How to optimize AI research questions?

The key to optimizing AI research questions is to be able to give direction as precisely as you would a bright intern - which reminds me of my experience last week using Gemini to analyze cryptocurrency trends for Q2 2024. I typed in "analyze cryptocurrency market" and got a general overview, but when I adjusted it to "compare the performance of SOL and ETH in the three dimensions of trading volume, institutional positions and on-chain activity in March-May 2024, with a focus on the impact of the US ETF policy. "After the harvest of a 38-page in-depth report with Bloomberg data citations. This shift is like going from fishing with nets to deep-sea directed fishing, with at least a five-fold increase in efficiency.

Replace the wide lens with a microscope

The most common mistake is to ask the question too broadly. For example, when analyzing the AI chip market, the question "What are NVIDIA's strengths now?" will put AI in a difficult choice - should it talk about the CUDA ecosystem, or the H100's arithmetic performance? Or the latest Blackwell architecture? And when you explicitly ask "Compare the energy efficiency of NVIDIA H100 vs AMD MI300X in LLM inference tasks using MLPerf benchmark data", the model instantly pulls up the most relevant technical whitepapers and benchmarks. A certain research group at MIT even found that questions that limited the scope of research boosted the number of authoritative literature cited by AI by 2,40%.

But scope control also requires skill. When I researched the generative AI copyright controversy, I limited the question to "the details of the data collection involved in the 2023 Getty Images v. Stability AI case," which resulted in systematically ignoring the entire industry context. Later, the report was adjusted to "analyze the representative position of this case in the discussion of the legality of generative AI training data and compare it with the relevant provisions of the EU AI Act," and then the report had real policy reference value. This lesson has taught me that limiting scope is not a matter of drawing a line in the sand, but of establishing a coordinate system.

Putting a Navigator on the AI

A recent report from the Stanford HAI Institute shows that questions that include a clear analytical framework can improve the accuracy of results by 57%. For example, to study self-driving safety performance, instead of asking "Is Waymo safe?", we should construct a matrix that "compares the safety records of Waymo Gen 5 and Cruise Origin in 2023 on three dimensions: crash test data, disengagement rate, and extreme weather performance". Instead of asking "Is Waymo safe?", we should construct a matrix: "Compare the safety records of Waymo Gen 5 and Cruise Origin in 2023 on three dimensions: crash test data, disengagement rate, and extreme weather performance, and ask for cross-checks that include the original NHTSA data". This kind of structured questioning is like putting GPS on AI, and I've personally tested that it makes a qualitative difference in the professionalism of the report.

Interestingly, this technique works especially well in business analysis. Last month I helped a friend research SaaS track, using the framework of "Compare the penetration strategy of Zoom and Teams in the enterprise market in 2023, focusing on the analysis of: 1) the procurement decision-making process of large organizations 2) the construction of the ISV ecosystem 3) the degree of openness of the API," and ultimately generated a competitive analysis of the company's CTO thought that he was a consulting firm. This reminds me: a good research question is one of the most important things you can do. This reminds me: a good research question itself should be an analytical model.

A little magic in expectation management

Have you ever had the awkwardness of an AI giving you a 20 page paper and only wanting a presentation? I've found that specifying the output requirements directly in the question is like choosing a cup size for a coffee machine. Instructions like "summarize in bullet points" or "generate a 5-page PowerPoint outline for board presentation", combined with Gemini's report length control, can save a lot of post-editing time. Even better, if specific formatting is required, such as "Organize results into a SWOT table" or "Label all cited DOI numbers," AI is now able to perform these tasks - basically flawlessly! -It's much more efficient than organizing them manually after the fact.

Finally, a counterintuitive finding: sometimes it's better to put a "human face" on a question. For example, when I was researching the ethics of brain-computer interfaces, I tried two kinds of questions: the stereotypical "List the ethical risks of neurotechnology" and the curious "If Neuralink is approved for civilian use next year, what three types of social controversies do you think are most likely to be triggered? Please analyze it from the perspective of a science and technology ethics expert". The latter even cited more cutting-edge discussions in academia, and even included a few preprint papers. It seems that even AI likes to interact with warm questions.