What seem like great initiatives are being watered down because nobody can keep up, debugging issues takes so much longer because everything is changing at once, and everyone is exhausted and hardly talking to each other which feeds into a cycle of having no idea what is happening.
We actually talk more now which helps, but it is still hard to keep up when everyone is barreling ahead doing their own thing. In addition to more talking, there needs to be a semblance of strategy that everyone is aligned on and understands their role in.
A high-agency, high-functioning team has always been a superpower, but this capability is what will make or break organizations that are trying to run lean with AI. It's a "people problem" at its core, and no amount of technology can fix that.
It's not for things where you can just ask some expert to tell you what works or decide for you.
It is permission to trade inaccuracy for autonomy.
In other words, privatized profits and socialized costs. Again.
But slow laps also look slow.
"Move fast and break things" is about conquering the second kind of slow. Not idealizing breaking things but not being legitimately slow tied up in bad attempts not to break things.
Step two is being slow in the right way.
Be thoughtful, be methodical, be aware, be comfortable, and be decisive. Made a lot of sense when I caught a 2-hopper off the line at 3rd and didn’t have time to think about how to field it or where to throw.
Time and time again proven true.
You can spend 100M tokens/week and generate something that is good enough for end to end demo to paying clients in 1-4 weeks depending on complexity of the project. Doing this feels like being on drugs, in that the creative process is a high, and that you will be mentally exhausted at the end of every day (the crash).
Inventions that were initially useless but found application later, are still in the very small minority.
The way to view LLMs is that it is a better google search. This allows you to speak to coworkers only when they have the context you need. For other trivial things there’s AI.
Optimally you must only disturb your coworkers when necessary.
These guardrails are precisely what should be laid down in advance to enable workers to run safely with AI. Write all the rules in your AGENTS file, and point your AI reviewer at it. Encode whatever you can describe algorithmically in commit hooks. This will get you 90% of the way there, and peer review will take care of the rest.
I am hopeful that AI will empower smaller companies, where there is less deadweight, and consensus can be formed more quickly. Discussing what to build is not wasted time; it's one of the few things that favors humans.