An idea from Keith Adams on Still Burning, and neither of us thinks it's settled: If software quality is proof-of-work, the moat isn't cleverness. It's gigawatts. Keith calls it the frontier labs' quiet bet. Maybe true, maybe not.
Keith Adams on Still Burning: programming lost its flow state. The agent world feels more like air traffic control. What I miss is feeling oriented, a sense of mastery. That loss is emotional, and when software gets fiddly, practical too.
I asked Jessica Kerr what keeps her up at night. "Two kids to put through college, and I can't say I'll have a job in any city that pays what I want. The market is different now." Her answer: demand will surge, and someone has to design the IKEA furniture. So get great at that.
Jessica Kerr's pinned tweet, and the heart of our Still Burning episode: "I don't want to build software so much as build understanding and express it in software." The new craft: designing a system that produces working code and being able to say why you know it works.
"Human in the loop" — until the loop becomes a noose. Jessica Kerr on Still Burning: if you try to understand everything the agent did, you're stuck. Don't verify the work by hand. Make the agent build the thing that verifies itself.
Agents can forget. People can't. "Parenting works like this now? 'That didn't work, slash clear. Rewind. Reload my game from before you came in the door.'"
Programming isn't going away, it's splitting in two. Jessica Kerr on Still Burning: "If you want to build a bookshelf, that's a craft. But there's also IKEA." The code-crafting got commoditized. Knowing what to build didn't.
experienced engineers: one change, test, one change, test junior engineers: batch everything because they're in a hurry this is exactly backwards the person least capable of batching is the one most likely to batch
Slices fail when they're too small to provide feedback. But no amount of features fixes a flawed premise. Go smaller, but ask questions with what you build.
I've been trying for years to get a decent shot of a bird of paradise flower. I finally got one. Still not perfect but that's what the quest is for.
How Long Should Functions Be? Hint: The Answer's a Distribution, Not a Number tidyfirst.substack.com/p/how-long-s... I'd love to be contradicted with data.
One way to look at tests is as a filter that lets (most) good changes through & keeps (most) bad changes out. Tests make (most) mistakes reversible. Layer these filters & you get better discrimination at the cost of longer delays & more maintenance.
I'm not anti-metric. I'm anti metric abuse. Data mostly asks questions, not answers them. Here's an example of using data to ask questions about who are influential programmers on a project. tidyfirst.substack.com/p/measuring-...
Here's a sequence I haven't heard discussed. You have a system with tests. You break the system into parts. You add tests for the parts. Now you have redundant coverage. You have the opportunity to trim the system-level tests without losing any confidence in the behavior of the system. Do you?
"Space Needle", 12" x 4", acrylic on glass. I love painting scenes with strong emotional connection, in this case my parents taking me to Seattle frequently as a kid.
Untitled, 10” x 8”, oil pen on acetate. One of the cool things about drawing on clear plastic is the background completely changes the picture. Also, new gear inspires—I got the Overseas PMA-520 pen set.
In the 3X: Explore/Expand/Extract model people ask about the transition from Explore to Expand. How do you know when it's time? Answer: The market tells you. Like this. You can't force it.
Untitled, 150 cm x 45 cm, acrylic on glass. These abstracts often start with a single gesture, in this case the sort of squid-looking figure on the left. The rest fills itself in around that.
Untitled, acrylic on plexiglas mirror, 28cm x 23cm, 2022. Of my abstracts people ask, "What does it mean?" The best answer I've found it that this is what the inside of my head looks like. What does it mean? Everything & nothing.
Deepgeek. I found a new fat-tail distribution. If a line of code is deeply indented, well, it's more likely than a random line to need to be indented further. Voila!