Agents can write code faster than we can absorb it. Here’s why it still matters for humans to understand what they build — and some techniques for doing that efficiently: explainer docs, quizzes, micro-worlds, and shared spaces.

What lets the same algorithm work across all three domains is that they share two key properties:

1. Closure under negation and intersection — so the difference between two candidates is itself expressible.
2. Sampling from that difference — so the system can show the user concrete cases where the candidates disagree.

I built a custom feed for Bluesky that trains a model on my past interactions. Here’s what I learned about RecSys and social media feeds generally.

Hype.

let’s talk

Very beautiful hypermedia desktop environment powered by Patchwork. At the same time, I’m a bit disturbed by how involved LLMs have become in the I&S process.

Mark Weiser gets it. He understands exactly why the current gen AI/LLM moment is so frustrating to me. We should be building better software, not creating more software to control our existing mid software for us.

For example: being alerted of a potential collision

• agent: “collision, collision, go right and down”
• ubicomp: background presentation of airspace information for continuous spatial awareness, as in everyday life. You’ll no more run into another airplane than you would try to walk through a wall.

Damn; he’s even anti-racist:

It obsessively fascinates

• the human-like machine to which we give life
• the perfect, all-powerful, slave
• be careful when appealing to ancient prejudice

via: http://geoffreylitt.com/2025/07/27/enough-ai-copilots-we-need-ai-huds.html

Funny how Litt takes away a totally different message from this slide deck than I do.