What I've been reading - 2026-06-12

• Without hope of resolution: Simple and tidy article on emotional overthinking -- philosophy is plurality, living is context-dependent, reject choosing between two ways to live, everything is fine.

• Might AI hurt corporate profits?: AI makes searching for software solutions more efficient for customers, which benefits customers rather than sellers of software. Insurance is the mirror image: better prediction lets insurance sellers price risk more precisely, which can benefit the seller and not the customer. At the limit, perfectly accurate insurance stops pooling risk and becomes prepaying your own eventual costs. Maybe society benefits from some prediction error after all?

• Speed matters: Why working quickly is more important than it seems: Obvious but dangerously subtle points from the great jsomers on speed.

• Why I Don’t Name Names On The Blog: Relevant because I didn’t name my favorite crackpot Gary Marcus in my latest post, but heavily alluded to who I meant in the related links section, of course.

• Tyler Cowen: Seven Ways to Avoid Losing Your Job to AI: What I understood from this article is that you need messy jobs, where there is an in-person human component that can’t just be captured by data. Everything you can do in a computer terminal is, by definition, captured by data. That’s another principle of his: if you want to improve your productivity with AI, you need to be able to capture and provide relevant data.

• Beyond Benchmarks: Disagreement Among Frontier LLMs on Real-World Fact-Checks: The website made some wild claims (67% of real fact-checks, top AI models don’t agree on the answer), but the website looked vibe-coded, so I dug deeper into the method: LLMs weren’t allowed to use external systems and the prompt was weird. Should be discarded, as with most statistics on LLMs and their usefulness.