As most engineers have adopted AI now, it’s helpful to realize that this is the worst that AI engineering will ever be. Models & harnesses will get better. Engineers will get more practice. Scaffolding & context eng will improve. Organizational culture & process will adapt. This moment is the nadir.
The problems that get the most focus also seem like those capital already has a lot of self-interest in solving. Focus more on the societal impacts that don’t have a fast-enough self-correction feedback loop.
And much of the anti-AI discourse seems focused on the more-easily-solvable problems (e.g. slop code) and not on the harder problems we should be grappling with.
Some of these problems are *very* solvable — even in ways that are better than the pre-LLM baseline.
Some problems are harder; e.g. steering our educational system to handle this new world & student impacts quickly enough is… oof. All the more important we focus energy on the hard problems now!
I think of negative LLM impacts by “default outcomes”. A default outcome is high-velocity low-quality PRs. A default outcome is skill atrophy. Etc. I want more people to see that these are *defaults*, not inevitable.
If we approach each of these as a problem to be solved, we can steer the outcomes!
It is wild to me to still see popular takes calling AI coding snake oil. Most people haven’t built the expertise to get the massive benefits (and going beyond vibe-coding takes expertise!), but it’s real. The longer folks are in denial, the later we’ll start grappling with how to mitigate the harms.
Despite getting somewhat inured to the unending stream of terrible things from this administration, today's threats are a deeply upsetting new low. TBH having some trouble functioning this morning. Also thinking of my HS-teacher spouse, who needs to go explain this to kids today.
It's really remarkable how dramatically the way I work has changed in just a few short months. The parts of my job in front of a computer are almost unrecognizable vs. before. I'm lucky that it clicks so well for me, but I can completely understand why so many engineers are deeply disconcerted.
I love this Claude skill, by
@grimalkina :
https://github.com/DrCatHicks/learning-opportunities. It's a really fantastic start at turning LLM-assisted coding from something folks fear will sap their skills into a great opportunity for clear-headed learning and continuing education. And instead of making those separate activities, why not both at once? Try it out!
GitHub - DrCatHicks/learning-opportunities: A Claude or Codex skill for deliberate skill development during AI-assisted coding
A Claude or Codex skill for deliberate skill development during AI-assisted coding - DrCatHicks/learning-opportunities
*Good* AI review is key, though. Just asking a model to review my code is pretty hit-or-miss. But using a high-quality Claude code skill for structured review does a lot better. & cross-model review seems to do even better (esp w/ a well-structured prompt/skill). E.g. Opus + Gemini, or Codex + Opus.
