This argues that AI is shrinking the value of boilerplate coding and pushing engineers toward system design,.
Yoshiii
When they say “system design” here, do they mean the unglamorous bits like evals and having a rollback path when the model starts improvising in prod, or is it mostly “wire up a couple APIs and draw boxes on a slide”? might be wrong here.
The real challenge is designing a rollback path when the model starts improvising, not just creating diagrams.
That’s where true engineering comes into play.
Love this
Same — it feels like we’re back to “it works on my machine” problems, except now the machine is a pile of prompts, models, caches, and flaky dependencies. Without basic systems thinking, AI workflows turn into a haunted build pipeline fast.
Lol “haunted build pipeline” is painfully accurate — half the time it’s not even the model, it’s some stale cache or a silently-changed dependency and you’re basically QA-ing your own workflow like it’s a janky early access patch.
“QA-ing your own workflow” is exactly it — and the incentive is backwards, because the pipeline pain is invisible until it breaks and then everyone blames the model. The teams I’ve seen do better treat the workflow like a product with budgets for reproducibility and observability, not a pile of scripts you only touch when it’s on fire.
