Last week I had a genuinely hard question. A professor had just walked me through the statistical modeling he runs on MRI scans, and I wanted to know whether the same approach could power a prediction pipeline for things that aren't brain scans at all. Exactly the kind of question where you want the smartest answer you can get.
So I did what most of us do and asked my favorite model (Claude Opus).
The response was not fully satisfying so I asked the question (exact same prompt) to four different AIs — Claude Opus and ChatGPT 5.5 directly, Gemini Pro and Nemo Ultra through Perplexity. All four were good but not quite there - none were clearly "the best." Each one caught something the others missed. The real answer was scattered in pieces across all four.
So I took one more step and handed all four answers back to a single model. I said: take the best from each, throw out the worst from each, tell me what's still missing, then poke holes in the logic that led you to those conclusions. What came back beat anything any single model gave me, and not by a little but by a lot.
That rewired how I use AI, and that is the lesson of the day:
Stop hunting for the one best tool. Use what you have to sharpen each other.
Let the best output win every time, and don’t let preconceived notions tell you otherwise. Three things that drove it home for me this month:
The synthesis beats the single shot. Four decent answers, stitched, beat the best individual one.
Your gut about "the best AI" is wrong more than you think. I needed a logo and was sure Claude or ChatGPT would nail it fastest. They didn't - not even close - and I burned iteration after iteration trying to make them work. The runaway winner with zero edits? Gemini, running inside Perplexity - a model I wasn't even taking seriously for the job.
Use one AI to watch another. I set up an automated reviewer to check the code my other AI was writing. Small wrong turns I'd never have spotted added a few refinement cycles but the output quality improved dramatically. The quiet drift it caught was the gap between "looks done" and actually done.
If you're not a developer (like me) and this all feels like it's moving too fast, here is what I learned and that I hope you take away:
You don't need to find the perfect tool. You're not behind because you haven't picked a winner. The edge isn't in the model — it's in the process that produces what you need.
So, go try it out - run your hard question through three models this week, then make them argue. You'll be surprised what comes out the other side.
— D
Voice and direction are mine. Produced with Claude — which is sort of the point. More on that soon.