My most common advice for junior researchers

A viral post from an AI researcher and Inkhaven Fellow outlines the most common advice given to junior collaborators: perform quick sanity checks before diving deep. The piece argues that researchers often waste weeks or months on fruitless investigations that could be avoided by first validating core assumptions, checking for data bias, and ensuring their experimental setup isn't fundamentally broken. The advice is framed as a counterbalance to the natural tendency to overcomplicate or overlook basic flaws in the excitement of a new idea.

The post provides concrete, technical examples relevant to modern AI research. It suggests checking if language model agents in a scaffold are actually making successful tool calls, or if the reasoning chains in an LLM experiment are functioning as intended. A specific example involves analyzing why an LLM might fail a complex reasoning task like the n=10 Tower of Hanoi—not due to a lack of capability, but because it finds the task "extremely tedious and error prone" and refuses to engage. The author promises follow-up pieces on the other two pillars of their common advice: saying precisely what you mean and asking "why" one more time.