Don't Let AI Converge on Identical Answers

I compared asking AI directly versus adding a layer of cross-validation, and actually ran through it. In an Andrew Ng interview, there's a quote:

You brainstorm a problem with AI, and it usually gives you ten ideas. One or two are good, two or three are mediocre, and the remaining four or five are absurd.

The trouble isn't the absurd ones; it's that the one or two good ones get reused repeatedly, eventually turning into "everyone thinks this way." This is the scariest thing when evaluating valuations. Where is the ceiling for a track? What is the competitive landscape? If everyone uses the same entry point to ask the same set of questions, the answers will look like they were carved from the same mold.

First, pick a question you really need to judge. Don't make it too big, e.g., "Is looking only at monthly sales enough for million-yuan luxury cars?" Prepare three things: Chat AI using GPT-4 via OpenAI, a local small model using an edge-side model (runs offline), and a record table using task management software. Create a new table in the task manager with three columns: AI Original Answer, Counter-Argument, Evidence Gap.

Open a new conversation in OpenAI, ask it for 5 verifiable hypotheses, and instruct it not to draw direct conclusions. Seeing it list hypotheses rather than verdicts is key. Copy the response into the "AI Original Answer" column. Then open the edge-side model, ask it to point out the 3 most likely errors above, and provide a counter-example. Its wording might be clumsy, but the bias has value. Copy the local model's views into the "Counter-Argument" column.

I've been testing batch processing lately—throwing a batch of questions at a script at once—but beginners shouldn't use this yet; get the manual workflow smooth first. Then use tools like Insightify (which I've been using for a few days) to tag viewpoints. Create a new project, paste the viewpoints, and select tags like Sales Volume, Residual Value, Channels, Reputation. My tests show it can organize scattered talk into grids. Finally, do an anti-sycophancy test: ask it to play the role of someone disagreeing with me, using my data to find loopholes. Seeing it attack the original conclusion counts as a successful run.

The easiest mistake is changing the model but not the question. Asking Model A "How is this car?" then Model B "How is this car?" just results in the same sentence with two different accents. The second pitfall is treating AI's new jargon as evidence. If it says "residual value anxiety," it doesn't mean the market is actually pricing in residual value. You need to go back to public information: cumulative sales, user reputation, channel actions. I wrote an article about Maextro (Zunjie); hot search titles took things out of context, and single-month data cannot convict a brand. The third pitfall is discarding everything after running it. Keep the "Evidence Gap" separately, e.g., "Lacking comparable residual value metrics for competitors in the same price range," so next time you won't rely on emotion again.

Where is the ceiling for this track? You can't just ask AI. What is the competitive landscape? You can't let one model set the tone for you. You just need to break the answer into small blocks that are comparable, refutable, and fillable with data.

After learning this, take the three conclusions you were most proud of from the past month and rerun them through the same process. If they don't significantly weaken, it means you probably haven't been homogenized by AI thinking.