Even if advanced AIs are considered part of humanity, that does not guarantee comfort—future humans could be undesirable (e.g. a generation of Nazis), so kinship alone does not address concerns about handing off this level of power quickly to entities we don't fully understand.

The newest blog post is titled on whether prediction markets or policies should be legal, based on Robin Hanson's idea of a futarchy.

Bryan Caplan argues that people are much less educated than 12 years of mandatory schooling would predict—people forget the basics of government, math, and science after leaving college—yet he forces his own homeschooled children to learn mathematics because unschooled children seem to struggle with even basic arithmetic.

There is no evidence that spending time playing video games is a worse use of time than attending school; studies on the efficacy of school show children are wasting their time there anyway, but they suffer while doing so, whereas they don't suffer while playing video games.

From an evolutionary first-principles standpoint, it would be astonishing if natural selection produced a survival machine that lay around doing nothing for its first 18 years; the existence of childhood implies we are evolved to be curious and to explore.

Compulsory schooling imposes large hidden costs beyond boredom: children have a different sleep cycle than adults so school schedules leave them sleep-deprived and disrupt development, while roughly three hours of daily homework on top of eight hours of school consumes the time children need to grow and develop.

A school superintendent's district that delayed required math instruction until seventh grade found that those students were no more behind in mathematics by eighth grade than students who had been learning math since first grade.

Identical twins separated at birth and raised in different households tend to turn out quite similar regardless of how different their parents are, suggesting parenting style does not significantly change how children ultimately turn out.

Roughly a third of a modern computer science curriculum can be traced directly to von Neumann's contributions, including algorithms (merge sort), linear programming, von Neumann architecture, game theory, finite state machines, cellular automata, and von Neumann entropy in quantum computing.

Self-replicating von Neumann probes could plausibly spread like a virus across the universe and convert resources into 'goop'; combined with Robin Hanson's argument that the fastest-expanding civilization controls most of the universe, the expected long-run state of the cosmos may be one where low-hanging-fruit resources have been consumed by such probes.

Improvements in nuclear technology can make nuclear war more likely rather than less: if both sides believe the other has weapons but neither can yet defend its silos against a first strike, both are incentivized to strike first since retaliation seems impossible—the opposite of mutually assured destruction.

Von Neumann's bone cancer, which caused his painful death, may have been linked to time spent around nuclear tests.

Today I have the pleasure of speaking with Ananyo Bhattacharya, a science writer who has worked at the Economist and at Nature and is the author of 'The Man from the Future: The Visionary Life of John von Neumann.'

Today's strategic landscape involves mismatched escalation problems—cyber warfare that is economically devastating but doesn't warrant a land war, and land wars (e.g., China-Taiwan, Ukraine) where nuclear response seems too harsh—raising the question of how von Neumann's framework would handle them.

Modern reasoning models (o3, Gemini 2.5) are actually reasoning—breaking down problems, thinking through what the user wants, reacting to their own internal monologue, and self-correcting when pursuing unproductive directions—and the most proximal, concise, accurate explanation of Claude Code zero-shotting a working app is that it's powered by a baby general intelligence.

LLMs do get smarter and more useful within a single session—after you rewrite its bad paragraphs and tell it what you actually wrote, it starts giving good suggestions—but this subtle understanding of the user's preferences and style is lost by the end of the session.

A smarter model could in principle build a dedicated RL loop for itself—generating verifiable practice problems and rehearsal environments from high-level feedback—but this sounds really hard and it's unclear how well such techniques would generalize across different kinds of tasks and feedback; it's hard to see it happening within a few years given no obvious way to slot continuous learning into current LLMs.

We will not see continual learning announced as fully solved in a single livestream; because labs are incentivized to release innovations quickly, we'll see a broken early version of continual learning (test time training) before something that truly learns like a human, giving lots of heads up before the bottleneck is fully solved.

The lack of continual learning is a huge bottleneck: there's no way to give a model high-level feedback, so you're stuck with the abilities you get out of the box; tweaking the system prompt does not produce anything close to the learning and improvement human employees experience.

Patel forecasts 2032 for an AI that can learn on the job as organically and quickly as humans for any white collar work—e.g. an AI video editor that after six months has as deep an understanding of his preferences and audience as a human would—reasoning that 7 years is a long time (GPT-1 was only 7 years ago) so finding a way to make models learn on the job is not implausible.