Your Best People Can Show You What They Can't Tell You
Polanyi's paradox, four million labeled faces, and where your team's know-how really lives
Your best person is on vacation this week, and the new hire just asked you why an invoice got flagged.
You know the invoice.
You open the shared drive, the process doc and three old email threads, and the reason is nowhere in writing.
The person who flagged that invoice could explain her call in ten seconds, and she is on a beach with her phone off.
So you promise yourself that when she gets back, you'll ask her to write everything down.
We think that request is the wrong first step.
She'll try hard, and she'll hand you a careful document, and the reason she flagged that invoice will still be missing from the page.
A philosopher explained why sixty years ago.
In 1966, Michael Polanyi wrote a short book called The Tacit Dimension, and near the start he put the whole problem in one line: "We can know more than we can tell."
His example was a face.
"We know a person's face, and can recognize it among a thousand, indeed among a million," he wrote. "Yet we usually cannot tell how we recognize a face we know."
Picture someone you love, and you'll find their face in a crowded airport before you've finished looking.
Then try to explain how you did that.
You can list the eyes, the jaw and the hairline all afternoon, and the list still fails to find your sister in the crowd.
In 2014, the economist David Autor gave this gap a name, Polanyi's paradox.
He used the paradox to explain why computers took over some jobs early and left others alone for decades.
A computer needed someone to spell out the steps, so the work people could explain got automated first.
The work people knew without being able to explain stayed with people.
Recognizing a face stayed on the people side of that line for a long time.
Engineers kept trying to write the rules for a face by hand, and their systems kept falling short of a person glancing at two photos.
That same year, four researchers at Facebook took a different road with a project they called DeepFace.
First, people labeled more than four million photos of faces with the names of the people in them.
Then the AI learned from those labeled examples until the AI could tell one person's face from another's.
Finally, the team showed the AI pairs of photos the AI had never seen and asked whether both showed the same person, and the AI matched faces almost as well as people do.
On the standard test, the AI got 97.35 percent right, and the researchers put human performance at just over 97.5 percent.
Nobody in that project wrote down how to recognize a face.
The people who labeled the photos showed the answer, one face at a time.
Autor saw the same move from the economics side, and he wrote that computer science was trying to get past Polanyi's paradox "by building machines that learn from human examples."
Your best person works the way those labelers did.
She knows which customer needs a phone call and which one would rather get a short email.
She knows the invoice that looks fine and still needs a second look.
She knows the quote that closes if you send the quote before Friday.
Ask her for the rules, and you get the polite version.
Watch her work, and you see the real version in every email she rewrites and every exception she catches.
When we set up AI inside a business, we start with those examples.
We collect the work your team already does, like the emails they rewrote, the invoices they flagged and the reasons they gave when someone asked.
We keep those examples in a plain folder of files and instructions, and you own the folder and everything else we build.
The AI works from what your people have shown, and your people stay the judges of every call the AI makes.
Each correction they make becomes one more example in the folder.
When the models get better, your workflows get better with them, because your examples travel with you to whichever model comes next.
The next time your best person takes a week off, the new hire can look up a flagged invoice and read her reasons in her own words.
She gets to keep her phone off for the whole week.
Polanyi was right that we know more than we can tell, and DeepFace showed that what we know can still pass from one person to the next, one example at a time.
We give businesses back to the people who built them.
What does your best person know that nobody ever wrote down?
Sources
- Michael Polanyi, The Tacit Dimension, 1966, page 4 (University of Chicago Press edition, 2009). Both quotes.
- David Autor, "Polanyi's Paradox and the Shape of Employment Growth," NBER Working Paper 20485, September 2014. Names the paradox; full sentence: "Contemporary computer science seeks to overcome Polanyi's paradox by building machines that learn from human examples, thus inferring the rules that we tacitly apply but do not explicitly understand." https://www.nber.org/papers/w20485
- Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato and Lior Wolf, "DeepFace: Closing the Gap to Human-Level Performance in Face Verification," CVPR 2014. Training set of 4.4 million faces of 4,030 people, labeled by humans; a nine layer neural network with more than 120 million parameters, after a 3D alignment step; 97.35 percent on Labeled Faces in the Wild (5,749 people, 6,000 test pairs) against human performance of over 97.5 percent; most earlier systems used hand-crafted features. https://openaccess.thecvf.com/content_cvpr_2014/papers/Taigman_DeepFace_Closing_the_2014_CVPR_paper.pdf
- Meta research publication page for DeepFace, published June 24, 2014: https://research.facebook.com/publications/deepface-closing-the-gap-to-human-level-performance-in-face-verification/
Original: https://elutionlabs.substack.com/p/your-best-people-can-show-you-what