Hurricane Polo is a good example of why asking, “Was this caused by climate change?” is too simple a question. The immediate reason Polo became so powerful was an unusually favorable environment. The National Hurricane Center reported sea temperatures around 31°C (88°F), very low wind shear, abundant moisture, and, importantly, warm water extending roughly 50–100 meters below the surface. That deep ocean heat makes it harder for a hurricane to churn up cooler water and weaken itself. Then there is El Niño. The Pacific is currently in a strengthening El Niño, with NOAA reporting temperature anomalies exceeding 3°C above normal in parts of the eastern equatorial Pacific. El Niño did not “cause” Polo, but it helped create an unusually warm Pacific background in which storms can develop. And behind both is the longer-term question of climate change. A warmer climate raises the baseline temperature and heat content of the oceans. That does not mean climate change created this particular hurricane, and we do not yet have a Polo-specific attribution study telling us how much stronger it became because of global warming. So I would describe the chain this way: Climate warming raises the ocean baseline → El Niño adds another layer of Pacific warming → Polo encounters extremely warm, deep water plus low wind shear and moisture → explosive intensification becomes possible. The mistake is looking for a single cause. Hurricanes are systems. Several conditions have to line up at the same time.

I am watching RoboCop 3, and it struck me that this is another example of something that has been going on in movies for decades. The corporation is the evil empire. Corporate executives are greedy and corrupt. Ordinary people are the victims. Government and police become tools of corporate power. Eventually the people fight back, corporate property gets destroyed, and the audience is encouraged to cheer. And importantly, the movie sometimes gives this system a name: capitalism. One movie probably does not make someone anti-capitalist. But think about the cumulative effect of seeing variations of this story again and again for decades. Movies, television, books, and games can gradually construct a mental model: Capitalism leads to greed. Greed creates giant corporations. Corporations capture government. Wealth becomes concentrated. Ordinary people become powerless. Eventually the corporation becomes the government. Once that mental model exists, people do not necessarily distinguish between capitalism itself, competitive markets, monopolies, regulatory capture, corrupt corporations, or corporations exercising political power. They can all become emotionally associated with the same word: capitalism. That does not mean the problems being portrayed are imaginary. Corporate concentration, political influence, exploitation, and regulatory capture are real issues worth examining. But that is different from assuming they are the inevitable destination of capitalism. That is what interests me here. Popular culture does more than entertain us. Repeated stories can help construct the mental models through which we later interpret the real world. RoboCop 3 is just one particularly obvious example. Links [Between Fiction and Reality: RoboCop and the Critique of the Collapse of Neoliberal Society](revistas.pucsp.br/index.php/…) Academic article covering all three original RoboCop films and their critique of neoliberalism. [Washington Post review of RoboCop 3 (1993)](washingtonpost.com/archive/l…) Contemporary review discussing the film in terms of class conflict and corporate capitalism. [John Kenneth Muir: Cult Movie Review — RoboCop 3](reflectionsonfilmandtelevisi…) Explicit discussion of capitalism, OCP, and corporate power in RoboCop 3. [Property and Privatisation in RoboCop — Cambridge University Press](cambridge.org/core/journals/…) Scholarly analysis of privatization, policing, corporate power, and neoliberalism in RoboCop. [RoboCop — Oxford Academic / Liverpool University Press](academic.oup.com/liverpool-s…) Scholarly book on the film, including discussion of anti-corporate science fiction and Reagan-era economic themes. [The Persuasive Effects of Narrative Entertainment — Cambridge University Press](cambridge.org/core/journals/…) Meta-analysis on how fictional narratives can affect attitudes, beliefs, intentions, and behavior.

Jim Croce had an extraordinary gift for taking ordinary human feelings and turning them into songs that stay with you. “Photographs and Memories,” “I’ll Have to Say I Love You in a Song,” “Time in a Bottle,” “Operator (That’s Not the Way It Feels),” “Lover’s Cross,” and “These Dreams” are some of the songs I keep coming back to. There was something remarkably personal about his music. He could tell a whole story in a few minutes, often with little more than his voice, an acoustic guitar, and words that sounded as though someone had actually lived them. Croce died in a plane crash in 1973 at only 30 years old, just as his career was really taking off. That reminds me of Don McLean’s “American Pie” and “the day the music died.” It sounds related because Croce also died young in a plane crash, but it actually happened much later. “American Pie” was released in 1971 and was referring back to the 1959 plane crash that killed Buddy Holly, Ritchie Valens, and the Big Bopper. Jim Croce was still alive when “American Pie” became famous. A remarkably short career, but a remarkably long legacy. And one musical memory naturally leads to another. piped.video/8lpaJj_IOoE?is=zfK_…

Hidden information networks surround us: governments, intelligence agencies, corporations, criminals, and AI each see only part of the picture. The real challenge is separating what we know from what we can only infer.

He is starting to say things that others including myself have been saying. We don't need superintelligent toasters.

Hidden Information Networks and AI

Did she pass the bar? Or did she go into a bar and drink before this video?

I have built my own tools tol implement this idea. And its a little scary at times. It recalls conversations and manages documents and can quickly remind me of what's going on.

After all. Humans don't need a data center to do most of the work being done now. Humans are evidence that general-purpose learning, reasoning, language, planning, and adaptation do not inherently require a data center at inference time. A human brain runs on roughly the power of a small light bulb, yet it can learn from a book, remember what it learned, combine it with prior knowledge, form abstractions, and apply those ideas to situations it has never seen before. We do not retrain ourselves on the entire library every time we learn something new. That suggests part of the enormous compute demand in current AI may be a consequence of how we currently build AI, not a fundamental requirement of intelligence itself. Today we compensate for relatively weak continual learning, memory, and self-directed learning by doing enormous amounts of training beforehand. A more mature AI architecture might instead have a smaller core model plus persistent memory, efficient continual learning, a world model, reasoning, and the ability to selectively study new information when needed. In that sense, today's data centers may be analogous to an extremely inefficient way of building intelligence. The long-term breakthrough may not be "more compute." It may be figuring out how to make AI learn efficiently. The huge data centers may remain important, but increasingly as civilization's shared external memory rather than as the place where every intelligent act has to occur. They could hold the accumulated knowledge base: books, scientific papers, historical records, software, engineering designs, medical knowledge, images, video, experimental datasets, simulations, sensor data, and the provenance needed to determine where information came from and how reliable it is. Then the architecture starts looking more like this: Data centers = humanity's library and research archive Individual AI = the researcher Local memory = what the researcher has already learned Network connection = going to the library when more information is needed Large compute centers = specialized laboratories for unusually difficult calculations and simulations That's quite different from today's approach, where a tremendous amount of computation is used to compress a large fraction of that library into the parameters of the model. A future AI wouldn't necessarily need to memorize everything. It could learn the fundamental concepts and learn how to learn. When it encounters something it doesn't know, it retrieves the relevant material, studies it, integrates the useful parts into its persistent knowledge, and moves on. Humans already work approximately this way. A physicist doesn't carry every physics paper in his brain. He carries enough understanding to know what the problem means, how to reason about it, what he already knows, and where to look for what he doesn't know. So perhaps the long-term technological progression is: AI today: enormous training infrastructure → enormous pretrained model → mostly fixed knowledge. Future AI: efficient learning architecture → persistent individual knowledge → access to an enormous shared human knowledge repository → continuous self-education. In that world, we could still have gigantic data centers. But their primary purpose would shift from "this is where the intelligence lives" toward "this is where civilization's accumulated information lives." And that makes the Internet itself look somewhat different too. Instead of primarily being a network of webpages designed for humans to browse, part of it could eventually become a structured global knowledge infrastructure designed for machines and humans to learn from together.

I largely agree with this. I think the long term direction is toward AI that learns more like a person does instead of requiring nearly all useful knowledge to be compressed into one enormous pretrained model. The data centers would still matter, but increasingly as shared repositories of human knowledge and as specialized places for very large simulations, training runs, and unusually difficult computation. The individual AI would carry what it has already learned, maintain persistent memory, and retrieve additional information when needed. The key breakthrough would be efficient continual learning. An AI should be able to read a book, study a paper, learn something new from it, integrate that knowledge into what it already knows, and continue from there without needing to be retrained on everything again. In that sense, the goal is not to put all of civilization's knowledge inside every AI. It is to build an AI that knows how to learn, remembers what it learns, and knows where to find the rest.

Thinking, Understanding, and AI I was looking at a simple meme about honesty, and it turned into a useful example of what I think real thinking actually involves. At first glance, the idea seems simple. Honest people tell the truth. Dishonest people do not. An AI can respond to that immediately, and the first answer will probably sound reasonable. But once you spend more time thinking about the words, the problem becomes more complicated. An honest person can sincerely tell you something that is false because they genuinely believe it is true. A dishonest person may knowingly create a false impression, but the reason for doing so might be selfish, protective, compassionate, strategic, or something else entirely. Someone can even make a technically true statement while deliberately leaving out information in order to mislead you. So now we have to separate several things that initially looked like one thing: what is actually true, what someone knows, what they believe, what they say, what impression they intend to create, why they are doing it, and what consequences follow. That is where this becomes interesting for AI. A language model can often generate a very good first response by recognizing the pattern in a question and producing a plausible answer. For many everyday tasks, that is all we need. There is no reason to conduct a philosophical investigation every time somebody asks how to change a tire. But when the goal is understanding the world, the first answer should often be treated as a hypothesis rather than the conclusion. A more capable AI should be able to examine its own first answer. What assumptions did it make? Are several different concepts being treated as though they were the same thing? Are there counterexamples? What information is missing? Could the same behavior arise from different motives? Could a statement be literally true but still misleading? Does the explanation still work when unusual cases are introduced? In our honesty example, every time we introduced another case, the original explanation had to be revised. The important part was not simply arriving at a better definition of honesty. The important part was discovering that the original model of the situation was incomplete and then improving it. That may be an important distinction in AI. Answer generation asks, “What is a reasonable response?” Reasoning asks, “Why does that response make sense?” Deeper reasoning asks, “Under what conditions would that response be wrong?” Understanding asks, “What explanation still works after we test the assumptions, exceptions, motives, causes, and consequences?” Not every question needs that level of analysis. But if we want AI to help us with philosophy, science, economics, human behavior, history, or understanding the nature of the world, then simply producing a convincing first answer is not enough. The AI has to be able to keep thinking after the first answer.

Not every decision requires deep analysis. Much of everyday life works perfectly well with experience, intuition, and first-pass judgment. But if the purpose is philosophy, or simply understanding the world more accurately, the first answer should usually be treated as a starting point rather than the conclusion. You have to examine the words, assumptions, causes, motives, exceptions, and implications. That is where thinking becomes understanding.

Notice there's still no solutions provided.

Maybe the better alternative to DEI hiring would have been to define what “fair hiring” actually means. The law can say you cannot discriminate on the basis of race or sex, but that does not eliminate the underlying problem. Think about how companies actually hire people. Usually they do not write down exactly what is required to perform the job, test applicants against those requirements, and hire the first person who passes. Instead they collect applications, interview several people who are qualified, and then choose the person they think is “best.” But what does “best” mean? Often it means the person the hiring manager likes best, thinks will fit in best, communicates with best, went to the preferred school, has the preferred background, or simply creates the best impression during an interview. Some of those considerations may legitimately matter to the job, but they also create a large area of subjective judgment where discrimination can occur without anyone explicitly saying, “I am discriminating.” That was one of the real problems DEI was trying to address. But there is another way to attack the problem. Define the job requirements in advance. Put them in writing. Develop a reasonable way to determine whether an applicant meets them. Apply exactly the same requirements to everyone. Then, once somebody satisfies the predetermined requirements, hire them rather than continuing to search for somebody management happens to prefer. Notice what disappears from the system. Race and sex do not have to be considered at all. The question becomes much simpler: What does this job require, and can this person do it? That would not eliminate every hiring problem, but it would reduce the enormous amount of subjective discretion in hiring that made discrimination possible in the first place. Instead of trying to correct discrimination after it appears, redesign the hiring process so there is less opportunity for it to occur.