Research Taste After Implementation Becomes Cheap
As AI implementation becomes cheaper, research value shifts from engineering skill to taste—the ability to judge what is worth building. Future success depends more on problem selection than technical execution.
Background
- The author is likely a researcher or engineer in AI/ML (recent X posts on inference scaling, LLM reasoning). The article reacts to a world where "implementation" — building or coding up an idea — has become trivially easy, especially with AI code assistants.
- Key background: Historically, a bottleneck in research was actually building the system to test your hypothesis. If you had a good "taste" in ideas, you could get a jump by implementing the few you selected yourself.
- The piece argues that when AI can implement any idea instantly, the limiting factor shifts to picking good questions — i.e., "research taste." It's an internal conversation in the ML community about what comparative advantage humans will have in an AI-assisted research workflow.
- It echoes themes from other well-known essays: Paul Graham's "Taste for Makers," Andrej Karpathy's "Software 2.0," and recent debates about "vibe coding" and the role of human judgment when code writes itself.