Over the last week, I’ve used AI to, among other things:
- Write ~70kLOCs of integration and E2E tests for a codebase with hundreds of kLOCs and zero code coverage
- Have my PRs reviewed.
- Have AI address AI-generated PR comments.
- Debug product issues, GitHub Actions issues, and improve our observability.
- Research technologies and product categories to inform engineering decisions.
- Generate Anki cards for topics I’m studying using my Anki style guide.
- Create memorable D&D NPCs, create unique images for them, then turn them into 3D models for 3d printing.
- Learn Blender well enough to sculpt and slice those models.
- Build a Claude Code-like agent aware of my Obsidian vault to edit D&D synopses.
- Get coaching feedback on my communication style.
- Get useful comments on an important self-help topic involving spirituality, psychology, and engineering. I don’t know anyone I could currently talk about this topic.
As many have noted, even if AI progress stalled for 5+ years, it will take us a long time to max out the capabilities of what we already have.
I have no idea where are we going to end up. Last year I was using AI coding assistants as fancy autocomplete. Now I regularly manage 3-9 Claude Code agents that are independently solving as many problems as I can coordinate, not just bug fixing but also production issues, observability, CI, data analysis, documentation, everything. They even respond to my jokes on GitHub comments. I’m getting better at specifying problems, making them more and more autonomous, and coordinating them. And the models are becoming smarter. I only have to check Twitter to see the next steps ahead: I’m regularly seeing people building autonomous ai agents that are monitoring kubernetes clusters, production systems, monitoring Github issues and taking on them.
I loved doing many of the above tasks all by myself, but change and evolution are intrinsic forces of nature that we can’t escape and there is also beauty in being able to do higher level moves like unlocking a team’s velocity by going from 0 to nearly complee code coverage on a mature service, or creating tactile, 3d representations of imaginary characters that my friends and I love playing with.
Main risks I see:
- Growing dependence on fewer single points of failure for SOTA intelligence, at least until Opus-4.5-level models are realistically self-hostable.
- The digital divide.
- Monkeys with digital equivalent of nuclear warheads.
- Abuse: trust & safety, black hat hacking, scams.
- Add robotics, and a few more generations of 3d printers to the mix, then look 3-4 years out. What will life be like?