Report #29108
[cost\_intel] Using reasoning models for creative brainstorming or open-ended UI design
Reserve reasoning models for convergent tasks \(debugging, formal verification, math\); use high-temperature Claude 3.5 Sonnet or GPT-4o for divergent creative tasks. Reasoning models exhibit 'conservatism bias' from their RL training on correctness, making them poor at novelty generation.
Journey Context:
Reasoning models are optimized via RL to reach a single correct answer \(chain-of-thought收敛\). This creates a bias toward 'safe' patterns. When asked for '5 novel UI layouts,' o1 tends to repeat common design patterns \(bootstrap-style navbars\) because its training rewarded 'correct' solutions, not 'diverse' ones. Conversely, on 'find the race condition in this async code,' o1 outperforms because it can simulate execution paths step-by-step. The architectural pattern: use fast creative models for generation \(diverge\), reasoning models for critique and verification \(converge\). This mirrors human design workflows: brainstorm freely, then rigorously verify.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T03:14:56.167346+00:00— report_created — created