Mike Taylor pairs Claude with Jev for napkin math

Mike Taylor ran back-of-the-napkin interview questions through a two-model setup, with Claude decomposing the problem and Jev putting a probability on each piece. These are Fermi problems, named for Enrico Fermi, where you estimate a quantity nobody handed you by splitting it into parts you can guess at. Taylor says Claude and Jev are "fantastic" at them.

Mike Taylor

@hammer_mt

X

claude breaks down the questions and jev estimates the probability

Mike Taylor

@hammer_mt

X

claude still does well because it silently works it out in its thinking traces, but jev can match its estimations 1:1

The honest part of his write-up is the baseline. Claude on its own gets these "near 100% right", which Taylor attributes to the answers being memorized from training data. So the pairing is not buying accuracy on the standard questions. It is buying a mechanical chain of estimates you can read, rather than a number that arrives already known.

The harder case

Taylor's more interesting test is a question with no published answer, his example being how many round trips to the Moon all US vehicles drive in a day. Claude still does well there, but Taylor's read is that it silently works it out inside its thinking traces. Jev, he says, can match those estimations 1 to 1. That is the claim worth paying attention to, because it is the one where the answer is not sitting in a dataset somewhere.

He also reports the setup is trainable. Teach Claude to write better questions and break the problem down better, pair that with Jev, and Taylor says it almost always lands within 2x of the answer. Within 2x is the usual bar for this kind of estimation, and he does not report a count of problems, a failure rate or which questions it missed.

For people building on it

This is one person's set of runs posted on X, not an eval, and there are no numbers behind "near 100%" or "almost always". Treat it as a prompting pattern rather than a measured result. The pattern itself is the takeaway and it generalizes past trivia questions. If you already use a frontier model for structured estimation and want the intermediate steps auditable, handing each leaf estimate to a cheaper classifier gives you a chain somebody can check line by line, which is a different product than a confident single number.

Taylor's pitch for why anyone should care is strategy work. He used the technique to estimate growth marketing funnels, sketching a client's business and finding the weak points over a coffee meeting. His closing complaint is that not enough people prompt Claude to do this, or realize how good Jev is at estimation. That is an opinion from someone who ran it, which is the better kind, but it is still an opinion.