I struggle to believe that AI R&D is ingenious

The writer questions whether current AI systems can conduct genuinely ingenious research autonomously, based on experiments with models across several projects: building a top Scrabble engine, creating an Ultimate Tic-Tac-Toe bot, and researching philosophy and formal proofs. They describe recurring failures of judgment: models miss the most relevant literature, run experiments too small to yield useful comparisons, and spend extensive effort on trivial research directions. In the game projects, a model found a promising blog post but failed to recognize its importance, despite evidence that its approach performed well. The writer also notes that papers reviewed at a hackathon often shared obvious methodological omissions, suggesting a broader weakness in identifying straightforward analyses.

There is a notable exception: tasks with rapid feedback and clear measures of success. Repeated prompting helped improve a Scrabble move-generation system substantially, but the largest gains followed human suggestions about how to measure performance and divide the problem into cases. The writer further says model-generated reports can be so difficult to follow that interrogation is needed to uncover what was done and what mattered. Replies distinguish the limits of public models from their usefulness as research assistants, especially for implementing experiments, and cite claims that internal lab models are much stronger. Another commenter doubts that such claims have matched their own experience across successive generations.