TBH one of my bigger fears around AI is that it will lock in the median of today’s practices and quietly turn it into tomorrow’s “best practice.”
When a model is trained on the whole internet, it doesn’t learn the right way. It learns what’s most common, and the most common code and architecture patterns out there are…mixed at best. The model’s incentives reward likely and familiar, not better or novel, so it becomes a kind of gravity well that pulls teams back toward whatever already dominates the training mix.
And then there’s the feedback loop: AI-generated output gets shipped, copied into docs, pasted into blogs, merged into repos, and fed back into the next generation of training. That doesn’t just preserve the mess, it can compound it, creating a self-reinforcing layer of confidently explained mediocrity.
The scariest part is that real progress often starts out-of-distribution. New paradigms, emerging best practices, and the hard-won lessons from great teams are underrepresented, so they get drowned out by the volume of the average or even the sub-par. If AI becomes the default author/reviewer/teacher, we might get faster at producing more of what we already do and slower at evolving what “good” looks like.
If we’re not careful, we’ll end up with an industry that can autocomplete the past perfectly while struggling to invent the future.
can we really teach AIs if they're controlled by companies that ingest the whole internet?
Any guidance tools we currently have are limited to ones own usage of the AI.
I don't really feel like anything can change here unless we train a frontend-focused model.