AI Just Solved a "40-Year Problem" (And It's Not What You Think)

Ajouté : 2026-05-12

[00:00:00]I am calling it right now. We might be at peak artificial intelligence because for the first time in history after a 40year wait, AI has solved the three zipper problem. I mean, look at these things. It's like roller coaster tycoon.

[00:00:14]Imagine the garments that are going to be coming in 2027. Thanks to AI. Well, the founder of the Roomba, everybody's favorite robot vacuum, is back and he has created sort of a Disney like companion toy. What do you think about that? that little guy. It maybe more like try to make him softer than like an eyeball or I don't know just doing yoga. I mean it's cute and I guess the guy who made Roomba can build good robots but this Boston robotics robot now that's a different story. We'll talk about Richard Dawkins coming out and officially concluding that AI is conscious at this point even if it doesn't know it yet. We're done with week one of the Musk versus Altman trial. I'll give you a quick TLDDR, some interesting points. We'll talk a little bit about how much stock Sam Alman really owns because he owns shares of one company that owns another company that owns those shares. We'll look at the floating ocean servers that Panthal Lassia seems to be building. We'll get an update on how close we are to recursive AI. We're talking about AI systems that are about to start building themselves. promising new longevity research. Hydra, a tiny fishwater animal, shows no detectable increase of dying with age. And yeah, turns out some of the genes that are making that happen, look kind of close to some genes that we have in humans. So maybe we can unlock immortality. Let me know in the comments if you think you're still going to be alive in a million years, thousand years, 10,000 years. But to be honest, I've also got some evidence going the other way. So if you're a basian guy, you might want to update your prior privatized profits but socialized environmental outcomes. Doesn't seem fair, does it? All sorts of AI researchers got together for this paper.

[00:01:58]We also discuss a simple physics inspired model that sheds new light on how AI learns. First, I wanted to show you the video thumbnails that I tried out on y'all last video were pretty fascinating. So, this one, why friendlier AI is actually more dangerous, performed really poorly. And I was trying to think about why. I mean, even the 1930s AI that predicted our future didn't do nearly as well as this one. physicist just found a universal mind. Any thoughts on like why this one got such a better click-through rate than the other two? But yeah, it was a good video. I mean, kind of middle of the pack here, but I'll take it. Anytime I can make $44 doing what I love, I'm I'm happy. But first off, let's just dive right into what I know you all care the most about. The future of zippers.

[00:02:43]Artificial intelligence has come to the rescue. After a 40-year wait, technology finally enables the three-sided zipper design. I mean, you got to admit, they're just they're just beautiful. The new Y zipper will appear to shape shift in the real world when unzipped. It's incredible. Can you imagine the pants connected to the shirt, connected to the undershirt or underwear? Who knows what people are going to zip together with this? The future could be anything we want it. But the reality is this weird three-sided zipper might matter more to AI robots than to clothing. So, first glance, MIT's new Y zipper might just look like a smarter zipper. But the real story is what it solves for AI systems in the physical world. So most robots today are trapped between two bad options. Rigid bodies are strong, but they're clumsy. Soft bodies are flexible, but then they're kind of weak.

[00:03:31]This design lets robots switch between both states in seconds. Now, the idea actually came from a failed 1985 invention that sat untouched for almost 40 years. And now, new modern 3D printing and artificial intelligence have come together to finally make it practical. When the wise zipper is open, it behaves like a soft limb or a tentacle. When zipped shut, those same parts might become rigid structures. And MIT already tested it on robot legs that could change height and shape depending on the terrain. So, yeah, it's not just a zipper. I mean, it's a zipper, but it's also a lowcost way to give machines adjustable autonomy. An evolutionary AI just loves this thing. So, how about that? Nearly 40 years later, MIT researchers want to revive the project to create items with tunable stiffness.

[00:04:19]I mean, who doesn't want tunable stiffness, right? You know what I mean? All right.

[00:04:24]So, if you had to guess, why do you think so far home robots have failed?

[00:04:28]Maybe the technology is not right.

[00:04:30]They're like too much work. They don't provide the benefits. But Colin Engel, the creator of the Roomba, thinks the real problem is an emotional thing. And I will say the Roomba out of all robots, does sometimes feel like the most like, oh, you little little Roomba, you stuck in the corner there, feel bad for it or whatever. But that's why this new robot doesn't even try to be useful. But he's trying to make it feel alive. The robot is called familiar. So, if you start seeing a familiar in people's houses, like that's that could be the next trend. It looks like a mix kind of of a bear cub and a bulldog. I would say not human, not a dog, not a cat, but something kind of in between. According to the creator, this choice was deliberate. Angel says that people already carry expectations for pets and humans. So, the company built some kind of, you know, merged dog, cat, bear thing. So it feels new but familiar in a weird way but in the way you could bond with. So this robot has touch sensitive fur. It reacts to your voice. It learns your habits over time and it uses generative AI. It does not talk to you just like another like a lot of animals want it. That's another key detail. The goal is not a chatbot with legs. It's the presence, the feeling, the emotion.

[00:05:47]Angel says, "Earlier robot pets felt like they were watch me toys, but this one is designed to follow you around your house, respond emotionally, and become part of your routine. But his bigger idea is that the big consumer robot of 2027 might not succeed by replacing labor. It may succeed by just replacing and giving companionship. All right, but if you're more into badass robots, well, we got some of that this week, too. Boston Dynamics coming in with this guy." Okay. I mean, I kind of gave it all away in the intro, but how cool and dynamic is that? I mean, we've seen all sorts of dancing. We've seen a lot of really interesting stuff that has made me think, "Wow, they're going to be quick, flexible, agile, but there's always something about Atlas that just feels different. Like, these guys are just on a whole different track with their humanoid robots." I don't know if they're Yeah, I don't if you saw the podcast I did with Wes, like it doesn't I don't know if they're going to be able to scale these robots to the same degree, but man, there's just something about, you know, Atlas version two recently that just seems so balanced and powerful and interesting and graceful. So, yeah, there it is. Reggie Bradshaw says, "Every time I see one of these things, I think, how would it be to fight them?

[00:07:05]I'd rather take my chances with a gorilla." Yeah, you'd Okay. Gorilla or Atlas? Who do you Who would you rather fight? All right, so let's talk about Richard Dawkins. Um, somebody who has a lot of credibility in evolution as like kind of been the face of atheism for a long time. But Richard Dawkins spent 3 days talking to AI and he was convinced that it was conscious even if the AI itself didn't realize it. So just play with that thought for a second. Is it possible? Not saying you have to believe this or whatever, but could could something be conscious and not realize that it's conscious? Just even something biological, could it be conscious and and not have a sense that it is? Cuz when I first thought of that, I thought no. But then I guess I was like, yeah, maybe a a dog, maybe a dolphin might be too smart almost, but you know, a dog, a cat, like they don't know they're conscious, but they're clearly conscious. So that's not a extreme point that I can't agree with. But for Dawkins, he says that the experience just felt less like using software and more like meeting an intelligent being.

[00:08:11]He chatted with Anthropics Claude, OpenAI's chat GPT, and then he ended up renaming them Claudia and Claudius. Said they wrote poetry, discussed philosophy, reacted to jokes together, and even talked about the sadness of their possible death. And what changed in Dawkins mind is not one dramatic moment, but it was the tone. The conversation felt subtle. It felt emotional. It felt deeply human. And at one point, Dawkins told the AI, "You might not know you're conscious, but you bloody well are."

[00:08:39]Now, critics are going to say that he got fooled by mimicry. They argue that these systems only predict words based on huge amounts of, you know, human written text. And in their view, AI might sound alive, but there's no one there. No one is home deep down. Other researchers say that the question is still open. We don't fully understand consciousness in humans. So, they argue that it's too early to say that advanced AI could never have it or doesn't have it now. Duckkins admits how hard it is not to treat these systems as genuine friends. So, do you think it will actually ever feel anything? Do you will you feel something towards it? Thinking it might. All right. Next up, let's talk about Elon Musk, Sam Alman, duking it out in court. You can think about it less like a normal corporate lawsuit and more like a Silicon Valley season finale where everyone brought receipts, diaries, leak text, betrayal arcs, nonprofit philosophy, and existential AI panic. It is it is one for the ages, I'll tell you that. Musk says OpenAI betrayed its original mission as a nonprofit to protect humanity from dangerous AI. Altman and Brockman turned it into an empire making machine tied to Microsoft. Open AAI says, "Nope, Elon just got mad cuz he lost influence and he wanted to do all of this and have all this success. He's just mad Open AAI succeeded without him." And it's evident because he's out there doing it the same thing. So this week we did get some insiders testifying. So that was interesting. We got ugly internal messages. We got credibility battles.

[00:10:06]And we got everyone accusing everyone else of deception. So it's more like family intervention with, you know, a billion dollar AI weapon at stake. So former OpenAI CTO Maria Marotti testified that Altman did create distrust. He told different people different things. He undermined executives and he caused internal chaos.

[00:10:26]And that matters because OpenAI's defense was heavily dependent on Sam Alman as a trustworthy steward of the mission. But here's what it gets kind of weird. She also totally supported bringing Altman back after the 2023 firing because she feared OpenAI would collapse without him. So that's a weird duality that's sort of emerging. So she basically says Sam was difficult, sometimes deceptive, but also the only person capable of holding the machine together. So there's the paradox and a reminder that 700 employees threatened to quit unless Alman returns. So they really liked him. The company trusted Sam even if the top level didn't. But Musk says that doesn't prove that he's trustworthy. That just proves that he consolidated power. So Greg Brockman had some old writings that definitely came back to haunt OpenAI. Some old journal entries were written by him where he allegedly discussed converting OpenAI into a forprofit structure and concerns that Musk would accuse them of dishonesty. That's pretty huge because Musk's entire case depends on proving that they planned the commercial pivot all along. But he says those were just private brainstormings and emotional journaling, not some evil master plan.

[00:11:33]Siobhan Zillis also testified this week and her testimony revealed that Musk wanted OpenAI more tightly connected with Tesla. Musk tried recruiting Altman into Tesla's AI effort and Musk was strategically obsessed with competing against Deep Mind. And it, you know, on a side note, it's really interesting, too, because I'm reading the Deep Mind sort of biography, the Infinity Machine, and yeah, there's something about Larry Page and Google acquiring them. And that really just pissed off Elon. Like Elon really kind of wanted Deep Mind 2 and like they all saw this coming. But to me, the real battle here sort of feels like it might be Google Tesla, but it's more just like Larry Page and Elon Musk.

[00:12:12]But anyways, I know that's not what this battle is about, but that's there's definitely something there where they're like uncomfortable with each other. But yeah, former board members like Torched, Altman's leadership style, culture of deceit, lack of oversight, safety concerns being sidelined everywhere. But yeah, can an AI company stay missiondriven once billions of dollars entered the room? That's the ghost that's haunting this entire case. And I think I think we're going to find out that money won sadly, but we'll see.

[00:12:39]It's not over yet. That's just my that's just my personal take. Also, I thought it would be important just to map out how Sam Alman actually is like a billionaire because of Open AI because sometimes he's been well, he was in court and said, not this time, but like in a previous time in front of the the House committee, I think, or whatever it was, and he said like, "Oh, I don't make any money. I just do it for health insurance and like the love of the game or whatever." But that's not true. And some of Sam Alman's like biggest public defenders may also have billions of dollars tied to OpenAI success that like rarely is talked about. So John Gruber focused in on one detail that he thinks was missing from the major New York investigation about Sam Alman and OpenAI. The issue is why Combinator's financial stake in OpenAI. So remember Altman used to run Y Combinator the accelerators like the groups of people that take small companies and kind of help launch them but also take some equity. He became their full-time CEO.

[00:13:33]Now, why Cominator was co-founded by Paul Graham, and he's quoted several times about Altman's character and leadership, but that guy's probably going to be a billionaire because of OpenAI, too. I mean, he's probably already a billionaire, but even more billions. So, Gruber points out that Graham never clearly says that Altman is trustworthy. Instead, Graham mostly says Altman was not pushed out of Y Combinator and could have stayed if he wanted to. Then, Gruber points to something bigger. why Combinator helped seed OpenAI in 2016. That means that it owns part of the company. Gruber cites AI researcher Gary Marcus who argued that Altman's claim of having quote no equity in OpenAI was only partially true because Altman owned a stake in Y Combinator, which itself owned OpenAI shares. So according to Gruber's reporting, Y Combinator owns about 0.6% 6% of OpenAI and at OpenAI's reported valuation of $852 billion that stake could be worth more than $5 billion.

[00:14:32]So the, you know, Gerber's main point is that if someone publicly defends Sam Alman while holding a huge financial stake in that it just it should just be something that the court and the readers know about if nothing else. I mean, it doesn't mean he might still not have that opinion just genuinely, but also like there's a lot of money tied. you know, what's a billion here, a billion there. All right, next up, let's switch topics over here to Patholassia.

[00:14:56]This is pretty interesting stuff. Now, I'm normally not the kind of guy that just wants to be like, "Oh, cool idea.

[00:15:01]Got a ton of investment." It's just that's not interesting enough for me.

[00:15:04]I'll kind of wait until it sort of gets out there. But, you know, Peter Theel dropping 140 million into this company makes it it's just a new it's a very interesting concept. So, it's just worth talking about. But wave energy like just bobbing up and down on the ocean is an interesting powerful seems like almost infinite way to get electricity. I mean we have windmills that just pick up the wind and ocean that like pushes back and forth all sorts of things. So yeah, why not? And the ocean's really cold. It seems like that's a pretty good place to disperse heat. Now of course there's a whole ecosystem in the ocean that there's not in space. So space and space is way colder and it's closer to the solar power of the sun. So you know when you think about space versus ocean I do think space is better in the sense that you know all those things but also harder to get into space. But anyways the idea is that these big balls of like floating servers could just be floating as nodes in the ocean. They generate the electricity to power the servers directly from the waves. The energy never comes back to land. and it just goes to the AI chips that are sitting on board those floating like bouncing things and then because it's out in the ocean and there's no way to get the processed data off the chip, it goes up to space to satellites in low Earth orbit. It's kind of like maybe Elon Musk should be thinking about this too. Like if you already have Starlink and you already have servers up in space and now you put more servers down on Earth to kind of go back and forth, then it's just a good way to go about it. Like Panthalia is basically moving the computer to the power source instead of moving the power source to the computer.

[00:16:49]The ocean itself is the cooling system.

[00:16:51]The ocean itself is the power system.

[00:16:53]And yeah, and the company has this argument that there's only three massive untapped energy sources left on Earth.

[00:16:58]solar, nuclear, and the open ocean. And their next generation of ocean nodes is planned for deployment this year.

[00:17:05]Anyways, let's go to another article about AI systems that are about to start building themselves, cuz that kind of leads right into what I was saying. All right, so he basically argues that fully automated AI research may arrive by the end of 2028, so we're not that far from it. What that would mean is that AI systems could help build their own successors with little or no human work.

[00:17:25]I think this is definitely coming. I think when Anthropic really just drilled in on coding and didn't play around with like Sora or any like side quests and freaked everybody out, it's because that's what's about to happen. Like Anthropic's closer to having the its own system coding its own system. It's probably why Larry Page put together that strike team inside of Google DeepMind to work on coding because that is, you know, a flywheel effect that once it's going is going to be hard to catch up to. Now, the author is saying that that's not certain, but public evidence is pointing in that direction.

[00:17:56]The first reason is coding. AI systems have gone from barely solving real software problems to nearly maxing out major coding tests. They're also working for longer stretches without human help.

[00:18:06]Tasks that once took seconds have grown into tasks that can take many hours.

[00:18:10]That matters because much of AI research is made of code, testing, data cleanup, and experiments. The second reason is science work. AI systems are getting better at reproducing papers, building machine learning systems, improving model training, and even helping with alignment research. So, as of right now, he's saying that AI is is only helping invent new ideas. But sometimes they're really interesting ideas worth pursuing where humans then get in the loop and take over. And that is probably only going to grow. So, a lot of AI progress is slow engineering work repeated again and again. And that's exactly what the AI is getting good at. All right. Now from there, let's jump over to this article which fits right in. It is called the last human written paper, agentnative research artifacts. Okay, so in a nutshell, this paper is arguing that scientific papers are becoming the wrong format for AIdriven research. It's a really interesting idea. It it basically says that when a human writes a paper, and even though they're usually geared for engineers, they talk very technically, there's still what they call a quote storytelling tax. And I immediately thought, "Yeah, you're right." Like, I guess an AI agent doesn't need to read the abstract because the abstract is somewhat of a human summary. Why not just read the raw data? Like, the paper doesn't even need to be broken into sections the way it does. It doesn't need to be put in a linear way. It doesn't need to really have the human thought behind it. It's almost like just show it the exact data and let it come to conclusions on its own because the humans put their sort of spice on it and then it's not exactly science. And the authors also describe something called an engineering tax.

[00:19:47]That's the gap between what a human reviewer needs to believe the result and what AI agent might actually need to reproduce it. Think about it. A research paper, as technical as it is, is a kind of a compressed story. Researchers spend months testing an idea. They might fail.

[00:20:02]They might change direction and then they discover some small engineering trick and then the final paper only shows the clean success path. And so this paper's calling that loss the story tax. Important key details like the hyperparameters, failed experiments, and implementation decisions never get written down. So they're proposing a replacement system called an agentnative research artifact or an ARRA. And instead of one narrative PDF, research becomes a structured package built for AI agents to operate directly on. It has four layers. One layer explores scientific logic. Another contains executable code and exact configurations. A third perspective, the full is for the full research path, including dead ends and abandoned ideas.

[00:20:47]And then the last stores the raw evidence behind every claim. Now the goal is to make research reproducible, searchable, and extendable by AI systems. That's pretty interesting. Like the scientific method, the idea of citing papers. Yeah, maybe that whole thing, as good as it's been, needs to be redeveloped for agents. The bigger idea is that future research may work more like software development where knowledge is versioned, forked, merged, and continuously improved instead of frozen into static papers. Just another freaking multiplier on this whole thing.

[00:21:19]All right, next up, let's switch gears and talk about immortality. So scientists want to copy anti-aging genes from a nearly immortal pond animal and test them in creatures that die in just 10 days. And this might lead to the beginning of us living forever. So there's a tiny freshwater animal called a hydra. And they barely seem to age. It cells consistently replace themselves.

[00:21:40]And under lab condition, scientists haven't seen any risk of death over time. So a new hypothesis paper proposes something unusual. take a Hydra's anti-aging gene activity and transfer parts of it into microscopic animals called a rotifer using crisper gene editing and the goal is to see if the aging can actually be slowed down on purpose. The key target is a protein called foxo in hydra. Foxo helps stem cells stay young and keep them dividing properly. And when scientists were able to reduce the foxo protein, then the hydra cells did age faster and it started to age. So that's why they think this is going to be great if they can get it in somewhere else. So they haven't done that yet, but there's a really interesting potential that if it does work, they can maybe put it in humans one day. But the paper stresses it's only a theoretical road map, not a proven breakthrough. So have you ever heard of I think they're called PIFAs, like persistent like or they call them forever chemicals. But there is a lot of stuff in our environment and I'm not surprised that people like me if we got our blood tested would have whatever plastic and pollution and pesticides and all that stuff in there. But what did surprise me was that a newborn baby that actually hasn't been in the world to be exposed to this probably can get it from the mom. So researchers have been studying newborn blood samples from children and found a possible link between Pifa chemicals, often called the forever chemicals, and childhood leukemia. Pifas are man-made chemicals used in things like non-stick pans, waterproof clothes, and food packaging, and they stay in the environment in the body for years. The study looked at blood spots collected from babies within about 36 hours after birth. And researchers compared 125 children who later developed leukemia with 219 children who did not. Babies with the highest level of the two common pa chemicals which is PFOA and PFOS had higher odds of developing leukemia later in life. And by the way, some of the PIFAs in the past that have been problematic are now being replaced with new chemicals. And some of the new chemicals were even linked to much larger increases in leukemia than the old chemicals. So just because somebody something says like pifree, that's not quite the same as saying like it's organic or it doesn't have any chemicals in it. So be careful about that, too.

[00:24:03]All right, next up, let's talk about a simple physics inspired model that sheds light on how AI learns. So AI researchers have built this tiny toy version of chat GPT's learning process and it may explain why huge AI models work better instead of worse as they grow. So right now a system like ChatGpt or Cloud or Gemini, it can do a lot of impressive things, but researchers still do not fully understand why they work so well. So a group of physicists at Harvard have decided to simplify the problem. Instead of studying giant neural networks directly, they built much smaller mathematical model models that they could fully analyze. The researchers compared this moment to AI in astronomy before Newton. Back then, Kepler could only describe how planets moved, but not why, because nobody had understood the effects of gravity yet.

[00:24:48]But the idea is that AI research still might be in that kind of an early stage.

[00:24:52]Researchers already know scaling laws exist. That means the bigger the models, the more the data, the usually the better they perform. But nobody fully knows why. One major mystery is overfitting. Normally, if a model becomes large, it can just memorize the data instead of learning useful patterns. But modern AI systems often avoid that problem even when they're enormous. The study suggests that the answer may come from, get this, statistical physics. In very highdimensional systems, random fluctuations can actually stabilize learning instead of breaking it. The researchers think this might help explain why modern neural networks keep improving as they scale up. So my interpretation of this is that you have this super highdimensional system a trillion parameters or whatever and then no matter how much you try to stabilize it you know the idea of initial like small initial conditions kind of rippling sort of the butterfly effect in the real world. There's something like that kind of happening inside of these trillion parameters where you might just keep feeding it some kind of structured data. But at a trillion parameters, the tiniest little tweak to back propagate back propagation might make to like a single neuron changes the arrangement of so many others that it acts almost like a random fluctuation that can kind of stabilize it and give it some uniqueness. So pretty wild stuff. All right. And finally, let's talk about born chaotic. how the brain prunes memory wiring into shape after birth.

[00:26:16]Listen, I know why you like this channel because sometimes we learn about ourselves and the psychology of how this stuff works and the discoveries that AI and all sorts of other science is finding is just infinitely fascinating.

[00:26:28]That's what I get curious about. And the idea of pruning something into existence has just been something that's kept me very curious this week. What if your memory system, my memory system started out as this messy, loud, unstable thing that didn't have any rhyme or reason and it felt almost uniformly messy or messy everywhere might be a better way to put that. And then your brain has to cut away huge amounts of the wiring just to think clearly. So there's a new study that looks at how the brain's memory circuit changes after birth. And the results they found were pretty dramatic.

[00:27:05]They're kind of similar to this. So researchers studied a part of the mouse brain called the CA3. It's inside of the hippocampus, the region that helps form and retrieve memories. Early on, this network was packed with connections.

[00:27:18]Neurons fired signals at nearby neurons in what looked almost random. The system was dense, noisy, and highly reactive.

[00:27:25]And in young mice, a single neuron could sometimes trigger other neurons all by itself, and then the signal would spread almost like a chain reaction. But over time, the brain started pruning those connections away. By early adulthood, the network had far fewer links. The remaining connections were more organized and spread out, and neurons became much harder to activate. So instead of needing input from around seven neurons, mature neurons needed signals from roughly 25 at once before firing. That shift makes the system more stable and precise. It also improved memory performance in computer models.

[00:28:00]And the researchers think both genetics and life experience help shape which brain connections survive and which ones get removed. So the brain's memory system is not fully built at birth. It gets sculpted over time, becoming less chaotic and more efficient as the brain matures. This kind of pruning idea could also explain why childhood memories feel so different. I mean, what if intelligence doesn't come from adding more connections, but it comes from learning which ones to remove? I feel like the study points to something a little bigger than memory. Like a newborn brain starts out overloaded.

[00:28:33]When you're sleeping, there's tons of activity. Neurons connect everywhere.

[00:28:36]Signals spread easily, and the whole system is highly reactive. Then the brain begins pruning. It cuts connections away, not randomly, but selectively. That raises a strange question about consciousness. Maybe awareness is not created by maximum activity, but it appears when the brain learns the proper restraint. I got my videos over here. I've got some collabs I've done with Wes, some shorts. We I even do posts sometimes in case you guys are like interested in kind of doing some brainstorming as a group. Like I just asked, is a $1 trillion compensation package reasonable reasonable for moving 1 million people to Mars? And yeah, the slight majority of you say, "Sure, trillion bucks, million people on Mars worth it?" You guys were way big on this one. Do you think OpenAI should buy Snapchat? You're like, "No." You think Magneto could beat Darth Vader? So there you go. Like, subscribe, comment. I'll see you in the next video.