AI makes output cheap; review discipline decides whether you ship slop.

The core claim. Gabriel Martinez, engineering manager at G2i, defines AI slop as output that looks finished before the thinking is finished. Ambiguity gets cleared with guesses, trade-offs get left to chance, and no one takes ownership. The serious danger is not a bad machine choice. It is that an engineer never notices a choice existed, so the unknown unknowns stay hidden. Software is a record of choices, and teams that stop seeing those choices stop owning the system.

The cost shift. Generation cost has fallen; understanding cost has not. Martinez borrows Dijkstra's line that code should be counted as lines spent, and the SQLite creator's reply that an incoming pull request means maintaining, documenting, and testing it for years. With agents, a vague task becomes a large PR and a small feature becomes a pile of files. Output arrives with the shape of completion, while the actual thinking is handed to the reviewer. Evaluation, not generation, becomes the bottleneck.

The incentives. Martinez describes a familiar dynamic: careful engineers look slow, fast reviewers look productive, and performance anxiety pushes people to merge code they have not fully judged. He admits doing it himself. He argues the metric system is broken when PR throughput counts more than value, and that teams should reward well-factored design and clear test harnesses instead. Multi-thousand-line PRs are a warning sign nobody can review. This is the strongest part of the talk, though it stays anecdotal.

The guardrails. His answer is conventions, documented flows, small reviewable surfaces, and clear ownership. Diagrams matter as compressed review layers for service boundaries, data flows, and failure modes. Rails serves as the model because convention over configuration cuts the decision surface and lets engineers move across projects without re-learning intent. He adds that JavaScript environments with one obvious way are better for agents, because competing styles confuse both humans and models. The reasoning is clear, but evidence is argument, not measurement.

The product pitch. The talk closes with ORC, a G2i workflow built from these lessons: planning, review, adversarial checks, and context management. Martinez says it is not a magic box and does not remove human responsibility. He frames the future as orchestration with judgment, not better vibe coding. No benchmarks or production results are shared. The pitch signals direction without proof.