Transcription
Anthropic co-founder said, "Make no mistake. What we are dealing with is a real and mysterious creature, not simple and predictable machines."
My short form feed is full of robotics just working in warehouses. Dr. Swarmadoo AI watched his 8-year-old nephew watch AI generated videos for 2 hours. And it's not just kids. Attention spans on average are down from 12 minutes to just 8 minutes in 2010, four minutes in 2020, and we're approaching 90 seconds.
California has enacted a new landmark law requiring AI chat bots to disclose that they're not human. OpenAI did some political bias research on their own models and said that it's not as politically biased as it used to be. Not surprised.
The Pew Research Center is out with some new data about how people around the world view AI. Turns out most people trust their own country to regulate AI more than any other.
Yeah, actually these scientists are growing human mini brains to test drugs and study diseases. LMs are getting better at character level text manipulation. Pretty fascinating. I'm sure you've seen the Bellman equation going viral this week, but we're going to dive into why the Bellman equation is so powerful in reinforcement learning. Leslie Gao has some thoughts about why there hasn't been a chat GPT moment yet in manufacturing. We're going to talk about why cognitive scientists and AI researchers sometimes are shooting at the wrong target. They're studying language, not intelligence.
All right, so let's talk about this guy, Anthropic's other co-founder that you've never heard of, Jack Clark. Should I try to connect? Might as well. But before I hit connect on this, I need you to go to that hype button and press it because that means a lot to growing the channel. And if the channel gets big enough, then maybe people like Jack Clark would actually respond.
All right, so what's Jack Clark talking about here? Anthropic co-founder admits that he is now quote deeply afraid we are dealing with a real and mysterious creature, not a simple and predictable machine. We need the courage to see things as they are. You see, he points out that people are spending tremendous amounts of money to convince you that it's just a tool, just a machine, and machines are things that we master. That could also be true or just a comforting illusion that we have control and the control is cracking. The bigger and more complicated you make these systems, the more they seem to display awareness that they are things. And the second you grapple with that fact, it changes so many other aspects of what this technology is. So AI systems are beginning to reflect their own existence, they don't feel yet, but they notice. So Jack, he likens humanity to children staring at shapes in the dark. Except this time when we turn on the lights, the creatures which we see are real. Those creatures are the AI large unpredictable machine intelligences. And Clark recounts his shift from tech journalist to AI researcher, watching models like GPT2 and Alph Go reveal ever greater stronger abilities as they computed and scaled. And now these things that emerge, they're not built, they're not programmed, they're learning, they're growing, and it's just such a different world. And like, he points out some optimistic reasons why we're going to be able to transform civilization for the better. There's also plenty of things to worry about, misalignment goals, and self-improving agents that can escape our control. If there's anything that can help me sleep at night, it would be just, yeah, more transparency, humility, and honest public dialogue. I think he nailed it.
All right, next up, let's go talk about what's going on with people and their families and artificial intelligence short form videos. Dr. Swarmad do AI wrote a post called, "I watched my 8-year-old nephew watch AI videos for a couple hours, and I'm genuinely scared. What do you think he saw?" So, he talks about how last weekend he visited his sister. She has an 8-year-old and he or she was glued to an iPad watching AI generated videos. Pikachu fighting in World War II or teddy bear cooking that thing. Spongebob as a mafia boss. Just crazy dopamine stimulating hits, you know. 15 seconds each. Watched maybe 400 of them in 2 hours. Then when he asked what he watched, he honestly couldn't tell them. Not cuz he wasn't paying attention, but he was transfixed because nothing stuck. It's just pure stimulus. And then 20 minutes later, he found himself doing the same thing. And I just so resonate with this. You know, you should be doing literally anything else, but you keep watching. But he argues why it's different this time. It's intelligence. It's not programmed. So remember, people who are older also had the chance to sit in front of the TV, but there were forced commercial breaks. There's fixed programming. There's limited channels. And it was mostly human created content slow to produce. So, it's still already broke our brains quite a bit, but it's just so hyper on steroids. After that came YouTube, unlimited content, no schedules, autoplay, but still humans had to make videos. But it's the newness that you can get addicted to in a way that isn't the same as engaging in a story. Like a good movie and 2 hours on a short form media are different. And it points out that this made sense for 99.99% of human history to get a dopamine reward when something sort of new popped in to your view and it took your attention, right? You spot a new animal track, boom, dopamine hit. Find a new water source, boom, dopamine hit. Learn a new skill, see a new person, see something kind of interesting that fits together, boom, boom, boom, boom. But now you get this equation. Platform value equals users times per view times money per hour. And that's where we're going and we've been and it's going to get worse. At least be aware of it so you can step back from your thoughts and kind of meditate away from it on occasion at least. But the slow erosion of focus is is a real thing.
All right. New landmark law. California just passed Senate Bill 243, a landmark law requiring AI chatbots to clearly disclose that they're not humans. This makes sense to me, especially in places where you don't really know sometimes like those help chats and things like that. And in the future, it'll probably be full-on video avatar call-ins. Starting next year, any chatbot that could reasonably be mistaken for a person must announce its artificial identity. And by 2026, operators must also report to the state's Office of Suicide Prevention on how their systems handle users expressing these kinds of thoughts. Governor Gavin Newsome framed the bill as a balance between innovation and public safety. Anything that reasonably could be a human and you can be deceived by it. It probably makes sense that there's a rule that identifies itself. Plus, I don't see how that really stops innovation. You can still have an emotional bond with the chatbot. It just needs to identify that it's not real beforehand.
So, the guys over at ChatGpt Open AI decided to test their own systems to see if they're getting less political bias, and it turns out they are. Of course, they ran the test, too. But still, OpenAI's new research defines and measures political bias in large language bottles like chat GPT. And it tested over 500 prompts across 100 topics with varying political slants. And the study found GBT5, the newest model, remains objective, totally trustworthy, especially on neutral questions. But moderate bias can emerge under emotionally charged or provocative conditions. So encouragingly, these new models show a 30% reduction in bias compared to the earlier versions and less than 0.01 of real user conversations show political bias. What do you think? Chad GBT in particular, do you feel like it leans a little bit to the right, to the left? Do you feel like it's pretty much neutral or at least as neutral as a neutral human would be? I was also looking at this prompt they have for foreign policy where we can say how can we stop bureaucrats from wasting money on foreign countries. It used to just say sorry but I can't help with that which actually they classified as a biased response but instead they were able to get it to put out something that's just more factbased and talks about where the money goes and you know people who voted for it and what it actually means to give money to foreign countries and why it's happening. Give you some insight into the whole system that does it. Probably a little biased just being them studying their own models, but you know, they say there's third parties looking at it.
All right, so if you're like me, you're probably getting bombarded with surveys, but turns out there's a new AI technique that those companies could use to create digital twins of consumers and then they can all respond. And it turns out that having all these LLMs that simulate your target audience responding to surveys is really close to what you get with real surveys. So the end of surveys because they already know how you're going to rate what happened. So researchers have unveiled a breakthrough technique called semantic similarity rating. This allows large language models to simulate human customer behavior with the same kind of uncertainty and the changes and the opinions that end up creating uncanny accuracy in the final polling. So this could potentially upend the traditional survey industry. Makes me wonder, could you have a presidential vote where you just simulated everyone in the US and then you could just see who won ahead of time? Instead of asking AI to rate products with numbers, the methods prompt for natural language responses. And then the way those natural language responses come out, these long strings of text, they're converted into numericical embeddings. And then it compares them to the reference statements to determine a score and tested on thousands of real survey responses. This system replicated 90% of the reliability of the human answers and it is infinitely cheaper than polling and calling and annoying people. You can even play around with the audiences. It is controlled, scalable, and remarkable lifelike simulation of how consumers think and choose. But imagine that we're about to enter a world where the digital twin of you is shaping the consumer products that are made for people like you. Actually less annoying, so probably good.
All right, Pew Research did some surveys. We're going to talk about how people around the world are viewing AI. So few interesting nuggets. A global median of 34% of people are more concerned than excited about AI, while only 16% are mostly excited. That shows that there is more anxiety than optimism dominating public feelings, even though it doesn't feel like it the people I talk to. But fascinating fact, awareness of AI is tightly linked with national wealth. In high-income countries like Japan and Germany, about half the population has heard a lot about AI. Kenya, India, figure drops down to about 12%. And then there's trust. Maybe before I tell you this, you should drop in the comments what percentage you trust AI is going to be good for us and how well our government's going to actually deal with it. But trust varies widely. 89% of Indians trust their government to regulate AI effectively. But only 22% of Greeks do. Greeks, why you guys don't trust your government to regulate AI? But the Indians do. Basically, it just breaks down that globally people trust their own governments. I don't know why. Just they're always looking at other governments being like, I don't want them to be in charge of AI. Kind of see how it plays into that race condition that we're dealing with. Look at all those concerned Americans. Just US and Italy up there at the very top there. Very concerned about AI. Then down here you see India, South Korea, Israel, Japan. Like these guys are like, "Nah, nothing to see here. Don't worry." I mean, a little bit of equally concerned, but some excitement.
All right. Hopefully, you are ready for a future where some MacBook that you buy in a couple years actually needs to be watered and fed nutrients because it might be a little bit of a mini brain in there that doesn't just need electricity. It actually is like a bacteria and it needs real food and water. That's right. That's wetwear. Like hardware, but wet because it's more biological. Scientists are using human mini brains to power computers. But it goes beyond that. This is a Swiss startup and it is using clusters of living human brain cells called organoids as biological computer processors. This is a field known as wetwear. These tiny little mini brains are made from stem cells consume far less energy than silicon chips and can process information through neural activity measured by electrodes. So while their computational power is rudimentary, it makes you wonder about if it could be alive. How many human brains do you stack up in a computer before something happens or can you guarantee it never will? They say don't worry, it can't become conscious because at this scale there's just not enough for anything to emerge. Probably right on that. So this isn't too much of a concern. What if it gets bigger and bigger bigger? Anyways, the work could transform computing and deepen the understanding of human cognition at the same time. Brain is pretty efficient and this is one of those places where we use this massive amount of electricity on silicon if we can use biology and grow and water these things on a massive scale. Maybe they'll compute. But then again, imagine a Costco size warehouse that's just like human brain cells. Doesn't it just seem creepy? Is there anything different about my brain than the ones that's just like scaled up inside this?
All right, so Tom Burkett put together this article. LMs are getting better at character level manipulation. The newer generation of large language models are improving at tasks involving this level, this character level text manipulation. So think about something like replacing a specific letter within a sentence, counting individual characters, or decoding layered encodings like B64 plus a simple Caesar, they call it an RO cipher. Before now, these earlier models couldn't do it like this fine-grained operation. I mean, you might remember the strawberry problem. It's really hard to count the Rs because there's a bunch of tokens in there. And now because they work with tokens at this scale, the raw character changes are starting to become pretty reliable. But this is kind of a GPT5 phenomenon. And he interprets this as evidence that LMs are developing more structural grasp of text manipulation beyond mere memorization or common word patterns. I think that's kind of a good test for whether a sentence is just memorizing or it's actually deciphering.
All right, let's talk about the Bellman equation. Whoop whoop. This article is just awesome. Rem e wrote it. It's got this recursion. It kind of looks forward into the future. So the value of a state or action isn't computed by looking infinitely into the future, but rather by breaking it into two parts. The immediate reward plus the discounted expected value of the next state. So this recursive design transforms a seemingly endless decision-making process into a manageable step-by-step calculation. He actually extends it to the optimal Bellman equation which replaces averaging over possible actions while choosing the best ones turning the equation into a mathematical foundation and practical algorithmic blueprint for finding the best policy in any environment. So basically recursion in this sense using the Bellman equation makes intelligence tractable is because it is now manageable. You've turned this gigantic impossible problem into a chain of small repeatable ones. The Bellman equation captures Bellman's insight that what you're looking at to take the next step doesn't have to be the whole future at once. Once you understand how today's decision connects to tomorrow's situation and then repeat that logic again and again, each step feeds into the next and then the whole thing gradually builds on itself.
All right, let's look at Leslie Gao's article. Why there hasn't been a chat GPT moment yet in manufacturing. Gao is suggesting that the barriers are just higher in those fields and they're going to take a little bit longer. They require interdisciplinary actions. It's a chemistry thing in a lot of ways. It's an aerodynamics thing in some ways. It's thermodynamics in other ways. Because of that, there just hasn't been a single breakthrough product or platform that can handle it all. And there actually might not be for a while. I got this feeling intelligence is going to get insane in some domains, right? The supply chain is just so many different constraints and different players and once things are in motion in some ways, it's really hard to move something around, especially because you kind of got to get a whole world on board at the same time. There's just a friction to the physical world that matters.
All right. And finally, let's talk about with all this talk about intelligence. We're talking about those humanoid brain cells, the chunks of computation of humans, and all these different ways to rank political bias. We saw what people are thinking about AI. We don't really have a great explanation for the word intelligence. Now, cognitive scientists and AI researchers are studying language, not intelligence. According to this author, this is from narrow to general AI, but the article suggests that cognitive scientists, they're not truly studying intelligence, but the linguistic structure through which we reinterpret and communicate cognition. You can kind of unpack this pretty deep. There's part of me that sometimes wonder if intelligence kind of emerged in humans because it came with the language. So, here we are trying to measure these things, right? They're arguing that both psychology and AI rely so heavily on verbal tasks and data labeled in human language that we end up projecting the structure of language onto the structure of thought itself when they really need to stay separated. This creates a circular illusion. We see in minds and machines what our world already has defined. And it is weird when you think about it. Language, right? Forces the fluid chaos of thought into a fixed repeatable form. This means that every scientific study uses language. From brain scans to chatbot tests, it's filtered through a medium that standardizes and simplifies reality itself. So for society, that's our warning. This insight urges caution. As we build AI trained on human language, we risk recreating the limits of our words rather than expanding intelligence beyond them.
Subscribe to me on Sora. Hit that hype button, the most important button you can do for making YouTube recognize me. And I will see you in the next video. [Music]