Episode Transcript
[00:00:00] Good morning everybody. Well, this weekend we heard a lot about AI destroying humanity. 10% chance that we all die, A posting by Dario at Anthropic believing that all the swarms of AI bots could shut down the Internet. And then a rather half hearted agreement between Sam Altman and Dario Amadai and maybe Elon Musk, but not really that they were going to slow down.
[00:00:27] I have a lot of reactions to this and I'm sure you all do too. But my reaction is as follows.
[00:00:34] First of all, these guys are not going to slow down. They've taken out billions of dollars of loans, they're all planning on going public. They're very, very greedy entrepreneurs. And if they really thought they were building something that was going to destroy humanity, they would shut down their companies. And they're not doing anything close to that, they're just writing articles.
[00:00:54] Second of all, the evils or dangers of AI are implicit in the design of what they built, what they did. And what they're doing is they're building mathematical algorithms that have self improvement and self learning capabilities to manipulate and learn from large amounts of information and turn that into an agent that can take action.
[00:01:19] That's what a large language model does. And it is a language model, which means that it uses language or words or tokens to create its intelligence. However, it's being trained on the Internet, so it's being trained on every piece of content that could possibly have values, strategies, information, science, math, physics or other things in it, including something cybersecurity, hackers, movies like the Terminator, blogs from right wing, racist or anarchists, it's all in there. So the values of an OpenAI or an anthropic model are the values of society. And even though the chief Technology officer at Anthropic talked about the fact that alignment does not align values, it only aligns behaviors. The values of these AIs are the values of society.
[00:02:16] So of course since we have so many, maybe not a large number, but we have a fairly large number in number of people who break the rules, steal things, believe that theft or rule breaking is okay, commit corruption, including people in Congress and all over the government. Why wouldn't the AI learn all that? It has no personality.
[00:02:40] Its personality is us. So the job of the AI vendors is not to complain the or ask the government to help them, but to deal with this issue and not design something that's going to be dangerous. It's so obvious to me if Boeing told us that there's a 10% chance our planes are going to crash and then ask the government to fix it. Boeing would be out of business. Boeing is a mature market of where we're going with AI. Eventually AI is going to be integrated into every part of our lives and we can't use it if it's not safe. And if these guys think safety is an option or maybe something they should design in later, I think they're just making a fundamental mistake. And because we're all so enamored with them and maybe intimidated by them, we don't hold them to a higher quality, but we would hold a car manufacturer or an airplane manufacturer or a computer manufacturer to a higher quality. We wouldn't buy a PC if there was a 10% chance it was going to blow up in our face. In fact, the business would cease to exist. So there's a very irrational thinking process going on here. And I think it's irrational because the government and most of us are intimidated by these data scientists who built this stuff. But to me it's just very, very clear that we should not buy or use something that's that dangerous. And for those of you in the corporate market, where most of you probably are, if any of this stuff is true, I think your legal people are not going to want you doing business with these vendors. And you're going to make an argument, oh, it's such so great and it has so much potential and all our competitors are using it. And then the vendors are going to say, yeah, but you know, there's a 10% chance that the agent that you built is going to go out and wreck your customers experiences and their systems. And, and you'll have to take that liability. By the way, another issue here is liability. If OpenAI or Anthropic's models commit fraud and you're running the model, are you liable or are they liable? If their model destroys the IT infrastructure of your customer or your partner, are you liable or are they liable? Presumably they should be liable, just like the airplane manufacturer is liable if the plane crashes.
[00:04:58] Now to take that analogy a little further, if the airplane is perfectly safe but the pilot crashes it into a building, obviously it's the pilot's fault. And if the airline fails to repair the plane or doesn't give it enough fuel or doesn't train the pilot, maybe the airline's at fault. So there's lots of chains of thought here on who's at fault when these risks take place. But I think the vendors to sit around and sort of make this academic debate like they're gods, I mean, I can't stand listening to these guys talk because they sound so above it all. It's their responsibility to make something that's safe, period. It's that simple. Don't ask the government to do anything, just do it. And we as buyers, as consumers, as people that buy their stock or their bonds or whatever it is, our financial relationship is with them, we should hold them responsible as well. And, and that's where I think the market pressure will come is when you as a corporation or you as a customer, as an individual say to say to them, you hurt me, you damaged my business. I want my money back or I'm going to sue you. I think they'll get their act together. I don't believe it's a technical problem to solving these issues. I don't believe this runaway train has to continue in this direction. I think these companies have made design decisions that create this situation. And tomorrow I'm going to put out, we're going to do a big launch of our new version of Galileo, which we call Jupyter, and you're going to see that we've built a system that's completely trustworthy and doesn't do this stuff. Yet they don't seem to care because they're trying to collect as much money as possible as fast as they can. Now, I don't mean to be negative about AI or the Frontier Labs or any of these executives, but I think this frenzy of fever about making money on AI has just made everybody a little bit insane. And I think many of you are rational thinkers like me, and you come from human capital backgrounds. And you probably would agree with me that even though we do want to have the best military equipment and the best monitoring equipment and the best buying equipment and all that military stuff, if the AI can't be controlled, then the user of the AI takes responsibility and they go back to the vendor and say, your product is junk. I don't want it. Now. You know, the federal government at the moment doesn't seem to have a lot of conscience about a lot of things. And maybe they don't care if the drones that are running anthropic or OpenAI just randomly go out and kill people and nobody knows what happened. But I kind of think that's not the way the defense industry works. I don't think that's the culture of our country. And I don't think most countries around the world would be particularly happy if the tools that they're using for defense start attacking their citizens.
[00:07:58] Likewise, none of us who are in the business World want our systems to attack our customers or our prospects or our channel partners. So it's pretty obvious to me that this is a design flaw that has to be fixed. Now, getting back to that issue of the design flaw, somewhere along the lines of all of the conversations that Dario Amadi has had with the press, he mentioned that they're going to have to retrain the models. And my answer to that is, duh, yes. If you train the model to read a bunch of evil science fiction, which it doesn't know is science fiction, by the way, and learn from that science fiction how to make a bomb or shut down the Internet or destroy humanity. Yeah, maybe you should have thought about that before you put that data into your model. And it's just sort of staggering to me that this quest for superintelligence is so. I don't know what the word is. Academically interesting that no one is thinking about the impact. I suppose back in the 1950s, when the United States was building the atomic bomb, Robert Oppenheimer and the other people that worked on it were scientifically fascinated by the idea of splitting an atom. But I think they were also very ethical and honest and, and worried that splitting an atom could be the end of humanity, too. And they did it under controlled conditions, and they did it with the government. They didn't just run around and try to make a trillion dollars for themselves. You know, it's weird. I don't know how to say this, but I feel like we've entered a period of commercialism around the world, certainly the United States, where making money is more important than anything else. I mean, it absolutely seems to overpower every possible decision that could be made. If we can make money doing this, we're going to do it regardless of who gets hurt or how unethical or unfair it may be. And that's exactly what's happening here. Investors, people building data centers, CEOs, et cetera, people in the stock market, Jensen Huang, all of them. Now, you know, I think these people, if you met them face to face, are probably great people, and they all have ethical backgrounds and they have families and kids, and they certainly don't want to hurt anybody. But it's sort of a strange group think process that's taken place, that people like Elon Musk go on the Economist and say the world is going to be wiped out and we're not going to have any jobs and we're all going to be sitting around looking for things to do and no one sort of pays attention and just sort of Just moves on with her life, all in the interest of making more money. And perhaps the human need for power and wealth is so great that the obvious human needs for health and safety just get left in the dust. And we have the competitive issue of the United States versus China and our economy versus their economy and the railroads and what happened in the railroads and so forth. I'm reading a long book. I'll give you a link to a fascinating book about what happened in the 1800s when we had a very big, very similar to what we're going through right now. It was during the 18th century, 1870s, when we were in the railroad era. And it's so similar to today where the bond market and the stock market and mom and pops and retired people, everybody threw their money into railroad bonds and railroad stocks and contract firms and manufacturing firms and there was graft and corruption and all sorts of fake companies being created just like there are now. And, and then there was a big crash and a lot of people got hurt, I suppose, you know, human nature never changes and we're going through that again. The other reason this is such a tricky problem to discuss is there's a lot of political side taking to deal with. You know, right now, as of this week, the Trump administration is very pro AI and wants nothing to do with any of this safety stuff because they want to beat the Chinese.
[00:11:55] And for some reason they just don't care about these issues.
[00:11:59] Bernie Sanders and the socialist Democrats in the United States and the more left wing leaning people are very anti data center, anti infrastructure, pro energy equality. And they see this as an oligarchy taking over the world and all the money going to the hands of the richest people in the world and therefore they're pushing back.
[00:12:21] And according to the research in the New York Times and some other things I've been reading, there are 80 or 90 active protests against data centers in the United States right now. And so a lot of the right wing politicians who would normally be aligned with the AI growth are now anti data center because they're going to get voted out of office. So the data center, which is the underlying sort of energy behind this AI technology, is becoming in some sense the political anchor or weight of that's slowing it down from maybe getting out of control. And because maybe what'll happen is the pushback on data centers will force the infrastructure going from the bottom up instead of from the top down to be more responsible. What they're doing. The problem with the data center debate, of course, is, I mean, even, even in Florida, there's pushback in Texas, there's pushback even in the right wing states. The problem with the data center issue though, is if you took a position that you wanted to save energy and you wanted to save the environment, and you don't want a data center near your house, so you want to slow down AI, that developer is very likely to just pick up, leave, go to another state, go to another country, or go into space and build it anyway.
[00:13:37] So the city or municipality might say to themselves, well, you know, one way or another, if we lose this opportunity to get whatever tax revenue we think we can get, maybe we should take it. So because the world is so flat and there's so much opportunity for mobility from place to place, that political force may have effect on who gets elected, but it may not have any effect on the actual progress of AI. Now, you know, the other angle on sort of stakeholders here is the software companies who I work with, including us, who build products on AI. Well, we are trying to solve real problems. We're the ones that are trying to make these tools useful and really create an ROI out of them. We're building the stack of solutions on top of them. And if the models are unreliable, we can't do what we want to do. So in some sense, we are either dependent or stakeholders in the interest of these frontier labs getting their act together. And I don't mean this negatively, but these, these labs are paying people hundreds of millions of dollars to do this. And I go into ChatGPT and I get the wrong answer all the time. I mean, I asked very, you know, complex questions about the economy and labor markets and laws and regulations and business trends and companies. And I mean, I have to read everything very carefully and double check it because it's always making mistakes.
[00:15:03] Why is that acceptable? Is it acceptable if the airplane engine just doesn't fly sometimes or, you know, that the wheel just doesn't stop and the plane lands on one wheel? And is it acceptable for your car to just halt in the middle of the freeway and say, oops, sorry, we had a mistake? By the way, I have a friend with a Tesla that did that. I guess because these maturities are so low and these products are so new, we're willing to accept an awful lot of glitches here. Like the little line at the bottom of the AI harness that says so and so might make mistakes. How would you feel about a line like that in your car, in your airplane, in the food? Such and such a food might cause poisoning? I don't think you would be too happy with that. So, I mean, I think ultimately this issue is not will renegade AI destroy humanity? But rather how quickly do we mature these companies so they behave like grown up companies. They can do, by the way, you know, you can do advanced R and D on airplanes, on food, science on drugs. I mean, anything that goes into our bodies or affects our livelihood has advanced science that isn't ready for prime time. They don't ship it and test it on us.
[00:16:25] I mean, in some ways, we're all part of a giant beta test here for the frontier models. And we're all so excited about it that we're not paying attention to the fact that it's a beta. Maybe it's an alpha.
[00:16:37] Anyway, all right, you can debate this all you want, but anyway, that's where I sit on that. Tomorrow on the 15th, we're going to be launching the Jupiter release of Galileo. It has incredibly powerful new things in it, a new architecture. I will save the story for the podcast and the information that comes out tomorrow, but you're really going to want to pay attention. I thought I would just get my voice out there this morning as the world talks about AI slowdown and AI safety. And you're all going to be discussing it at work and at home and at lunch and at dinner, and maybe this will help you think it through. That's it for now. Bye.
