This is a collection of questions for which I don't have a satisfying answer. If you do, please email me! # AI - Can symbolic AI handle the [Sorites paradox](https://en.wikipedia.org/wiki/Sorites_paradox)? - Much of human activity is moderated by [social control](https://en.wikipedia.org/wiki/Social_control), yet these mechanisms are ineffective for managing AI, as current models are socially unembedded. For example, I would risk great consequence if I fabricated citations in a research paper, yet AI has no such exposure. How do we train models that care about their reputation? - Why are LLMs so bad at writing? How can we make them better? Why do [claudisms](https://tropes.fyi/tropes-md)" emerge? My hypothesis is that reading lots of text generated by a single human would have a similar effect. Take for instance prolific fantasy authors like Robert Jordan or Brandon Sanderson, [fans make fun of their repetitive phrases](https://www.reddit.com/r/Fantasy/comments/hxfxzs/does_anyone_find_brandon_sandersons_prose_a/). - Why aren't LLMs funny? This is what Opus 4.6 Extended Thinking comes up with when prompted for a joke: Why do programmers prefer dark mode? Because light attracts bugs. - Will model providers ever remove the "AI makes mistakes, verify important info" disclaimer? What differentiates technologies that require calls to oversight (e.g. self-driving) from those that don't? Is it just performance? Is it the possibility for catastrophic outcomes? CYA? # Philosophy - Why do people work so hard? Is hard work a moral good? - I know this question has been considered by Aristotle, Max Weber, Bertrand Russell etc. but I haven't gotten around to reading them yet as I've been too busy working. # Startups - I used to severely underestimate the role hype plays in startup success. To what extent is hype a self-fulfilling prophecy? Greater hype leads to more money raised which leads to more hires which (with large error bars) leads to a better product. Obviously hype isn't sufficient but is it necessary? Note that I am exclusively considering venture-scale startups. # Running - The women's 1500m world record is 3:48.68 by Faith Kipyegon[^1] while the men's record is 3:26.00 by Hicham El Guerrouj. How would their [rating of perceived exertion](https://en.wikipedia.org/wiki/Rating_of_perceived_exertion) compare after the first minute of their record-setting races? Intuitively, Faith's would be lower, because she has longer to run, but if so, why doesn't she speed up? This was a perennial debate on my high school cross country team. [^1]: Set in my hometown of Eugene, OR!