A thirteen-month field test of teaching an AI agent the receipt-collection work software never automated: which seams exported to prose, files, and code, what stayed human, and what that says about white-collar replacement.
Yann LeCun’s JEPA doesn’t reconstruct the world pixel by pixel — it predicts, in representation space, what a thing is. Why that shift matters for representation learning, compute, and world models.
A24’s Backrooms works best when you stop reading it as dimension-hopping horror — and that’s also exactly why it stops short of greatness. 6/10.
Writing skills while you work is compound interest on your labor. But what actually compounds isn’t the pile of skills — it’s the judgment of what’s worth encoding.
LLM CLI is the new REPL for agentic applications: explore interactively, export repeated and fragile behavior into skills, promote stable capabilities into MCP tools, and end with a small app runtime over reliable tools.
Sachi and Tamotsu are both lovable — and that is exactly what makes this film cruel. Fifteen years of a marriage undone by timing, capacity, and luck.
