rgambee

I agree with the central point that coding agents are sticky. We're unlikely to go back to a word where humans are writing the vast majority of code themselves. When I think about whether AI is over-hyped, I return to the fact that even if AI progress stalled tomorrow, the abilities of today's AI still haven't saturated our world.

I'm more skeptical that IDEs will be dominant a few years from now. It's certainly possible. But I also find it plausible that AI continues to do more and more software development work, and humans less and less.

> the tokens are already starting to come less cheap

It may be true that frontier models have become more expensive at times (there are counter-examples though). But the stronger trend is the cost of running a model with fixed capabilities falling over time. That can't continue forever, of course. But I don't expect the fundamental costs to rise. Are you expecting providers to significantly increase their margins?

I've heard stories lately of open source projects being inundated with vulnerability reports and PRs. But in this case, it seems like AI assistance was clearly a boon for root-causing and reporting this so quickly.

Looking forward to a Veritasium video about this in the future, like the one they recently did about the xz backdoor.

I'm curious what the verification and proof stages offer beyond automated PR checks. The verification example looks rather standard.

To me, the biggest bottleneck is the fact that automated checks don't catch everything. I still need to read through the code myself before I trust it.

I'm curious what it would mean to align the interests of an AI agent with a natural ecosystem. Alignment to human values is still an unsolved problem. Alignment to an ecosystem has a lot of the same challenges as alignment to human values, like balancing diverse and often conflicting interests (what's good for a rabbit might be bad for a fox). But it has its own challenges, like lack of data.

You can add a hook to your Claude Code settings to find and stop Docker containers. But it's risky since it might stop containers being used by another CC session.

Another option is to manually clean up Docker containers periodically, for instance by searching for all containers using a particular image:

    docker ps --filter ancestor=crystaldba/postgres-mcp

I definitely see a lot of these anti-patterns in the code that CC writes. Many of these can be caught at the time the code is written without needing to wait for a PR review. To me, it seems like most of these instructions belong in CLAUDE.md instead of or in addition to a specialized review skill. Are there things in the review skill that don't belong in CLAUDE.md?