Put me on a desert island with nothing but some clothes, some fishing nets, a few weeks of food, a good knife, a good axe, and a good whetstone, and I could eventually rebuild a reasonably good AI Large Language Model. This might take me 20 years, but I could do it. The island would need to have at least one large river with a strong current, since I would not have electricity, so I would need the current of the river to power the AI. Obviously I would need to spend the first few weeks building a home and setting up the nets to give me a reliable supply of food that I could get without too much effort, but then I could get to work. The "Large Language Model" would be restricted to what I have in my head, but that is still a reasonable amount. I would need to encode what I remember of John Doe, Satan's speech in Milton's Paradise Lost, the parts I remember of Shakespeare and Hemmingway and Fitzgerald and Doris Lessing. I have some of my own essays memorized, and I could write some more. All that matters is that I encode enough text that I can train the model. This might take me 7 or 8 years, but I could eventually encode a reasonably large amount of text. I would end up with a very small Large Language Model. Some people think you need electricity to build a computer but that is not correct. Back in 1936 Alan Turing described an "infinite state machine" which became our computers, but there is no electricity in his description. You only need something that goes around and around and around, so if you build an old fashioned waterwheel, like you see on any old mill, then you can build a computer. This is why I said the island would need to have at least one big river with a strong current. I would spend the first two years setting up the infrastructure: a waterwheel powered by the river, a drive shaft, some building to house it all, multiple "state machine" wheels that would hold a large number of uniform sticks, which could be pushed left (0) or right (1) which would set me up to start binary programming. I would probably use different wheels to encode different "software libraries", such as one wheel for math operations, one wheel for text operations. I would need some kind of clutch to move the "reader" from one wheel to another, as I engage different "libraries." Figuring out the clutch would take some careful thought, but if I hardcode the number of wheels and what they do then the work becomes easier (that is, if I hardcode "wheel 4 is for math operations, wheel 5 is for text operations" then this becomes easier). I would have to get good at wood working, but presumably, if it is the only thing I do for 20 years, I will get good at it. The most difficult trick would be figuring out the knots that, on the one hand, are firm enough to keep the sticks in place but, on the other hand, are flexible enough that the sticks can be pushed to the other position, when necessary. I would have to get the knots right so they could imitate the yes/no behavior of semiconductors. Once I had the machinery working, I would have to develop the "software," though the "software" would be patterns in the sticks that can be either left or right. I would need standard patterns to handle strings of text, and other standard patterns to handle numbers. This would take me 6 or 7 years, but I could do it. I've read the original LLM paper from 2017 (Attention is all you need) so I get the basic idea, and (in real life) I'm working on specialized n-gram models that are informed by the NLP advances of recent years. I would eventually have working AI. This might be a stupid way to spend 20 years on a desert island but at least it would keep me busy. But here is the important point: there are now tens of thousands of engineers who can all do the same thing. The knowledge is widespread. The idea that AI will go away is ludicrous. Now that we know how to do it, we realize how easy it is. I've told this story before, but a few years ago I was on the subway and I was sitting next to a Hispanic woman who was traveling with her daughter, who I judged to be 6 or 7 years old, and the mom was drilling the daughter on some of the math the daughter had recently learned: Mom: "What is 3 plus 4? Daughter: "7" Mom: "What is 9 plus 0?" Daughter: "9" Mom: "What is 7 minus 0?" Daughter: "7" The daughter clearly knew what zero was, and how to use it. By contrast, many of the most brilliant people who ever lived didn't know about zero: Plato, Aristotle, Ptolemy. They lived and died with no concept of zero. Some ideas are like that: before they are invented the most brilliant people alive don't know about it, but after they idea is invented then a 6 year old girl can easily learn it. AI is like that. The 2017 paper (Attention is all you need) described the "self-attention dot-product" technique, which is a branch of Linear Algebra. And now that we know it, we realize it is relatively easy. Put me on a desert island and I can re-invent the whole thing. And many other engineers can do the same. Therefore, AI is not going away, it is with us permanently, like the concept of zero. People need to get used to this.

Russia isn't winning, but Putin has shown he has political capital that no one had guessed. From 2000 to 2022 if you had asked, "Could Putin lose 1.5 million troops in a war and survive?" people would have laughed at you and pointed to Afghanistan. The assumption was that if Putin lost even 50,000 troops in a war then he'd be overthrown. This last decade has been very good for right-wing populists and no one has benefited more than Putin.

"The evil, selfish boomers have ruined my future!" they shout as they vote to ruin their own futures.

It's stunning how good this was. Nathan Fielder understands better than most how much a camera lies, and in some sense he makes the cameras a central character, in that most of his scenes raise the question, "Is this person lying for the cameras?" or "Is the camera lying to you right now?" Maybe once a decade we get a film that attempts this meta of analysis of analysis of analysis. In particular, watching Amanda Seyfried watch Elizabeth Holmes watch Amanda Seyfried portray Elizabeth Holmes was a level of meta I have not seen before. It felt like a magic trick, the way Nathan Fielder kept piling up the levels. And at each level you had to ask, "How much of this is true?" To me, attempting this in a documentary is much more interesting than attempting this in a science-fiction movie such as Tenet.

I assumed that Trump's re-election in 2024 meant the end of American democracy, but so far everything has been fine. American democracy is surprisingly resilient. Trump II has been surprisingly normal.

France has a very generous social welfare system. The idea that the people are being oppressed by austerity is not quite accurate. This is a very good comment. Still, there is a plenty of room for government action, but it is not the action that people are talking about. To the extent that France is undergoing an economic shock, it is due to the trade policies of china, which @Brad_Setser has been writing about. And the French government can take some action there. But "Do more to confront China" has not yet been seen at any protest, on any sign.

Me and my friend Lisa are going to see You Can See Everything today. The first several movie theaters that I looked up were completely sold out, so this movie is very, very popular.