I genuinely appreciate that the YouTube algorithm surfaces inane, extremely-low-view-count videos through the Home feed and Shorts. It feels like public access television.
https://www.youtube.com/live/j_IjQbuZfYs?si=fvzgc6yDTG6jGqrB
YouTubeWell, great, my job Googling “disregard” is now taken by AI.
The way we will deal with the Einstein in a Sweatshop problem is by making sure that no human is the smartest thing on earth and that intelligence is infinitely replicable. Then we can have all humans work in sweatshops without consternation.
It’s hard to articulate what my agentic programming workflow looks like to use all of these tokens. I would say it is weighted toward hardening and verification: I spend a lot more time polishing, whittling, and testing than I do generating. Here is what Codex analytics shows for skill use.
The naive, bitter-lesson-backed response to the question of “where are all of the hit vibecoded apps is”: there won’t be any for a while, then one will break through, then it will be most, and then we’ll be tripping over them. All accelerating. That’s just the way of everything digital.
Helluva day in the mines. Back at it again tomorrow.
I am really captured by the scale we’re about to hit with LLMs. Right now programming agents can produce on the cusp of superhuman performance: a solid few thousand lines per day. We can’t be but a couple years away from 10-100x that. It will be incomprehensible and feel like superintelligence.
We are in the twilight of human-computer comprehension, measured in months.
What really blows my hair back about programming agents right now is their tenacity: it is a terrifying marvel to watch a computer autonomously operate itself for hours to solve a problem of any difficulty level within itself. One example is the willingness to muck around with SQLite databases.
