A model-choice heuristic: pick a model based on whether it will disagree with you, not just on raw capability or agreeableness. Ankur Goyal and his co-host both reach for Codex specifically on hard technical problems, describing it as, at the time, the only model that will regularly push back on the user rather than defer to them.
When a problem is hard enough that an agreeable answer could hide a wrong assumption, deliberately choose a model known to disagree rather than one optimized to please the user — useful for the adversarial validation that Agent-Driven Exhaustive Experimentation (Practical vs. Theoretical Quality) depends on.
"I think it's currently the only model that will disagree with you regularly. And I think if you're working on hard problems, it's very important."
Agent-Driven Exhaustive Experimentation (Practical vs. Theoretical Quality)
Beyond picking a model known for pushback, Ryan Nystrom uses a prompting tactic to surface genuine reasoning instead of blanket reassurance: he tells the model "you're wrong, defend your argument," forcing it to either produce a substantive justification or concede the point, rather than reflexively agreeing. He applies this especially in domains where he says he lacks deep expertise himself (e.g., CI internals), using the model's forced defense as a way to sanity-check claims he can't independently evaluate ("I literally don't know what I'm doing here. You got to explain it like I'm a five-year-old.").
Nystrom extends model choice beyond raw pushback to "personality" fit for the task at hand: he describes Codex's blunt, non-affirming tone as well suited to "explain it to me" debugging sessions, while he finds Claude Code's warmer tone better suited to other modes of work — a task-fit judgment, not a general capability ranking.
He separately claims Codex tolerates long-running, context-heavy tasks better, saying it "can grind for hours," whereas he says Claude Code "would just kind of like lose the plot really quickly" once its context window fills — offered as his own first-person comparison rather than a benchmark.