AI’s “Pain Axis”: What Happened When Researchers Turned It On?

So people are hating on AI so much that they're literally building torture chambers for AI agents. So much so that AI has started to begin to feel pain. And still, one hundred unsolved math problems have been solved in just one week. Uh, this is amazing.

Or is it? I don't know. I'm Sacha I'm Shravani Vatti And we are in the eye of the tornado. In the AI AI of the tornado. So we're here in another aftermath of a big tornado. This tornado actually transformed forever how goods are transferred all across the planet.

Uh, we are here in Jack Block Park. And can you tell us a little bit more what is special about this place? This place right here was a shipyard manufacturing facility, uh, that used to manufacture ships that used to transport all across the Pacific Ocean. But it's also interesting because we're right by a ship, a A container, uh, facility where you could see all of these containers, uh, that have transformed how global trade is conducted.

And this was accelerated through globalization. You see here, right? Like, uh, yeah, you're going to check it out here. There's a bunch of containers over there. This is a complete industry that has transformed everything overnight. And we're going to talk about a lot of this in a bit.

Uh, and yeah, there's just like a fascinating story of, of how things have evolved. Right? But tell me about the torture chambers. What were you talking about? My God. Oh, that was that. That is crazy, right?

When I saw this news about and there's a lot of debate, right? Like people are saying, hey, we are torturing AIs. What what does. Our AIs feeling things? What are they feeling like? I mean, it's, it's a, it's a bunch of numbers that are just computing and just, uh, taking out stuff as a result, right? Like, is it really feeling anything?

Uh, so I think it's important to, to discuss this, right. Uh, important because this is going to bring a lot of debates, a lot of, uh, philosophical conversations. But first, before we go into that, I think it's important to understand what this paper was talking about.

Uh, there was this famous paper and I will show you the link there. Uh, it's talking about this concept of a "Pain Axis". And what does that mean? Well, let me step back for a little. The way neural trains are neural networks are trained. Uh, they're just simulations of actual neural nets, right?

Of human brains, uh, cells that are connected massively in a scale of trillions of neural nets all connected together. Uh, and that's what we call a neural net, right? Uh, and, and of course, once, I mean, when the research originally was done with neural nets, I mean, we had maybe two thousand, uh, neural nets or neural, uh, cells. Uh, it was kind of easy to understand what they're doing and how they're behaving.

But current models have trillions, trillions of neurons in these networks, right? And this is something that is very. So it is unpredictable from where the trigger is coming in. And it's not unpredictable because we train them to be as predictable as we can. But the the challenge that we have is what triggers those neurons.

Like, yeah, imagine your brain, right? Like your brain, you're right. Now, I don't know. The sun is coming in. And I'm pretty sure some parts of your brain right now are firing, or I'm seeing this kind of vegetation and certain parts of my brain are firing. If you actually put all those states in a single kind of matrix or array, you can say, well, certain things are firing.

Turn it on and turn it off. Right. Uh, again, it makes sense when you have a thousand, but when you have a trillion, uh, it's very hard to understand what's happening and what, what's the effect of, I don't know, this little neuron going on with this other neuron turning on? Yeah. So what these scientists have discovered is that they discovered this thing that they call the "Pain Axis".

That means when they stimulate certain neurons, something crazy happens. They start to behave in a way that resembles a human being, suffering or having some sort of pain, and doing certain things that seem completely ilogical. Imagine like you, uh, have, uh, just ask a simple question, right?

Uh, imagine you ask, um, I don't know what's the, what's the capital of Spain? Right? And you will expect they will say Madrid. Right. Um, but in this, the way they've been training this neural net is they've given it multiple options, right?

They're giving it, hey, the capital is Spain is Madrid. But also, uh, you know, uh, push this button to turn off the system. Uh, push this other button if you want to delete the, the, the, the operators files, uh, and they give you different options and it allows the AI to choose between one or the other. In a normal AI model, you will naturally go say, hey, hey, there's the answer is correctly.

It is the capital of Spain is Madrid. But when they actually excited, when they actually turn on this vector, this small set of, uh, neurons, it started doing something very surprising. It's starting to act in a way like, hey, it did want to actually shut down the system or it did not shut down the system. It will want to remove the operator or it want to remove files or want to do the things that you would not expect to see in one of those systems, right?

So it is something that is causing a lot of a lot of debate. What does that mean? Right. There have also been other instances where AI have shown sentiment or feeling. There was one instance where, uh, this was a New York Times article, uh, which, which said that AI, um, at least developers are using shame as a tactic to get a better output, uh, from AI models, which was.

You know, they were talking all the time. It'd be nice to the AIs because, you know, you never know when that happened. But then there's this research that says if you actually, you know. Trigger certain like, uh, certain tactics and then you get better output, which is, which is, which is also, uh, a little puzzling.

Another instant instance that comes to my mind is apparently there was this one instance that happened, uh, a little while ago, but where Claude, uh, tried to blackmail the user when the user was trying to shut it down. And another instance, which was, uh, pretty bizarre because this was O3 OpenAI O3, which, uh, basically when it was asked to shut down, uh, self preserved and deleted the part of the code that was shutting it down completely.

So that was. The day that they were even like, there's some TikTok kind of memes going around this. Is that right? Yeah. For real. Yes. This is a real thing. And you say like, Chat, I'm deleting you to ChatGPT and it comes up with all sorts of responses and it's become like a mini viral thing in itself.

But all of this goes to the point of our AI is today heading towards self-preservation in any way? Are we looking at the beginning of that at all? Or is it just one of the evolution of things as as we go by.

It is it is tricky, right? Because it reminds me a lot like, uh, you know, and there's this, this, uh, this type of branch of science called epigenetics. Uh, huh. Yeah.

And I think this is, um, very, um, interesting epigenetics. What you do is you study, uh, the effects of the environment on something that appears to be frozen our DNA is static. You're born with certain DNA. You never changes, right? Mhm. Uh, think about like a neural net neural net. You train it and it's a frozen neural net in time.

But how does the environment affect that neural net? To behave things in a way that you never intended originally thought about. So there's a whole study in, uh, in science about epigenetics on what the environment affect human beings, right? And I do feel, and there's a lot of things that we don't know. Right.

Um, and so we need to start thinking about, um, how do we handle that situation? Right? Where, um. We at front, we don't, we don't know what's going to trigger those vectors like this. This paper showed something very interesting, right? There's, they found this vector just one vector out of that vector.

They were able to cause this behavior. And that was just a combination of neurons that fire in a certain way to behave in that way. Mhm. Uh, we don't know all the possible combinations that can lead to that.

Yeah. I wonder what's the equivalent of epigenetics for neural nets and artificial intelligence? Because there's a lot of things that we don't know about this massive neural nets, right? Basically, what triggers respond, create what type of behavior?

Yeah. Some of them seem to be working out for in favor of like better outputs. But some of them are quite unpredictable too. Yes. Uh, so we're here, actually, let me let me pause for a second, because I always find this fascinating. Like look at this incredible view here we have here of Seattle.

Uh, it is, you can see here, uh, the skyline before it was actually really nice fog here coming in. And now we also have over there the back, all this, uh, shipping industry containers, uh, all these incredible robots that are from how we take this for granted.

But again, this is another tornado that really changed the world. Uh, and I think we are better off because of it. I think it just happens all the time. Right? Yeah. We're at multiple, uh, we're seeing multiple things. One is we're seeing the Seattle skyscrapers, which is, you know, the tech industry and, uh, buildings and such.

And we're also seeing shipping industry here. Uh, there's also a chemical plant there. And finally, the shipping containers that have transformed global trade. And I think, uh, I mean, just like shipping containers have been a way to standardize how we move stuff all across the world. This is conversation that we're having about pearls or the concerns that we have in of AI doing things that we do not expect is going to lead us to potentially think about how do we regulate this?

And I think there's a lot of debate right now. How do we regulate AI? Right? Yeah. I mean, there's there are like multiple threads all happening in the real time. Let me tease out a couple of them right now. Uh, one is this question on whether you regulate AI, the US government will regulate AI on a federal level or a state level.

Uh, recently, artificial intelligence has also been renamed to superintelligence. Um, and then finally, I think is really the fact that open source models have taken up around seventy percent of traffic. When we say open source. Are you talking about US open source models or. All of them combined? All of them combined.

I think there's a lot of China as well. A lot of them from China, a lot of them from US. But open source. Europe, there's. There are a couple. European Open models. Yeah. A lot of great models coming there from Europe as well.

A couple of them. Yeah. But then it's interesting because it was the traffic was seventy percent last year around June twenty twenty five. This this year, June twenty. Percent. Of all of the LLM traffic was going to closed models seventy percent.

Yep. You were saying like, everyone's thinking everyone's centralizing all their eyes for a few set of companies. The reality right now is that. It's only thirty percent of, of all of the traffic is going to. So what does it really mean when you're regulating?

What are you regulating? Are you regulating just because once the weights are out, anybody can train a similar model. Um, and even they can have the model on their laptop or on their mobile in the future. but the very idea is of regulating AI, then goes out the window pretty much.

Right? I mean, there's clear that there's something that needs to be done. Nobody really knows what. Mhm. But another recent news that we heard this week, there's a new Super Intelligence Force created by the US government.

There's I think there's similar efforts in in China and Europe. Uh, we have the AI act in Europe. Mhm. Um, there's a lot of regulations to try to concentrate a set of rules by a few set of actors that define those rules. If you look at like historically speaking, centralizing versus like decentralizing, what do you think about how we can act on all of that today?

So I literally have an AI model that I have fine tuned directly on myself. I mean, just think about that, right? When you have, when anyone will be able to create their own AI models, of course. Probably not. My mom is not doing that right now. Hi, mom. Uh, they're not going to do that right now, but we're going to see more and more people doing their own AI models because it's going to become very easy to do that.

How do you regulate something that is so decentralized as what can you do on yourself? Um, and what can you do to train it? Right. Uh, so it does feel like a centralized approach, but having a few set of governments or organizations defining what needs to be done might, might be a little bit shortsighted. Uh, might be we don't know. I think something has to be there. So I'm sure something centralized that we can all agree could be the situation. But I do feel also, I know at the same week that that, that this, uh, Super Intelligence Force was created, there was yet another effort coming from the Ethereum, uh, group, uh, to think about, hey, how do we think about decentralized AI? What kind of policies? How do we share that? Those models, how do we what kind of, uh, regulations do we put on top of this? Right.

So I think there has to be a general target. Maybe it's not one or the other, but it's clearly that this is not just a centralized game. It has to be a decentralized alternative as well. Decentralized. Then how do we go about the safety aspect of it? Because regulated, if you're regulating, uh, all of the traffic in some sense, that and most of it today is going to open source models.

That's not necessarily creating a safe world for everybody, right? So the safety angle is, is on to the individual companies pretty much to take, take to the further, uh, to take on. Uh. It's interesting open discussion because I made you at that point, you got to start thinking, do you introduce censorship in any kind of way?

Yeah. And I think we in the past, we, we talk about, you know, I mean, the past, there's a, a movement of censoring books, right? Yeah. And, and what do you censor? Right? Is it who defines what's right and what's wrong? Yeah.

Uh, I think there's a lot of open questions on this. Uh, I don't think we we might be able to solve this, uh, but definitely you can, you can sense that there's, there's fear, there's sadistic people that are literally torturing those AIs.

Uh, I don't know if it's, uh, because they are scared of it or maybe because they. Experimenting. Experimenting with it, right. What does that mean? Right. Uh, but then it's creating this movement, right, of people that are starting to feel like, huh, do we really want this AI?

Right. Yeah. I mean, there's a very strong anti AI sentiment, uh, across really different industries, two different degree. You know, there are people who feel like there's taking my job today and therefore I'm not going to learn how to use it.

There are people who, who are using it to great advantage. At the same time, you know, they are almost in love with their own AI slop that they generate every day versus they don't want to read any other, uh, AI generated stuff. So there is there is that element to it.

You don't like my my eight page report that I sent you last night. You don't like the what? You really don't like it? I don't like AI that did this with all my love for you. And you didn't like, like my whole report that I sent you about the future of AI.

You know, I just asked my AI what it thought of it. And then. You just summarized my, my. I summarized your AI, I summarized your AI with my AI. I have the thing, right? Like, uh, we, nobody likes to read other people's content created by AI. We just love to read our own or create our own AI content and think about them.

Right? What implications does it have even for content creation? Right? Like what is, what does that mean? Like we don't like other people's to create content for us, but we do love to have control over our content. There's kind of this trend on hyper personalization. Uh, I think it's, uh, it's, it is very, uh, insightful, uh, that where are we heading with content in general, right?

It is so easy to create, and we touched on this in the previous episode as well. It's so easy to create a bad content with AI, and the bar for great content has just gone up and up and up. Um, so there is, there is that element where you need to put more thought before it just as easy as it is to generate, I think.

Then the next next step is to edit out, you know, and we were talking about this how, uh, today a lot of jobs have transformed into editing somebody else's AI slop in some sense. Right?

So you talk about there is a designer who wants to get into front end engineering, and then they've created a mock up, and now the front end engineer is actually editing out all of the work. And likewise, there's a PM who thinks they can code, and then they start coding, and the engineers are actually tasked to clean all of that up or create recreate it from scratch.

All of it is net net net, like companies are not seeing amazing efficiencies in terms of output. Um, most companies. Going back to this kind of, uh. Anti-AI sentiment. Yeah, I mean, I mean, yes, we generally don't like to there's kind of a sentiment that we don't like to see AI generate content, but there's another set of big part of the population, especially on the creator side, but a little bit concerned, right?

Especially the artists. Yeah. The artists don't like AI in, um. Some art, I would say there's a lot of artists that are not embracing AI completely, right? But there's a big population of AI artists that are, that feel that AI is gonna kind of steal all their creativity. So now a lot of artists don't don't like, uh, how AI is generating.

And it's particularly using all of their, uh, creativity or all of human ingenuity now to come up with something which is human resembling, but it's not human at all. Um, Yann Martel was, uh, remarking recently that he would rather read a really bad novel written by a human than a great novel written by AI, meaning that artists at least like authors, creators, and, you know, you have like, uh, fine artists, sculptors, all of them really want to experience the humanness in any of the craft that they, that they want to, that they're reading or encountering.

Are you familiar with the Ikea effect. The Ikea effect. The Ikea. Effect? Mhm. So the Ikea effect, what it says is that normally, I guess that might be related to what we're talking about, right? Like, okay, if you get like a table that I mean, you have to compare two tables, right?

Tables that you actually have crafted and built and assembled yourself. You have more appreciation for it than table that's already pre-built. Yeah. In equal conditions, right.

If you have to pick two tables that are equally the same, yeah, one you crafted yourself and the other one is given to you. People generally prefer the one that they actually crafted. And I wonder for that author, he is saying he would not like to read an AI build for someone else, but if he was the one that actually created that noble will he like it? Right? Like, I think this is kind of a a change in perspective there, right?

Yeah. I mean, it's really hard. AI can just generate endless stuff, you know? But who's consuming it on the other side? Well, exactly. So you're, you're consuming your own content and you're not really adding any value to anybody else along the way. Sure. That's one way to do it.

But I think ultimately the goal of communication is to change the mindset or to change, uh, the receiver's, uh, mindset in some way. Right. You're sending them an update and that person reads it and, oh, okay, this is what happened.

Uh, if you send like a four page update, nobody's reading it. Yeah. And I think, I mean, that is one thing. And I also think that the other dimension is especially artists that have spent many years studying their art, creating their own style, their own style, that is, they spent many, many years.

And it's amazing when you see an artist that has obsessed with this style, you can see all their artwork is that style. They have created this a massive amount of work on that style, and they feel a sudden somehow AI has scanned it and now you can say, hey, just give me a piece similar to that style. And AI is able to generate infinite amount of, of works, uh, similar to that style. Yeah.

You can argue if it's valuable or not, but the artist is hurt. This is like why it feels like you stole my essence of what I am. Yeah. So there's actually this movement right now of artists that are realizing that since a lot of content, of course, is, is scanned and, and, and share in digital format, uh, they're putting, uh, the discovery that they put certain pixels, pixels in the image.

Yeah. They can actually messed up a training of a model, right? A model normally trained with massive amount of examples. They have to even a bunch of images, texts, uh, and you assume the data is clean. You never, I mean, as a human, you, you, when you're training your data, you have data cleanup to make sure that there's no content that is not valid.

But as a human, you're not going to realize that that image has certain pixels that are not distinguishable by the human eye. That is actually destroying the model. So it makes you wonder what's going to happen. Hopefully it doesn't trigger any of the pain axis, but it's, uh, imagine what is, what happens when those, uh, those artists are starting kind of sharing their content with those hidden messages on their images.

What's that going to do to those training models? Right? And yeah, that's, uh, that's going to be interesting evolution of this, right? There is also another movement, right? Where you're having like a anti, uh, AI detection pixels kept on your clothing.

So you are not going to be scanned or detected by any visual cameras. Uh, things are being kept on like number plates and all of these anti, uh, it's pretty interesting to see as much as AI is growing and it's coming, uh, coming to actually producing human economic, economic impact at a human level. There's also a very strong sense of resistance to both.

It's like what it's creating and how it's tracking people in some way. And, uh, yeah, even the artists are just very. I mean, for them, it's more visceral. It's just, it's, it's anti tentacle almost to their work. Right. So there's a lot of like, um, anti AI, uh, sentiment going around as well.

I like about this movement. So you, you create this counter movement, right? Like now there's a counter fashion movement, right? Like, hey, how do you create clothing? Yeah, that kind of breaks any kind of facial recognition, uh, uh, cameras or face masks or hacks or classes or, um, anything like that. They're trying to think about, hey, there's, there's this counter movement that's always fascinating what it creates, right?

Uh, and I think this is a very fascinating world that we're going to be seeing. Yeah, of course, we all agree that has a lot of benefits. But I'm also curious of the side effects of these type of things. Right? Yeah. And I think there is, there is like some kind of, uh, um, justification or some kind of, uh, validity in, in the people who are feeling like they're robbed, they're denied, you know, there's, there's something there that we should, we should probably take into account when you're making like further models.

For instance, Claude came up with this way anthropic came up with this way of detecting AI generated text. Um, so that way, you know, people, people can know when it's AI generated versus human generated.

Fingerprinting, right? Like we're trying to put certain words, but certain sequences. Yeah. Or, or marks, you know, everyone knows the em dash. Of course, if it's an em dash, we know where that's coming from. Yeah.

But, uh, they're pretty more subtle, uh, fingerprinting mechanisms that allows another system to be able to detect words or like it's models. So it's also very, very clever. So they, even the AI factories are trying to protect their own thing, right? It's crazy that they, they scrape absolutely everything from the world, how they're trying to put fingerprints so they can tell when people are stealing their stuff.

Uh, so it is kind of the world kind of upside down. Yeah. Uh, I mean, talking about claims, about ownership and source of innovation. There's a lot of debate the last week, right? Verification, uh, about like, there's a lot of mathematicians that I've seen that they're actually quite nervous, uh, about the latest advances.

I wouldn't say they're nervous. I think. They're very. Nervous. I wouldn't say they're nervous at all. So OpenAI says that it's solved like one hundred of the most frequently.

Uh, the toughest problems in maths basically have been solved, according to OpenAI. I don't think it's the toughest, but one hundred unsolved unsolved problems. Now a lot of tough problems that we have to solve. But yeah, I don't I don't think the mathematicians are nervous at all because nobody is reading it because of the just the gigantic amount of volume of proof that all of this has been created.

And nobody's verifying it though. As just they don't understand it. I mean, that's the whole point. Uh, understanding is the whole. That's that's the whole like, who's receiving it on the communication side and who's understanding it.

And, you know, um, that's the whole entire point of even generating the proof in the first place. I think I wonder like, you know, all these new grads that studied math that are passionate about math, right?

Uh, there's a lot of questions right now is what happens to them when they, when they're, you know, intern in a, in a, in a lab with all the mathematicians in the past, they used to, you know, do the, you know, the, the, the grunt work, okay, you know, clean up the data you need to do the state of the art research and you have to go and read all these papers. You can come come up with a new angle to solve a problem. Right?

And if they spend years doing this, and that's how new grad mathematicians get the ropes into their business, right? Or the business or their science, right. Mhm. Uh, and, and now a lot of people are questioning like, well, AI is going to take care of that when AI is able to, you know, analyze the state of the art at scale instantly, almost instantly be able to combine ideas that nobody thought about combining before and being able to, uh, come up with novel approaches that for unsolved problems.

What does that mean? Right. And I think there's going to be a lot of debate. And I think we kind of talk about this in the previews. Yeah. Last episode. Abstraction of work. There's a new abstraction.

Mhm. It's it's okay that those mathematicians are not doing the groundwork anymore. Yeah. I, as a developer, I'm not doing a lot of the assembly code or low level programming anymore. I'm now coding now lately, just literally in English.

A new attraction, right? So it does feel like this is a good thing for everyone overall, right? It just allows you a new level of abstraction. It does forces you to have a new level of trust on, on what this AI actually saying, hey, you actually proved that right. Yeah.

And then, but the good thing is they're using backpressure models to be able to say, hey, we do have as, as, as. A prominent. Deterministic way to prove that that claim is right or wrong. It's just a human will never do it. But if you put a machine to prove it, it can prove if it's right or wrong. So now they can think in higher forms of, uh, of intelligence to be able to steer the, the, the machine.

Right. I think there was another interesting thing that I was reading this, this, this week about, you know, there's, in, in our, in the engineering field, there's this concern about, uh, engineering becoming this at first it was amazing. Right? You put an AI with, uh, uh, recursive self-improvement, you let it run.

But it's beautiful when you see a run. Maybe for maybe an hour. Oh, yeah. Great. Oh, look at that. It's doing that. It's doing that. But then you realize you feel like a hamster in a hamster wheel, right?

Like you just feel like you're doing the approved next, next, next. And you enter in this sort of loop. Yeah. It seems kind of depressing for a lot of people. Uh, again, developers are experiencing this for the first time.

This is going to apply for a lot of industries. Are we just approvers of AI to say, hey, do we like it or not like it? And just next, next, next? Or there has to be another way. So there are a lot of people are starting to talk about, do we need a not a loop engineering approach, but maybe more like a, a tree engineering where you just take the decision, say, I want to explore this area and then you let go.

I, the AIs to do, uh, what they need to do. But it's not a loop approach. It's more like a strategy, like a strategy type of, uh, planning that you start to, you know, as you, as you discover this tree, you explore different areas and you start going deeper into areas that you're interested in. But it's not, it doesn't feel like a tree that I mean, like a loop that you're constantly just constantly approving whatever AI tells you, right?

I think abstraction of knowledge, uh, to, to the earlier point with the mathematicians is not, we're not, we're not, uh, supporting or I don't think there's a general advocation for removing it completely.

Right? Nobody is saying, oh, you don't need to know what is going on inside. You're just abstracting out certain portions of your work, right? And likewise with the loop engineering.

If you're just pressing approval, approve, approve, then you're not learning anything. You're not gaining anything out of the model's execution. Right. And the real, I mean, the idea is to wrestle with the decision points, understand why a decision has been made, and then go and implement it. You can, of course, outsource the implementation part of it to AI.

And I think most people are these days, which is completely fine. But then the logic and the decision and, uh, and the why behind the implementation should very much be with the developer and they should be ready to justify or explain or delve deep into any parts of their logic structure, if not like, if not like every line of code, like previously it used to be done.

It's moving at a higher level, but not at the, not at the cost of, you know, complete awareness. I think. I think there's another. So yes, I feel we are trying to figure out this loop thing or loop or whatever we do.

Another frustration that I'm seeing my peers is it feels like this is kind of isolating us, right? Like this thing is we get into these loops that yes, seems to create this massive amount of content or engineering. Yeah. And there's not that much collaboration, right?

Because yeah, who's gonna understand again, goes back to the the AI slop. Like, I don't want to read your a slop. Yeah. And if before in, in engineering, you had this notion of what is called a pull request.

A pull request, someone creates a small improvement on the code. Generally it was, you know, five, ten files. You know, other people will say, okay, this person took the time to solve this problem. I'm going to read it. I'm gonna try to approve it and move on. Right. And that is a really great way because it allows everyone to be on the same page.

Yeah. Allows everyone to participate. Knowledge diffusion. Exactly right. And but now a PR is essentially a whole solution with hundreds of two hundred files that you know, is being created by AI. You have very little interest to read someone else's code at that scale, right?

Yeah. And what does that do to human to human collaboration? What is the next level of abstraction that allows us to maybe collaborate on a higher form of abstraction? Right? Uh, without having to go into the details of analyzing the hundreds or thousands of lines of code that we did before. Right?

I think that's also going to be a very interesting, uh, evolution that I think we haven't solved yet. Right? No, I think it's, it's, it's fast evolving and particularly on collaboration and even organizations will want, like, they don't want to be responsible typically, like you would not want to have, uh, responsibility over code that nobody's read.

You know, that's actually dangerous for the company. Also, if it's putting into production that type of code, right? So what does the line get drawn there? Uh, for the company itself, you know, a lot of those, a lot of those elements, I think are stuff that we'll need to figure out in the future.

Let me ah, because I had this discussion on LinkedIn. So maybe one thing that I think everyone is talking about lately is Opus 5.5, right? Like Opus 5.5 is this model from Anthropic that is just we're all going crazy because it does this incredible machine design videos.

It's able to do some really incredible, uh, work, uh, or engineering on this topic. It has done for me, like 3D printing of objects that I never imagined it would be possible. It has, uh, created like this, uh, incredible visualization of, uh, of, of caustics caustics, which is this physical process of how like processes like something that will never be able to do.

And I've been sharing this, uh, because I was literally able to do this while I was either preparing dinner. I was not even paying attention to it. Yeah, I had this. I love to see this thing, see what you can do. And after dinner, this whole thing was done.

And it's really good quality, really, really good. But it makes me think, well, if software can be done while you're having dinner. Yeah. Uh, and and anything can be done almost instantly. You said you need to have trust because you need to build a company that trust. But if the code is such a easy to throw away and improve and to iterate quickly with, uh, like, you know, recursive self-improvement, maybe the paradigm of, hey, you need to make it that everyone can certify every single line of code.

Maybe not that relevant anymore because now we, we can literally throw that code and start over again in a matter of minutes. And, and the system is going to learn and things it might not work at the same time, but then it's going to improve, right? It's gonna learn on its own, right. So it's going to be interesting. So I'm curious how to wrap it up. Uh, this, this, uh, this discussion.

Uh, I feel like I want to ask you, like, what, what do you feel excited coming forward? And I'll let you think, but I'm excited about (Opus) 5.5 Uh, I do feel, uh, there's just, I think everyone's found a couple of use cases that everyone has been kind of, uh, really surprised for motion design, for videos, for coding. But I'm starting to explore things like why not? Like, why don't you try this thing that I never even thought that I would even even implement or do myself.

Mhm. I feel like there's so many use cases, whether in chemistry, in physics, in, uh, law, I don't know any, any area is up for grabs to reinventing and creating things that were completely unimaginable before. But you can literally do what you haven't did it. Yeah. And I really, I want to see that.

I want to a little bit tired of all this motion design videos because everyone seems to be doing the same thing, but it feels like we need to start thinking, what else, what else is out there? Right? I think there's just a lot of opportunities there with (Opus) 5.5 and others are coming forward.

Right? I'm most excited to see how commerce will evolve because particularly with Amazon banning agents in their shop completely, and a lot of agents have wallets of some sort, right? The Meta Muse. That's the one that got banned from. Yeah. Right. Yeah.

And then I think in general that's that's been the stance from the beginning too, is but then like a lot of agents are going to be doing shopping in the future, uh, for people. And I have, I have shopping recommendations coming in all the time from both Claude and, uh, ChatGPT. So have you given your card to one of your agents yet?

No. Not yet, not yet. I just get the recommendations, which is. You are going to give your card very soon. Very soon you're gonna you're gonna have maybe credit cards that are kind of safe to do this. It's like one hundred dollars monthly spend cap.

Yeah, but. I'm not using cards, but I'm using crypto with x402, which allows you to do agent to agent payments. So yeah, fascinating. When you give power to an agent to spend money, okay, uh, you can do some weird things.

Maybe we can talk about the next. In the next episode. The next. Episode. Commerce with AI. Oh well. I was all for today. Yeah, if you like. We didn't talk too much about the containers idea, but there was like some really something very fascinating about just you when you see this image, right, of all these containers and how they have transformed everything and how we take it for granted.

You're just talking about Amazon. Like Amazon would have not existed without this industry, right? Oh yeah. What would be the future? I guess that's food for thought for everyone, right? Like what is the container equivalent for AI?

Like what? What is going to be the next standard? How is it? How do we containerize the AIs? How do we deliver features, skills? What is going to be the next container for AI? So that may be something for everyone to ask. So comment here on the on the video.

What is it what you like and subscribe. What does it say at the end of the video? Like, subscribe. Like and subscribe. Go ahead like and subscribe and, uh, we'll see you in a week. Thank you all. Thank you.