Anshuman Biswas

Totally Chris, that's a genuinely uncomfortable extension of my own argument — and I think you're right to push it. I hedged. "Never zero" was comfortable to write because it felt safe. But you've identified the actual fault line: if AI reaches the point where it can solve Millennium Prize Problems — problems where we don't even fully understand the solution space — then "you'll always need humans who can verify the logic" stops being a guarantee and starts being wishful thinking. The honest answer is: I don't know where the ceiling is. Nobody does. What I do believe is that we're not there yet, and the path between "very good at logical reasoning" and "capable of novel mathematical discovery at that level" is longer than the current hype cycle suggests. But "not yet" is very different from "never" — and I conflated the two. We'll get there someday! The deeper point you're making is essentially: if the verification itself can be automated, the human role collapses from "understands the logic" to "understands that the requirements were correctly specified." Which is a much thinner foothold. I'll sit with that. It's the most honest challenge to the piece I've encountered.

This is exactly the kind of pushback I was hoping for — thank you. The tests vs. architecture distinction is on point. Tests tell the agent what, architecture tells it where and how — I'm going to be thinking about that framing for a while. You're right that the real leverage isn't in cloning existing logic, it's in extending it. And extension without architectural clarity is just chaos at scale. On agentic engineering vs. vibecoding — fair. I think we're in that uncomfortable middle period where the terminology hasn't caught up to the practice. What you're describing (structured context, explicit rules, human review) is meaningfully different from the prompt-and-hope era. It deserves a better name! The world model point on AI slop is the most interesting challenge to the evolution analogy. You're right — evolution is the real world. LLMs are pattern-matching against this real world abstraction. The test suite and the experienced human aren't just quality filters, they're the grounding mechanism. That's doing a lot of work that we probably undercount. This line is absolutely killer: "The tools have changed. Good software engineering hasn't." SOLID, DDD, TDD — those aren't legacy ideas, they're the scaffolding that lets agents reason cleanly. Couldn't agree more. Thank you for reading!