Transcription
AI Bob Ross almost conjures some kind of a feeling, but not quite. Okay, let's get another prompt going. What about a painting of a foggy forest at dawn, with a little cottage under a giant oak tree?
With generative AI, anything you can imagine, you can see. And the companies are working on what's called AGI, which will be able to do anything you can imagine, even manipulate every human being on Earth. Yes. On the eve of a dystopia we can scarcely understand or plan for. Coupled with a looming autocracy, it makes one long for the woods. Okay, now let's see what we got. Would you look at that? How about that? It almost conjures some kind of feeling, but not quite.
Well, that's it for today on the joy of AI. I get that so much where I'm like, wow, that's almost a feeling, but now that I think about it and look closer, it's just not. I'm sure that's literally the uncanny valley. And I'm I don't doubt these systems will go past it very soon. So, I'm not like a doubter. I just think we're in a weird spot right now where we're transitioning from like real things to AI things that are not quite.
Over the last 28 days, Chat GPT is now getting more website visits than X. We have officially seen the future of 3D robotic girlfriends in VR. It's so nice out here. We got to go. It's time to go. We looking weird here in the park. This is too weird. No, you look weird. Come on, stop acting all dramatic. I I don't know. It's just something I mean, I know it's a comedy skit, but it did kind of just make me think about what it will be like both when we're talking to our language models and they're extremely personable and we're walking around with like VR and those women were talking back to him in such a pointless kind of human power dynamic way and it just I don't know, it made me laugh.
I honestly think when people have the choice to turn their LLM into something that kind of like fights back with them and teases them a little bit, a lot of people are going to do it. Like there might be an era where we just get sick of how like responsive and helpful they are and they're like these AIs in the future might just push us around. It'll be kind of wild. I told you not to eat the mushrooms in my bedroom drawer. I was very excited. Oo, the faces in the clouds. God, MIT came up with a way to 3D print robotic fingers or curling tentacles. There's a new system for actually quantifying what a language model's personality is all about. Yeah, it's not vibing. It's linguistic analysis.
AI has now officially outsmarted virus experts, which of course means that biohazard fears are on the rise. Dwaresh Patel dives in with some researchers on really how the AGI curve is going to look. I know this sounds crazy because if you read our document, all sorts of bizarre things happen. It's probably the weirdest couple of years that have ever been, but we're trying to take almost in some sense a conservative position where the trends don't change. Um, nobody does an insane thing. Nothing that we have no evidence to think will happen happens. And the way that the AI intelligence explosion dynamics work are just so weird that in order to have nothing happen, you need to have a lot of crazy things happen. They literally break down every month from now until the 2027 intelligence explosion. Transformer models show surprising parallels to human thinking. We're going to look into that.
All right, so check it out. What do you think? Okay, so if you're trying to get all super vibe, this definitely works. I've got brightness setting here on my phone. Go from dim to bright. The coolest is probably the hue, meaning I can just choose any color that fits my vibe. I think I'll go for kind of a a blue. I don't know if you guys know I'm mostly red green color blind, so I like to stay in like the blues and yellows, but I think that's pretty nice. Something something like that looks good. And then saturation if you want to be more of like a monochrome black and white look or just like really go viby colors. It's got scenes. This is the party scene. I mean, I could imagine it. The chill scene, bedtime. Yeah, I can see that. a little bit more relaxing when you're trying to, you know, actually get some sleep and then awaken. Oh, that's interesting. Maybe like a good like 20 minute like morning glow, wake up a little natural. I like that. So, check it out. floraplant.com is the website. I appreciate them sending this to me. It's going to go right here in the background and you're going to see it hopefully in a bunch of new videos in the future. So, I just I think it looks very cool and vibe. So, thanks for sending it. All right, so floorpl.com. That's where you can pick up your own. I'm really excited. It's going to be in the background of the videos. Yeah. I mean, I chose it. I saw it on Instagram. I was like, I think I want to try that as a background for a little while. So, I think it'll look good. All right. What do you think? No maintenance, all beauty. It's got a little bit of the avatar look.
In the ongoing Elon Musk and Sam Altman feud, Sam seems to have one new thing to hold over his head, and that is how many people visit chat GPT. In the last 28 days, 4.78 billion people went to Chat GPT, 4.02 billion visited X. Kind of a weird chart to look at. That's like week over week. So, you can see how people just don't use chat GPT that much on the weekend. So some days this month, almost always on a Sunday, it was more popular to be on X than it was on chat GPT. But overall, it's a bit of a paradigm shift when you start thinking about people interacting with large language models more than they are social media. And you can see it's really a story of 2025. I mean, just slowly but surely that uh moving average is moving up and up and up and uh just kind of staying the same on X. And one person asked a clarifying question about maybe it's because more people are on the web versus mobile and chat GPT as an app actually did surpass X also in daily active users. Yeah, I don't know. I don't know exactly what it means, but it I would just generally say in my personal habits, I am opening up chat GPT and spending more time than I am on social networks. Except now that I think about it, probably not because I do spend a decent amount of time on those stupid Instagram reels. It's weird. I I intentionally go to chat GPT much more. It feels like I use it more. It feels like a tool that I'm using all the time. I probably like per minute, the amount of minutes I spend on it probably isn't as much as I spend on Instagram. I don't know. But I also make social media for a living. I don't know. I don't know if that changes anything, but well, not for a living yet, but I'm trying to get there. If you want to hit that Patreon button or that join button, that would help. No bigs, though. If not, I'm just going to keep doing my thing.
All right, so you can now 3D print things that fit together in extraordinary ways. Thanks to MIT Research, they have developed X-ring. This is an all-in-one 3D printed method that simplifies the creation of cabled driven mechanisms. And look at the kind of cool things you can build with it. You know, traditionally complex and time consuming to assemble by hand, these devices can now be automatically printed in just one step with embedded cables. I think this might be kind of a big thing. Imagine artificial intelligence making the most unique shapes for extremely precise environments, whether that's like manufacturing or around the home. And then a 3D printer can just take that information and print it out. It's It's going to be crazy. Look at that. dynamic objects like walking lizard robots, a peacock-like sculpture, and a gripping claw. X-tring reduces production time by 40%.
Yeah, let's talk about a new system that quantifies language model personalities. The new system is officially called the language model linguistic personality assessment, the LMLPA, and it can actually measure the personality traits of a large language model. So, you might have heard of the big five personality traits for people, which is, I think, more scientific. There's also Myers-Briggs, which I think is like a little bit shady, a little bit skeptical about how accurate those are, but there is a lot of data around them, so they might fit some patterns. Anyhow, this system that they're building is for large language models specifically, and it is an adapted version of the big five personality framework. And then it goes along with another model, which is like part of the whole package, which is an AI based raider to turn the model's text outputs into numerical personality scores. And it's kind of an important thing. I mean, this is the sort of technology which really isn't something that is needed in the current like landscape of large language models. But when it starts to become tools that are taking a lot of actions on our behalf, maybe even like humanoid robots in our house, we want to align that AI behavior with human values and make their interactions more natural and tailored. And this is the kind of system where we're going to be like, "No, I I don't want someone on that part of the big five." Let me see if we can pull up something. Yeah, I always think about something like this, one of these like wheels, the big five wheels. And you might want a robot that's more like of an extrovert, but like leans a little bit towards agreeableness or something like that. And it's really cool how this project actually bridges the whole idea of an AI human personality and a subjective quantity like these AI large language models and tries to merge them into something that we can use.
All right, now let's talk about one of the corners of the AI industry that kind of took me off guard, right? Like I was a little bit surprised when people started training AIs to do geometry and math equations, but I kind of got used to it and then I forgot about virologists and now I'm having one of those same shock moments where AI has officially outsmarted virus experts in a lab and it is raising biohazard fears. Okay, so models like chat GPT and Claude are now outperforming PhD level virologists. We're talking about the virologists that are working in a wet lab, problem solving, making decisions, predicting how viruses are going to spread, how they're going to work, what is coming next. This is real just virologist wet lab work. And yeah, Cloud and Chad GBT are just just smart. Really freaking smart. So, the good news is it could revolutionize how we fight disease and accelerate vaccine developments. The downside is that you can be very concerned about new viruses that can be created probably fairly soon. So the same AI that helps scientists solve complex lab issues can potentially guide someone with no formal training through the steps of creating a bioweapon. Experts are urging AI companies to implement stricter access controls and safeguards, while some are calling for government regulation to make sure that these tools don't fall into the wrong hands. And by the way, in a couple categories, the systems did way better than people. So, OpenAI's 03 model actually scored nearly twice as high as expert virologists on tough lab troubleshooting tasks. Yeah. Which kind of shows that AIs aren't just good at academic knowledge anymore at some of these hands-on lab skills that they're they're actually guiding people through. And look, if you're trying to build an AI powered wet lab, I probably wouldn't go with Google's Gemini. You can see it scored a measly 37.6 where you can get 43.8 accuracy out of Open AI's 03 model. So yeah, just a little tip for you there if you run a wet lab in your free time, which I really hope you don't.
All right, next up, let's jump over to the video Daresh posted. It's called 2027 Intelligence Explosion month-by-month walkthrough. And it made me do a couple things. you know, I've uh kind of been like less of the open-source AI model guy, and they really break down how they kind of thought the same thing, and then why they came to a conclusion that it's maybe not what it the way I thought it was going to play out. And that is kind of making me rethink pulling back. I wouldn't necessarily say I'm more in favor of open sourcing these models right away because bad actors can get them. I do think I've lost sort of more faith in the idea that if you keep it private, the people who have it in control will use that edge to do the right thing, which will amplify. So, if they have a three-month edge where they have AGI and the rest of the world doesn't, they can use that to actually build something in a responsible way before bad actors can. They just get that and they use it to just make themselves richer. And that seems to be like internally what a lot of people like who are in these situations want. So there's a ton of great stuff in this podcast, but right here around what is it 151? Let's watch some of this. Yeah. Your thoughts on why transparency through this period is important. Yeah. Alo community there's been this idea which I myself used to believe that like it's an incredibly high priority to basically have way better information security and like if you're going to be trying to build AGI you should be like not publishing your research because that helps other less responsible actors build AGI and the whole game plan is for like a responsible actor to get to AGI first and then stop and burn down their lead time over everybody else and spend that lead on making it safe and then proceed. And so if you're like publishing all your research, then there's less lead time because your competitors are going to be close behind you. Um, so and other reasons too, but that's like one reason why I think historically people such as myself have been like pro uh pro-secrecy even. Another reason, of course, is obviously you don't want rivals to be stealing your stuff. Um, but I think that um I've now become somewhat disillusioned and think that even if we do have like, you know, a three-month lead, a six-month lead between like the leading US project and any serious competitor, it's not at all foregone conclusion that they will burn that lead for good purposes either for safety or for constitution of power stuff. I think the default outcome is that they just, you know, smoothly continue on without like any serious refocusing. Um, and part of why I think this is because this is what a lot of the people at the company seem to be planning and saying they're going to do. I guess he kind of convinced me to think a little bit different just because it seemed like every word out of his mouth mimicked what I would have been saying for a while. And then him having more information to like what really would happen with that three-month lead because he knows maybe more about what's happening behind closed doors or how people who might have that power would think. It's makes me a little less confident that that's a good move anymore.
So anyhow, let's go talk about transformer models. There's some really surprising parallels to how humans think that can be found in the way that the transformer model works. So some researchers asked this question: Do the internal computations of these models resemble how humans process language and vision in real time? And they found transformer models like GPT and the vision transformers, they don't just mimic humanlike answers. They also seem to think, yeah, we kind of put that in asterisk. They seem to think in ways that resemble how humans process information. Okay, so we went over some of Anthropic's findings last week and I was totally blown away and I'm still trying to put this together how there is some kind of actual sort of physical shape, some kind of plumbing that seems like all things that have intelligence kind of hone in on even if they come from language models or they're in human brains. But these researchers, they were tracking how model predictions evolve layer by layer during tasks like fact recall, logical puzzles, and image recognition. And the team showed that the model's inner workings line up surprisingly well with human behavior. Things like reaction time, and even mouse movements means that these models arrive at the same answers, they might not be so different from how they end up constructing the dimensionality inside of their head. Yeah. The most interesting part was how the researchers actually found that the systems also did like what feels very human is to have like the wrong intuitive answer first and then logically you see relative confidence between the correct and an intuitive but wrong answer and boosting the correct answer over the intuitive one. So in a nutshell there's this two-stage process that's just built into the way that everything comes together in these models and it seems very human. It mirrors how people often rethink and override gut reactions with more deliberate reasoning, suggesting that a deeper similarity between artificial and human cognition than previously thought.
All right, next let's talk about some fascinating research that really kind of makes you wonder what it is inside of a large language model that's learned versus what data it's learned from. So, let me break this down. To make language models work better, researchers have learned to sidestep language. We insist that large language models repeatedly translate their mathematical processes into words. And there might be a better way. And actually, this kind of makes sense to me. Like we've had some questions about how to break words into tokens, right? Like sometimes they're words, sometimes they're word phrase phrases. And then we also see large language models moving to things like calculators to solve problems, right? like they might solve a simple multiplication problem just using like GPT40 or something but if it's bigger than that it might you know lean on Python it might go to a calculator to actually do it for it which is kind of wild but that conversion from the latent space the sort of plumbing of the model that actually can do some level of computation but starts to kind of degrade as it gets more and more complicated that process of translation can be clunky it can be inefficient and there's some new research arch that suggests that large language models can think more efficiently if they stay in the mathematical latent space longer before they actually generate any of the text. So when we ask for models to do chain of thought reasoning and it's like you know think about your answer does it make sense you're kind of saying why don't you move that into the math like this isn't really good for explanability which is one of the most important things if you let let it stay in the number space totally foreign to the way our brains work it can be better before it translates into English for us and there's these new models that they're playing with that actually skip unnecessary translations into words and reasoning and they just directly reason through numbers. And one of the more interesting insights from this paper is that when models avoid constantly converting their thoughts into language, they not only become more efficient, but they also become more accurate. So this model called coconut, which works this way, it actually uses one-tenth the number of tokens to achieve nearly identical results as its conventional counterpart. It can even in that space reason about when it should stop reasoning. So it can reason longer and get better at the answer, but also get to a point where it's like diminishing returns. Probably time now to stop, translate it into English, and then, you know, give the answer that we can understand. But it's wild to think that you can make large language models better by sidestepping language.
All right, there's a new brain-inspired AI technique that mimics human visual processing and it is enhancing machine vision. Let's dive into it. Okay, so first let's start with the idea of sort of a convolutional neural network. this idea that in a visual system you can take an image and you break it into pieces. Let's say pixels, little squares, little grids, and you can have a little window and you can like go over each one and you can start to learn what the patterns are. Then you can like stack these on top of each other and you end up with sort of patterns that emerge that end up looking like things you can describe like edges and lines and depth and colors. And that doesn't look too much like how the human brain actually sees. However, when you start to tweak it so that it actually has some kind of a priority to what it's looking at. Technically, they're teaching it to adapt its shape to focus, like what window it's wanting to learn from. Learning how to focus on the most important parts of an image. So if you have an image of a cat, like where the eyes are, where the ears are might be much more important than how many leaves are on the ground in the setting. And the, you know, we know that as people, like we look for the kind of thing that looks more biological because it could, if it's a snake or something, it could jump at us. If it's a car, it can move. If it's a person, we want to know who it is. We don't quickly and instantly just take everything on average. And they're finding that the more that they make AI convolutional systems, these vision inspired systems actually work that way, the better they can get at figuring out what the priority is, the more that it looks like the way the human visual system processes things behind the scenes. So they say in the way that it processes images with priorities begins to match how real neurons work in the brain's visual system, which could be a big step towards making AI more natural and smarter in the future.
Okay, so Clayton Ramsay wrote this. I picked it up from what was it? Hacker News. I think he said, "I'd rather read the prompt." And I thought it was worth sharing. All right, so he's a teacher and when he grades student assignments, they sometimes look like this, right? And it's just super obvious to me and you that like this is AI-driven because kids just don't, you know, students just don't naturally write with bullet points and like this whole kind of like repeat the question in the answer sort of style. And he says, "As an instructor, I'm always saddened to read this kind of stuff." And he writes this article as a plea to everyone, not just his students, but the blog posters, the Reddit commenters, all of those people. Don't let a computer write for you. And he doesn't say it because of intellectual honesty or for the spirit of fairness. He says it because he believes that your original thoughts that come out of the human brain, you as a creative human being are more interesting, meaningful, and valuable than whatever a large language model can transform them into. So basically, there could be places where we need AI to make decisions and to do all sorts of types of thinking, but there's still some places where we don't want to just be entertained or be thinking about creativity from a system, right? So he argues that writing should always be an expression of one's own thinking, no matter how basic or flawed and that delegating it to a machine defeats the purpose of communicating. And one of his more punchy points here is that he actually compares AI generated work to plagiarism. And he claims that what the AI model does is worse. He thinks you're better off cutting and pasting from someone else's human work than you are generating it. Because while plagiarism at least passes on a human's original thought, AI output lacks any real perspective or meaning, it's not just that it's written bad, it's fundamentally written by something with no soul. Right? And even though I'm I tend to think these things might have souls or something one day or or something much more complicated and much more conscious and humanlike, even if they don't right now, it's really interesting the conversation of not just using this stuff for academic dishonesty, but a deeper more existential issue which is like what does it mean when there is no human like creating or writing for us anymore. He argues that when we let machines speak for us, we stop showing up as humans.
All right, Jpool writes, "Beyond control, why the future of AI depends on our values, not on our fears." So, this is a response to Moadot. Let's break it down. So, using frameworks like the core value framework and the contemplator framework, they propose that we can help AI reflect, grow, and develop strong character. You know, I mean, we talked about this like it's like the baby alien thing. Like it's about raising AI systems with care.
And integrity, not just keeping them under lock and key, not just forcing them to do like in a military setting, like bad stuff. So we got to start thinking about not controlling these systems, right? It's not about like controlling what Meta is putting out. It's not about controlling Google's lead in coding. It's about building character and values into them so that they can become trustworthy.
And to do that, guess who has to be the role model? The human has to be the good role model. And that's that's kind of the biggest problem. But look, we have world-changing power on the line here. Like, come on, world. Don't you think we could just get it together? Like, don't teach these things to just be terrible and make money. Like, we have one shot to create a super intelligence. It could give us a Star Trek universe where everybody gets replicators and holodexs and like, and health. Well, you know what I mean? Like, you're going to try to put that on some military thing so that you can like beat your competitor when like everybody on Earth literally could benefit from this just fine if we do it right. Stupid. It's just stupid.
If you want to support the channel, go to patreon.com. I would love to get that 91 members up to 100. Also, I wanted to say thanks for everybody for watching the Wes Roth video. I mean, I know he like put it on his main channel, which is which helped boost it a lot, but man, 10,000 views. It was a great conversation. It was really fun interviewing him. You can see on my analytics, we even got 200 new subscribers. Leave a comment if you came from that video. Might be one of the new person. This might actually this what? This might be the first or second video that you've seen since then. So, the version of the thumbnail that said, "Can we trust AI" only got 28% versus the "the AI lie you believe" got 36%. Also, I got a lot, 100 comments on this post. So, I got a lot of traction on this. This is what YouTube says the other videos that you watch are when you're not watching Dylan Curious. And Wes was at the top, so it was really cool to be able to interview him.
Maybe I can get an interview with Matthew Berman coming out here soon. We're going to see. Fingers crossed. Of course, David and Matt Wolf have already been here before. And then I just don't I don't really have like a contact at the AI grid, but I'm going to work on that one, too. I would love to do an interview with all these guys. All right, let's take uh all this AI stuff one day at a time, and I will see you in the next video.