buy XYZ. And so this is the the wrong order, and I have a feeling that with the whole fashion bit, the order is reversed. Yes, and it's the and I'm rambling now a bit. No, no, no.

SPEAKER_02

This is exactly what I intended to go because I I think so. I think this is uh worthwhile exploring, yeah. Making connections that are um that can guide us rather than that can dictate futures.

SPEAKER_00

There's also a I'm not not really gonna stop this. Uh um she has a great uh Substack paid for on us, a great book called The Business of Aspiration. It's Anya and then a uh Eastern European difficult name for me, I should not pronounce. Yeah, but she also and the reason why I subscribed to her substack is also in the is in the fashion business and in the aspiration business and in the status business, and the ability to take this whole system and articulate it in what f to a large degree of lightness could be next, or might be next, or should be next, or might be interesting, or might be interesting, but in a in a much more grounded way, instead of having an opinion and just articulate it like I mean it it is just like these these fortune tellers, right? Oh, I can see I can see oh Ben Affleck is gonna marry uh J Lo again. Yeah, okay. Is that a prediction? That's that's a blot on his copy book. I'm sorry.

SPEAKER_02

Uh oh yeah, yeah. Probably uh doesn't need to do that. No, please, please don't do they have children? I'm not sure.

SPEAKER_00

But I saw a great uh uh clip of his where he bought a Yamaho R7 that's a motorcycle. Oh yeah, so he also has this motorcycle thing. So uh and I'm so happy that Ben Affleck doesn't sell tequila like the other ones. Oh, right. In the next episode, we should talk a bit more about this Ben Affleck uh uh uh thing, because I think it's also a great pattern in how to deal with AI. Okay, but it does two jobs, but that's for episode 14.

SPEAKER_02

14. That's uh that's a great uh I mean if if there's a cliffhanger, this is it this is it. Isn't it?

SPEAKER_00

Yeah, we're the king of cliffhangers.

SPEAKER_02

So uh please meet us at uh episode 14.

SPEAKER_00

When Dr.

SPEAKER_02

Bob says oh I don't know any of this reference. Anyway, so uh see you all in uh episode 14. This was uh It's just a model 13. Uh if you're curious about what we actually talked about and was went a little bit too fast for you, go to our website, it's just the model.com, and you'll find show notes uh made by Ron. And um uh that's uh very informative. Uh it might um uh uh uh give you food to wonder or wonder about, depending on which level.

SPEAKER_00

It's more informative than the podcast itself.

SPEAKER_02

Well, I wouldn't say that. This is my and let's say this is the entertainment combined with the informative. Oh no, no, no, don't be don't put me in the box. Break out of the box, anyway. So, Ron, thank you so much. Thank you, Pete, and uh we'll see you next time. Ciao, guys.

Head Shot

Peet Sneekes

Host
Head Shot

Ron Kersic

Host