Aug 5, 2026

An essay on whether skilled prompters exist when everyone talks to the same models. It points to Terence Tao pulling discussion out of ChatGPT that ordinary users cannot, and argues the real prompting skill is domain expertise, not phrasing.

Why it matters: The paradox that LLMs turn everyone into a generalist while raising the value of deep expertise. Read as a debate about where developers should invest their learning.

A 110-word post from 2024 resurfaced into a huge thread: an AI-generated hero image on a personal blog makes the reader suspect the text is AI too, and stop reading. A crude hand-made drawing proves a human better.

Why it matters: Evidence that reader aversion to AI content runs stronger than writers assume. Even for indie devs, one AI image can tax the credibility of everything under it.

A Swift + Metal runtime that keeps only the dense core of an MoE resident and streams routed expert weights from storage. The 80B runs on a Mac in 4.3GB of RAM at 4.5 to 5 tok/s, and the 35B runs on an iPhone 17 at about 1 tok/s.

Why it matters: A realistic path to huge MoE models on local hardware, with the honest caveat that only about 3B parameters are active per token. Weights published under Apache 2.0.