Two's Complement - The Tinfoil House

Episode Date: August 15, 2026

Ben's friend is locked out of their AWS account and Matt's AI has decided he's a hacker. The escape route runs through a cheap computer, a mesh network, and Neal Stephenson.

I'm Matt Godbolt. And I'm Ben Rady. And this is Toos Compliment, a programming podcast. Hey, Ben. Hey, Matt. It has been almost a month to the day that you and I have sat in front of a microphone and tried to talk. And at least I'm sat in front of a microphone now and not actually recording on my earbuds in my wife's bedroom, childhood bedroom. So back in the right continent.

Although I do have a funny story on that, I set myself up for a work recording the other day. and for an internal like chat kind of video casting. And I had my pompous microphone in shot, obviously, because I'm one of those kind of podcasty people. And of course, I forgot to actually select it in the drop-down. So it was using my terrible earbuds as the recording. And so that's going to go out internally, and I'm going to have the absolute nicky taken out of me

by being this, you know, hey, I thought we got somebody who was like media savvy, but no, no, I'm not. But that's not what we're going to talk about. today. We don't even really know what we're going to talk about today, but something happened to be this morning. It was the first time that was fun. I was intrigued by a super cool new exploit that has been discovered. Exploits strong. It's like the, do you know, I can't remember the dude's name, unfortunately. Zaw, EAX, EAX, EAX is their handle, his GitHub handle. He did sand sifter, which was like a comprehensive search. through the entire space of all possible x86 op codes looking for undocumented op codes, which was fun. He did mob for skater, which was, hey, every instruction, a mov instruction is Turing complete on its own with certain caveats.

And he wrote a compiler that compiles to just MOV instructions, which is fun. And then most recently, he's discovered this cool facility there. If you have ring zero access, like, you know, if you're like the lowest level access, you might just think, hey, yeah. I've got root. What more do I need to know? There are layers below that that like the firmware of the system runs at and other things that you're not supposed to be able to see about your own computer that you own under your desk. And he's found a way to get into it by misconfiguring the RAM system in some particular way in a really cunning way. Anyway, I was like, this sounds cool.

I'm going to, while I'm in the shower, I gave a quick query to Clause saying, can you just summarize this for me? I want to learn about it a little bit. Give me some pointers. I had my shower and I came back out and I looked and for the very first time I had a massive alert saying, no, you can't do this. You are clearly some evil hacker. We will not allow you to use our AI to hack things. You're a bad person. I don't think that's exactly what it said. I'm sorry if I'm anyone from Apple because listening, but it was it was something like that. And it gave me the option to use, you know, like a watered down model, which I did. And it was perfectly good for summarizing a document, which I then went off and read anyway. So, you know, it was fine. But it made me think,

are we entering a new age where we're being gate kept by things that we don't have any control about? Mm-hmm. Mm-hmm. Yeah. You know, this is, it's sort of like a new problem and also not a new problem. Right. You know, I had a similar experience this week where a friend of mine was locked out of their AWS account for reasons that are kind of opaque to them. That is a nightmare situation for me, because.

Because, yeah, that goes down. And then you're like, well, how do I, who do I call? Right, right, right. And I mean, we have this sort of digital dependence problem in many facets. Like, I'm going to give you a hypothetical scenario here. Your choices are you can continue to own and use a mobile device, a mobile phone, or not. But if you continue to use a mobile device, then you need to cut off one of your fingers. It can be any finger, whichever when you want, you can use whatever method

you want to cut it off, but you need to cut off a finger. Right. Okay. What do you choose? You lose the left pinky. Like, that's the choice that 99% of people are making. It's like, yeah, I kind of need a phone. Yeah. I don't like the way this analogy is going, but I understand wait, I think I understand your point. Yes, I get it. Yeah, okay. Yeah, it's, it's this dependence. And it's like, like, it does, it's not even good enough to have the device. You need, like, the network, right? You need the connectivity. You know, teleport me back to 1970 with the top of the line phone.

And it's useless. It's an interesting, yeah, device, but it doesn't really, yeah, it's not one. Yeah. So we have this, we have this digital dependency, right? And it comes in many forms. Like, AI models are the most recent form of it, you know, cloud services. Right. All of these, like, you know, your Google account, right with your email and everything else.

You've got probably a lot of identity tied up in that. Yeah. I have a story about that, right? My vanity domain, godbolt.org, is a Google apps for your domain when that was a thing. That was like a thing that if you were a Google fanboy like I was, and I was an early adopter of everything, you could say, hey, I would like to bring my own domain to a Gmail account, effectively. Subsequently, when I worked at Google, I discovered that is not how it works. There's some massive internal hack to make that work.

And then it sort of became Google Apps Enterprising type thing. But unfortunately, those who, like early adopted, like I do, are stuck in this hinterland between it's not really an enterprise account, and it's also not a Gmail account. So I can't do half the things in my house that are internet connected with my Godbolt org account. And so I want to just move. And I can't just move because so many things are tied into my login identity

as Matt at godbolt.org that goes through like the Google login process, that now I can't, I'm stuck. So I'm like, ah, what do I do? But yeah, sorry, I interrupted. You were in the middle of the flow of talking about other digital dependencies. And yeah, no, no, well, I mean, that's another great example, right? And so, like, I don't think it's realistic to try to eliminate all of these things. But to the extent that you can be more resilient,

to some other entity, just unilaterally deciding, maybe in a way where they think there, you're acting in your best interest, like my friend in his AWS account, maybe was. Or not. It's sort of like you don't know and you don't really have a way to know and you don't have a way to validate what it is that they're doing and be like, actually, that's not how that works. to the extent that you can limit that liability, it sometimes makes a lot of sense. And so like, you know, I feel like open source was like a great example of this where, you know, I certainly remember early on my career the pain of having to deal with, you know,

third-party software vendors and you would try to understand how their software worked and they would give you answers that were clearly wrong. and then you know you maybe tried to dig into it a little bit and then you know i had the experience very early on in my career of like you know i don't even think that it was any more complicated than like reading through like the scripts they had sent us but they were like don't modify those right and i'm like okay i'm not i'm not modifying them i'm just i'm just reading them like i'm trying to understand like this thing that you sent me that we paid a bunch of money for that doesn't seem to work and like god forbid if i had actually tried to like reverse

engineer and their software, right? I can't imagine what their reaction would have been to that. Like, maybe this is a lawsuit. But, like, you're stuck in this place where you're like, you know, I don't understand how this works. I am not empowered to change it. And yet I am dependent on it, right? Empowered is the word, the active word. That's really interesting. Yeah. It occurred to me that, like, another thing that was like early Google search engine optimization. Once Google became the monopoly, like, you know, you had this thing where an algorithm would change and people would drop off of the listings. And again, it's like, but they couldn't choose that. That's not something they could choose to be in or out.

They were like dependent on something, but not because they chose it, which is even, but yeah, the empowerment is, is an interesting aspect. I do wonder, yeah. So what can we do about this, Ben? Well, I mean, I'm kind of stuck, right? I mean, I will have my own internet search. I'll just, you know, my own data center. Right. Mesh network internet for everybody.

That's the solution is we're all just going to get some powerful antennas and put them on the top of our houses and mesh all the houses together. That's how the internet's going to work. It's going to end up in some crypto scam, though, isn't it? At some point, yeah. Anything distributed nowadays ends up in some, well, there's a, we can make a change. There's a coin, you got to buy the coin, and we're issuing the coins. AI coin.

It's something from these coins, exactly. Or token coin. I'm sure it's been done. I'm so it has already been done. The terrible idea that launched a thousand slide decks. Yeah. So, you know, like, I think it sort of depends on how far down the sort of, you know, crazy tinfoil hat, you know, internet libertarian, you know, route that you want to go. And I say that, recognizing that, like, a lot of the things that people, you know, from, you know,

like the Electronic Frontier Foundation and, you know, some elements of the open source community, you know, 20, 30 years ago, a lot of things that they were saying, you know, with panic in their voices would come to pass, have come to pass, and nobody seems to care. So like the amount of spying that like all the devices that you pay for do on you is like off the charts. Somehow, somehow that's okay when, I mean, the thing is there is sort of, I've always thought of it as there is some kind of implied social contract between me. And I'm going to pick on Google, again, just because I'm a bit of a fanboy for it. And I feel like having been there, I've got some standing in there.

And I've always felt like, yes, it serves ads, yes, it reads my emails or does something with them. And I don't mind that because I get a decent email service that I don't have to pay for, and it's integrated with everything. And the slightly spooky face recognition is great when, for example, you are preparing a family board. And I'm like, hey, let's just find pictures of my father-in-law. Here's hundreds of pictures of Tony.

And like, oh, that's great, right? That was, that's kind of an amazing thing. But yeah, I do realize that I've given up something for that. But I've probably given up more than I think I think about. And I've probably given up more than I would if someone set me down and said, well, what about these things? And I was like, oh, yeah, I hadn't really thought. Oh, that's a good point.

And, you know, but I'm getting something back out of this. And like some of the things though, I'm not getting something back out of it. So like, you know, automatic number plate recognition stuff, the flock camera stuff that's going on around. You know, there's a lot of things where it's like, yeah, I'm that's, I'm uneasy because it is not a, ironically, it's not directly a company making money. I mean, it is, obviously that is a flock current, but it's like law enforcement, a government thing that is for the benefit of the people has decided that they would like this information and then they would be happy to pay someone to put cameras up to do. all this kind of stuff. And that feels like, no, that's a step too far somehow. Now, now, I don't know where the boundary is there. I mean, it feels obviously, while I'm sort of choosing or at least choosing not choosing not to,

but then I get so much benefit and how will I get the same benefit without, you know, putting together a whole bunch of open source software, which is maybe where you're going with this, as you mentioned, the mesh network of like internet stuff and a big. Yeah, I mean, this is kind of what I'm saying is it sort of depends on how far down the rabbit hole you want to go, right? And I think for me personally, first recognizing that 99, I'm just going to put two nines in that, percent of people are not going to want to go down the rabbit hole at all, right? Like, they're going to want to do the minimum amount that they need to do to keep themselves safe and secure from like scammers and fishers and, you know, people like that. And, you know, some of them,

like, don't even really care that much about it from some of the, some of them. some of the things I've seen. But, you know, that's like the minimum amount that they're going to want to put into this. And the rest of it, they're like perfectly happy to be completely dependent and exposed in like a data sharing type sense. Like I'm very happy to trade my data for free email, right? Like that's fine. Yeah.

Yeah. Right. Like they're very happy to trade their data for just about anything. And in fact, many of them, what was the, was it snow crash that had the gargoyles in it? You know what I'm talking about? Did you read Snow Crash? I did read Snow Crash. And now I'm just trying to think about, I don't remember.

I think it was Snow Crash that had basically like there was a whole profession where you just walked around with a bunch of like data gathering equipment strapped to your body, right? Like cameras and like infrared and like, you know, microphones and all these things. And like that was your job. You just walk. It was like you're like a little mini like Google Maps car, but you were just a person. A person doing that. And that's what you did for your, you know, for your livelihood.

And they referred to these people as gargoyles, right? And I think like the great trick, it's like, you know, so much of what is being built is like, you know, like 90s cyberpunk dystopia. Right. Literally using those books as like a user manual for like how to build a society. Yeah. And I feel like one of the great tricks of social media is, is basically. like turning everyone into gargoyles for free. Like you don't even have to pay them, right?

Like you just walk around and you say like, hey, you should post us to your Instagram and then people take pictures of everything around them all the time and just post it with, you know, geotags on it to social media and then it's like, all right, cool, we've done it. We've gathered all the data that is. I remember like,

the surveillance things that like found all of the photographs of like a monument somewhere in Paris or something that were able to reconstruct a full, high detailed 3D model from like just that. thousands of tourist photographs of it, for example, that they've been posted to their Instagram. So, yeah, we're in, we're in that world, aren't we? Yeah, gosh. Yeah. So my point is, is that 99% of the people, maybe 95, I don't know, some very high percentage of the people out there are not going to be interested in going down this rabbit hole at all. But I think it is important, personally, I think

it is important to be able to defend the ability to do it if you want because, A, it just might make you sleep better at night. Like, I know a lot of anxious people. I know a lot of paranoid people and they're just like, that's who they are, and they just want to know that these things are going to be safe and secure, and then they reduce their dependency as much as possible. Right. And they just want to be able to do that. And I think that you should. And also, there are lots of situations in which, you know, people who may be in situations where they can be easily exploited by governments, by other people need to be able to keep themselves safe and secure. And having the tools available to them to do that is important, right?

And, you know, I don't know that I can go in and enumerate all of those situations, but I think we could, we could certainly imagine lots of situations in which that might be. not only something that someone would want to do, but be reasonably morally justified in wanting to do it. And so for all of those reasons, I think it's important to be able to do this, but it is also really hard. Right.

I mean, so like, going back to the beginning of this or not nearly the beginning, as you mentioned your friend's AWS account. I mean, so if they can't recover that account, do they go and buy servers and rack them? or, you know, is there a happy end to this story, by the way? Right. I mean, immediately what you probably do is hope that you did everything with

infrastructure as code and then try to port it all over to, you know, Google Cloud or Azure, I guess. I don't know. It's just switching one. Yeah, you know, the same category. One feudal lord to another feudal lord. Yeah, like, okay. Bezos, feudal lord.

You heard it here first. So, you know, like that's like on day one. And that's probably as good as it gets. Yeah. But I think this is why, like, you know, a lot of times, like, especially in our industry, people will wind up building their own data centers, not because they strictly have to, but because, you know, when things go wrong or when they have problems or when they want to try to do things that are maybe not standard, they want the level of control to be able to do it,

right? Like, they want to be able to say, like, yeah, we can put whatever hardware in these machines we want. We can, you know, configure them however we want. We can use whatever network technology that we want. We're not limited by what a third-party vendor is willing to offer to us, right? Obviously, that is out of the reach of most typical people. But I will say, I think one thing that is an off overlooked, you know, kind of like not that expensive way to get, yourself a lot of digital independence is just get a cheap computer and put it on your network

and leave it on all the time. Right. And, you know, if you want to take that up a level, like get some remote access for it, right? So you can go something, something to make it. Yeah, exactly. Yeah. Yeah. Most of those types of computers that you can buy use very low power. So you don't really need to worry too much about how much power they're drawing. They're quite powerful in terms of like the compute that they can you know supply and you will probably find that most of the things that you would want to run in some other service provider you could run on those types of machines today yeah and in a weird sort of like everything is backwards and i don't understand how anything works anymore a absolute trick that you can pull is you know use the lLM of your

choice which you are absolutely dependent on these days for at least To build you the software that you need to run on that computer to replace whatever it is that you're dependent on the third party services to do, right? So if you have some like, you know, awesome note-taking app that depends on some cloud service and you're like, wow, if this ever disappeared, I'd be totally screwed. And you want to go down the rabbit hole, which again, 95% of people don't. Something that you could do is you could build your own and run it on this computer that runs in your house. and maybe, you know, have some backups that, you know, write to a USB drive or, you know, go back to the cloud. I was going to say that's one thing where maybe it isn't holding you to ransom

unless two things happen. One, you need your backup and two, the cloud provider decides that they're not going to let you. Right, right. So that seems like a dual uncorrelated failure or whatever we used to say. Right. So, yeah, that is an interesting call to action. Now, I wonder, and we have touched on this before, because I'm getting a slight deja vu feeling here, but I wonder, is the fact that we have been so lucky for so long, like 20, 30 years in the sort of software profession that we're in, where essentially we've always been able to just, inverted comments, at quotes, just build it ourselves. And with open source, that has been free.

Like, is that spoiling us? because like if I worked in the manufacturing industry and I made grommets on a machine and then the vendor that made the machine said, no, you can't use this machine anymore, that's a real problem that many grommets manufacturers may have always had. And they might have to switch to another provider at short notice because they've gone out of business and now they can't get the report parts to repair them. So there's a little bit of like a, you know, other people have always had to deal with this dependence. It's like unless you want to become a core, your core competency is building lathes to make grommets.

as well as the Grommets themselves, then you've always had to outsource some significant part of your endeavor out, whereas you and I have been always like, oh, an email client, yeah, I can't be bothered to write an email client, but I could in principle. I feel like I could. Is it, are we just spoiled? And then does LMs now move us back to where we were to start with, you know? Yeah, I think a lot of the difference there is that once you've bought the Grommets, they're yours, right? The company who built the Grommets, Denver shows up and says, your license to these Gromits has expired. And we're going to go into your...

Maybe in that instance for the Gromit manufacturer. But yeah, either way, buy both of those things are true. Right. Like, it is possible to own the thing and it's yours, right? Really interesting. I mean, it's kind of like servers, right? Like, you buy a server. You own the server.

Modulo, maybe some weird things that happen when you ask LLMs about how those chips work. You know, you can do... pretty much whatever you want with them. Yeah. And like, you know, you want to screw them up and break them to experiment? Go ahead and do that, right? Like, there will be no ramifications other than the fact that you lost however much money you spent on the server.

Yeah. Right. So I think that is sort of like the fundamental difference. And some of it gets back to like software as a service type models where it's like, you know. You never quite own. And I mean like, you know, video games these days, right? You know, hobbies are my hobby is old computers and I've got like,

buckets and buckets of cartridges and CDs and DVDs even that I can put into various old machines and they work fine. But like nowadays, if you want to play a game for like PS4, PS5, it's like you don't really own it in any meaningful way. It exists as bits on the disc. Maybe it doesn't even exist in a hard disk anymore. It just exists in the cloud, which means that it goes away. When you don't own it anymore, you're renting, you've got a license to own something, which is like, right, right. Yeah. I mean, you buy a book and you read the small print in the front copy about, you know, you may not, you know, give it away or whatever. There's some liver restrictions are full for what you can use it for.

But like that's, that seems like those seem very reasonable compared to end user license agreements in software or whatever else. Yeah, I don't know. How do we get here? How do we let this happen? Well, the 90s cyberpunk set the model and all we're just going to do all these things. I mean, I don't know. I'm being, I'm being a little. Is this Neil Stevenson's fault?

Is that what you're saying? Yeah, exactly, right. Oh, man. William Gibson can be. William Gibson. Oh, yeah, yeah. Yeah, neuromancer as well. We think, yeah.

Yeah, Neuromancer, Mona Lisa Overdrive, and, oh, God, what was the? My nerd credentials are definitely going down here. I don't think I've ever read Neuromancer. Like, counter zero, count zero? Count. That would. Count bin face? Count bin face.

who sadly lost today at Clacton on C. I love my country. Oh, did he? Yeah, it didn't work in. 26% of the vote, though. It's not bad. I mean, the other guy, the guy who resigned and who's what he named here did win. Nigel Binface.

Yeah, no, don't besmuch the bin face. Yeah, I mean, it's, yeah, you've got to love the Brits for not taking things too seriously, but still, you know, I mean, someone pointed out that like the people of Clacton, the place where, you know, this is real for them. So we're all treating it as like a bit of fun. But, you know, and it is making some political points. But like, yeah, it's a thing. Anyway, that was not on my list of things to talk about with you today. But you gave me an opportunity to talk about it.

So count. What was you saying? Count Basie? Count zero. I think that's right. I think it's like counters. I forget with the middle. I love to count.

But, but I was, but. All of those, I mean, you know, every... Okay, I think... What is it? Who's your favorite vampire? The one from Sesame Street? He doesn't count. Oh, I assure you that he does.

That's all right. That's brilliant. I don't know how we're going to tell that now. We've peaked 20 minutes into this conversation. We've peaked. Oh yeah, no, we're talking about 90s cyberpon. So you were saying that this is the blueprint. So what's coming in the future?

Actually, no. Is there like so I think we've talked a lot about how few people feel about LLMs. And I think you and I have both been on various roller coaster like ride journeys with this is great to this is terrible to this is I'm going to lose my job to know. I'm now the most powerful programmer to now this what a mess I've got. And everything in between, right? And that's not even take into account like all of the moral and ethical aspects, which are also super important, but I have sort of compartmentalized. But, you know, not being able to run an LLM if, for example, tomorrow, Anthropic decides that I have been a bad boy.

Or my company is decided that, you know, they're not liked by Anthropic or whomever else they're providing or whoever, you know, whatever, that's a concern. So you mentioned about running your own stuff. Can we do that with L&Ms? Do I just need a warehouse-sized computer to be able to run something? Yeah, you know, I think it's arguable that the end state of all information is to be free, right? You know, it's like all of these models of like copy protection where you have like a private key that only three people in the world have. It's like someone is going to figure that out. Like ask Sony, right?

Like someone is going to figure that out. This is not going to work, right? And, you know, the model weights in LLMs are just too valuable for people not to try to figure out what they are. Right. And especially when you consider the fact that there is no like world government entity that can, you know, enforce any sort of of protection for any of this information in any sort of like global way. Are our model weights different though in the same way? They're more like a compiled program, right?

If you imagine I give you the EXE of my code, but not the source code to it, I give you the XE and you can go away and use that program as much as you like, but you can't really do anything more with it. You can't extend it or build on it or anything like that because it's kind of like already been compiled. It's very difficult to reverse engineer. And I feel like the weights or even the open weights that you can go and get on the internet have that flavor. It's like, sure, here's the result of compiling.

But, you know, we won't, you can't buy the compiler because it's like 45 data centers full of machines that ran for like two years to compile the program. And now here's the program. So, you know, the fact that the program is dressed up to me as a service is, you know, problematic. But even if they handed me the weights, I guess that's what I asked you in fairness. But like, I'm just now thinking ahead, right? you know, you could give me the, you know, 24 terabyte disk that contains the weights to the best model and I could somehow cobble together running it. But I'd be stuck with that until the next time somebody hands me, the 48 terabytes of the next model. Sure. It feels like, I don't know,

I've always balked for the term open ways because it's like open. Yes, I guess. I suppose so. Right. Yeah. Sorry, anyway, I cut you off. But yeah. Where were you going? I mean, you know, But that's a fair point and that is definitely, you know, an avenue that we could explore here is like how open our open source model. What does it mean? I mean, because like, yeah, open, I suppose that's it. I compare it to open source and open sources, you know, like, hey, here's the, here's the source code. Now, that might as well be alien to most people on the planet, but in principle, you can look at it and go, this was meant for us to build on and continue. And in fact, the people who gave me this or gave me access to this, they are using this to continue and get better.

the next version of the product will be built on this most likely, you know, obviously there are edge cases to these. Whereas the weights feels like so derived, it's not really the cool and interesting bit of it from the point of view of continuing and being able to build on and not being locked into this brain forevermore. Yeah. I wonder what, what, but then the resources are so, so massive that I can't imagine. Yeah.

That you could do much with it. And but at least you could, right? So in principle, and have you, have you tried this? by the way, have you experimented with your local models? When you're running local models? No, it's been on my list of things. It's like there's going to be a tipping point probably in the next 18 to 24 months for me where it's like I'm going to like part of my day job is probably going to be figuring this out.

And so like I'm kind of, I'm just waiting for that to happen because then it'll it'll be like more of a focus, right? You know what I mean? Like it's not just like something I'm monkeying around with. It's like, oh no, I need to make this work for real. You actually have to understand it. Yeah, you know. And like, and the world will be completely different in 24 months. Technologies will be different.

Everything will be different. So I might as well figure it out for real. Why I need to figure it out for real. Yeah. Yeah, it's interesting. I suppose, I mean, we're, sorry, excuse the dog. We are, we're meandering all over here. But like, you know, like even getting the equipment to run these things on,

obviously one can find GPUs around or, you know, in the house. I have a few around, but they're not really up to scratch for this kind of stuff. And then, you know, like, I know from my own. day job that sourcing, you know, top-end GPUs and the RAM that goes with them is incredibly difficult right now. And I wonder if that's going to change. But one can only hope that there are some advances in the way that we evaluate, not evaluating to make better, how we run inference or whatever on hardware. Because GPUs are designed, well, for graphics. I still feel sad for rows and rows of graphic processing units in a in racks that have never had a video display

plugged into them ever they've never generated a frame of a game which is the point of a GPU in my opinion but you know it makes me sad it's like battery farmed hens you know it's like it's not right be free do what you're supposed yes i want i need some free range GPUs please yeah well i've got a drawer of um i've got a voodoo two in a drawer somewhere i think that'll help right we feed them a steady diet of doom eternal and oh that's too new in my GPUs by my people more like VGA adapters with a bit right yeah anyway um yeah no I there's I mean well and this is kind of what I'm saying about the um you know the information wants to be free argument on models is like yes it does take a lot of resources to train them and generate those weights um but once you've done it like

it's not, they're not going to disappear. And imagine a world where there is an open source, uh, agenic coding model that is better than say, Fable 5, right? Like that is probably going to happen at some point, right? Like, do you maybe have some situations in which you want to use a better one? Sure. But if you're just trying to like vibe code the notes app to run on your person,

personal server. That model will be more than sufficient, right? And I can see a world in which that kind of open. The thing that I'm interested in, and now this is going to be part two of a massive thing, is that after the base training of a model, like, where you've just fed some subset of all human knowledge in through however so you've acquired it, which is its own massive asterisk and footnote and ethical dilemma. But you've got some amount of knowledge.

should you train a neural net on it. And then there are so much post-training that happens to make it valuable, make it aligned, make it safe. And then tuning after that to make it, you know, like, hey, don't write code that's terrible. And that's the bit I'd be interested in. I'd like to be able to go like, hey, I don't want to have to explain in text over and over again to a model that don't write comments like this, for God's sake, stop it. Because, you know, you and I am sure have done that a million times.

But I want to be able to say, like, why didn't we just. turned it, can I personalize the weights so that it doesn't do that in the first place, rather than me having to exhort it every single time to do it the way I want. And that feels like the last steps of an open model where you're like, okay, that's tractable. Here's the 20 terabyte intermediate model. You can add layers of tuning onto it and then like maybe that is something I could do. Yeah, maybe I'm just beyond. And you know, yeah.

I have to admit, I don't know nearly enough about how that process works. That process works to know whether that's a feasible thing or not. No, but it seems like I can imagine. an open company is starting that process or you know maybe maybe this stuff exists on hugging face maybe this is the kind of thing that I you know we would I would if I spent a bit more time looking into it I would discover it can actually be done yeah but no it's interesting and yeah I just so on us the irony here that the sort of leader in open weights as far as I can tell is the Chinese government that seems like wow yeah that was not on my bingo card for this year is like champions of

openness and freedom yeah yeah I'm sure you can find a, as you said, we need an AI company that's open. What would we call it? No. I just can't think of a good name. No, there's no, I don't think of the domain. I'll just check some domains.

Yeah, see if whether you can find one. Yeah. But yeah, like it's entirely possible that in the limit, the only thing that matters is GPU's power and cooling, right? Because like models will get good enough. Their information will leak out into, the world as all information does and you will reach a point where the you know you don't necessarily have to be beholden to a company to use their models as a service you can go and either purchase

or maybe just acquire for free because they're just freely available a model that you can run on your own hardware and then you can purchase that hardware and then you own that hardware and no one can take it away from you and then you could have you know it's like that that that server that's running in your house maybe becomes a little bit more beefy and uses more power. But it is yours and it's yours right for you. No one can turn it off and no one can tell you not to use it and you know. And you're using solar panels so that the electricity company can't cut you all further and further down the rabbit. And you got the mesh internet that you're not connected you don't have internet service provider. Except that only works when I'm not inside the

tinfoil wrapped house. Yes exactly. It is right gosh um well. Well, Well, friend, we've gone round and we've covered an awful lot of things that I wasn't expecting to, but you and I haven't spoken for ages, so this is not too surprising. It's rambling in some ways. But I think perhaps in the tinfoil wrapped house with our own utility-providing systems in place is where we should leave this with our own intelligent. Well, what I would say is about the power on cooling thing is that at least there is an existence proof of something which doesn't require that much cooling because I'm using it right now to talk with you. And I don't mean my mouth. I mean, although that is the primary way in which I think, to be fair, as you've probably realized.

But like, so it has to be possible to have something which has at least human intelligence that runs on what, 80 watts, right? Yeah, right. You know, something like that. Maybe I'm not as fast as, as a claw it is. But like somewhere in there, there's a scaling thing. So I'm sure we can make it happen along the way. And that's the hope. Somebody will come up with some cool 3D silicon hardware thing that is vastly more efficient at computing whatever it is. Or just fusion power. That works too. That would be awesome. Yeah. Yeah, we still got a lot of exhaust heat to put somewhere which doesn't seem like the best thing to have knocking around. You know, so yeah. Anyway, with that, on that bombshell, on our fusion powered tinfoil wrapped

home, I will see you next time, friend. Next time. You've been listening to Toos Compliment, a programming podcast by Ben Radie and Matt Godbaud. Find the show transcript and notes at www.2.2.2.complement.com. Contact us on Mastodon. We are at tuescomplement at hackyderm.io. Our theme music is by Inverse Phase. Find out more at inversephase.com.