Thariq Shihipar’s closing caution, aimed squarely at his own audience of AI engineers: “building is easier, but generating value is still hard. I think this is something we run into as AI engineers sometimes, where we think so much about the process of building and our setups… but the point is to generate value… it takes a lot of swings, it takes a lot of tries, to find the valuable stuff.” The implicit failure mode: mistaking a shrinking build-time cost for evidence of progress, when velocity in construction was never the actual bottleneck — recall connects why he made an entire conference deck in four hours: fast building is now cheap and repeatable, but that speed only counts for something once you’ve found which attempt is the valuable one.
This is a check on Don't presume the tradeoff — do all of it and force reality to show you where the real constraint is — attempting everything is only worth doing if you’re honest that most attempts still won’t be the valuable one, and the point of doing more is to increase your shots at finding it, not to congratulate yourself on shipping more. It’s also a variant of Verification is the single highest-leverage practice for agent-assisted coding one level up the stack: verification tells you the code works, but nothing in that loop tells you the thing you built was worth building — that judgment still has to come from outside the construction process itself.