The Spiraling Basketcase

The useful technology is not the same thing as a sensible price tag As of October 31, 2026 AI is producing real revenue, real productivity experiments and real technical progress. It is also attracting extraordinary capital commitments, complex financing and expectations that may outrun what customers can pay for. Those statements are not contradictory. They are the starting point for a more useful question than “Is AI real?”: which parts of the AI economy are earning their cost of capital, and which parts are being priced on a story about the future? That distinction matters because the current AI-bubble debate mixes together the usefulness of software, the profitability of individual companies and the value of financial claims built around future demand. A correction in one does not automatically mean the others disappear. The headline corrections matter S&P did not downgrade Meta to junk. On October 30, S&P affirmed Meta’s AA- rating but revised its outlook from stable ...

Ask a large language model a question and it gives you an answer with total confidence. Ask it how it got there, and it will happily make something up — the model has no built-in ability to narrate its own internal process, because its internal process was never written in a language anyone can read. That gap is exactly what the field of AI interpretability exists to close, and the way researchers are closing it looks less like reading code and more like neuroscience. Why You Can't Just "Read" a Neural Network A modern language model's actual substance is a long list of numbers — weights — connecting artificial neurons across dozens of layers. Even a modestly sized open model has billions of these weights, and none of them were hand-written by an engineer. They emerged from training: the model read enormous amounts of text, repeatedly guessed the next word, and had its weights nudged a tiny amount every time it was wrong. Multiply that nudge by trillions of guesses...

The AI jobs future will probably be uneven: fewer routine entry routes, more work done with software assistance, and a long argument over who receives the productivity gains. Dylan John’s video “MIT Just Discovered What Jobs AI Actually Killed” makes that case through a useful contrast between a sudden “crashing wave” and a gradual “rising tide.” The evidence suggests that both possibilities deserve attention, but neither should be treated as a settled forecast. The future may arrive first at the bottom of the ladder The video’s most important idea is not that every job disappears. It is that the first visible damage may happen to the jobs people use to enter a profession. Junior analysts, assistant writers, entry-level programmers, translators, and customer-support workers often perform structured tasks under supervision. Those tasks are precisely where a language model can be useful: summarize a document, draft an answer, translate a passage, find a pattern, produce a first versio...

Daniel loved lists because lists did not argue. They waited. They held still. They told the truth in rows. * * * Nails. Travel bag. Passport. Rent. Diapers. Milk. Wipes. Groceries. Electricity. Internet. A missed day of work. Daniel typed the items one after another. Not everything at once. Priority first. Then the rest. He told Elena they needed a list. Most important to least important. Necessary things. Things that could wait. Things that would change as they moved through different stages. Otherwise, he said, he would lose track. * * * Elena agreed. Of course she did. A list was practical. Her life had never lacked needs. What it lacked was money arriving at the same speed. * * * The conversation had begun with her hands. Factory hands. Elena wanted her nails done before the trip. Daniel asked to see them. She sent a photograph. Her hands embarrassed her. She thought they looked ugly. Daniel saw something else. Work. He told her they were w...