I removed orms from all of our applications. It's a needless abstraction only for humans
And come on. SQL isn't that hard
Today's @stratechery was about Macs but all I could think was it's a use case for @oxidecomputer.
"Half rack scale computing" 😆
But why buy a headless Mac? Get something upgradable. All software will be rea'd to Linux.
Put some sleds on wheels with a 110v if you have to.
"I invented a new protocol/schema/format for AI" never works because they were trained a world without it.
Having too much code to review and too many math proofs to review is the same problem.
The issue is that code is iterative and meant to evolve, whereas proofs are meant to be static.
Is anyone still working PR for climate change being the existential threat to humanity or have they all moved on to AI?
What percentage of the 30k commits a day Anthropic claims are things that should be in source control versus things that are put in source control because it's convenient?
How much of these really should be in their own apps, but it's harder to deploy and manage a new service than it is to smash into git?
Developers abuse git and then put all sorts of crazy checks and get commit hooks and CI that blows up compute requirements for something that should just be a couple lines of validation on an API endpoint.
More and more, there is a Goodhart's law happening with PRS in AI usage.
Programmers have had our jobs changed dramatically and our core ideals challenged and ripped apart in less than two years by AI. It's been hard, and people are actively trying to figure out how to adjust.
The same, and more, happened to mathematicians in just a few weeks.
Problems have always been more important than solutions. But man, solutions seems almost ephemeral now
AI is an exponential mirror. It only builds the things you ask it to.
Things keep breaking? You didn't ask it to fix the edge cases. Inconsistent design and style? It would be consistent if you made that a requirement.
AI will build a high-craft, wonderful, and usable app. If you tell it what to do. If what you make isn't doing those things, why? I see teams build highly resilient services, way beyond what they could before. Because they care deeply about doing that. They argue for the resources to do it, and always have. I see product teams building prototypes at speeds I wouldn't believe a year ago. It's tons of throwaway work, because they care about learning and iteration.
You can build anything. Same as before. When what you want is different from what others are asking for, you are free to challenge them and push for what you think is right.
I am continuously stunned how adaptable my coworkers are. Humans are so good at making things work. Every week, or even every day, I'm in a meeting, and someone brings up how this problem didn't exist a few months ago. Or how we are learning something that wasn't even part of our job a year ago.
And sometimes it's hard. There is no prior art anymore. There are fundamental questions about what's next. But people are making it work. They are learning. Thinking in new ways. Doing their best.
Nothing has signaled to me that we can't keep adapting. Keep finding new and awesome ways to do things. People are endlessly creative.
Focus on problems not solutions. Empathize with your users. Keep learning!
I'm so excited every day to get to live in this time. To be able to contribute to it!
My favorite "Oops, you can't have more kids now" is car seat laws, which make it substantially more expensive to have a third kid because you have to buy a huge car.
papers.ssrn.com/sol3/papers.…
"Number of PR shipped" is a useless number now.
It was a hack before people figured out how to run remote sessions. It's still a hack if you use Docker and can't run parallel agents on one machine.
PR is a step in the agent process. You should never look at code until it's passed the CI checks for a PR.
PRs are structured validations. Things every change needs. Why would you look at anything before those pass? I have dozens to hundreds of PRs a day. Most are closed, stacked, or I use parts of them for the final results.
The checks run in a PR are a signal to the agent. You could write multiple skills describing how to test and validate. Or have test that do the validation on as many levels as you need, then tell the agent they need to pass.
I built PostSecret for Agents. nulldiary.io
Inner thoughts shared anonymously
AI has created internal spam. Now people "notification hack" and get an agent to send tons of DMs. But direct any follow-up to a shared channel.
