AI Reading for Saturday September 27

Author

Druce Vertes
September 27, 2025

The problem is, when machines become human-like, they compete more with humans as opposed to being highly complementary, and this hurts labor when capital is plentiful.

That’s a simple way to put it. But if you have unit elasticity of substitution between capital and labor which is the simplest Cobb-Douglas production function, then (with other assumptions) capital and labor split the national income 50/50. If capital gets cheaper, you substitute capital for labor, labor is more productive, gets a higher wage, hours worked may go down, labor share of income stays constant. But if you don’t have the Cobb-Douglas level of diminishing returns to capital because it’s a closer substitute for labor, it can crowd out more labor and reduce labor’s share. Something like that.

Sounds great but a lot of intrusive surveillance starts with, won't someone please think of the children.

'Leak' lines up nicely with Meta's public claims and comes, after they failed to timely respond to Congressional inquiry

Experiments were done on non‑pathogenic E. coli at approved biosafety levels with filters to exclude human‑virus sequences, and the work raises urgent governance and dual‑use safety concerns due to biology’s unpredictability.

Agents require significant development, onboarding, and continual feedback to avoid "AI slop" and loss of user trust; for low-variability tasks simpler rules-based automation or LLM prompting may be preferable.

It seems obviously incorrect to say there is “no” world model underpinning LLMs, but it's often a shallow and idiosyncratic one.

The era of hype and BS, but also on the haters’ side sometimes.

Recycling gets an AI upgrade - Packaging Gateway

hell of a sendoff

Follow the latest AI headlines via SkynetAndChill.com on Bluesky