Gee, I’m terribly reflective this week. Everything seems to be a "Thought O’ the Day". Oh well. Feel free to heckle me in the comments if this is getting boring!
Earlier in my career, I learned the hard way that what works the best is NOT the same as what’s the smartest.
I have my moments of damning over confidence, but for the most part, I like to think I get technology. In general, it would be fair to say that I see a broad role for technology in our society and I believe in the many benefits it brings.
Given the choice between some kind of Luddite/Amish utopia and the mess we live in today, I’d take the latter. For example, I love the products from these guys, but I’m not so infatuated with technology that I’d ever buy one of these (I’m brave — I’ll touch my trashcan).
Enough staging. On with the story.
I used to work for a company that made tools to integrate paper-based drawings with CAD software. You could scan your blueprints/architectural drawings and load them in to AutoCAD, where through the magic of technology, they would behave like a typical CAD drawing. All very clever.
As with any good software product back in the day, we were trying to placate the gods of feature creep by staying on a regular release cycle.
One of the "big" features that no-one did well at the time was symbol recognition. Just as OCR recognizes scanned text and converts it in to words, symbol recognition recognizes the scanned squiggles that denote doorways or windows, or electrical outlets etc.
Good symbol recognition is absurdly hard. Matching handwriting is bad enough, but symbols come in all shapes and sizes and aren’t usually separate from the rest of the drawing in the same way that text is.
As I recall, on a sunny day with a light wind, the best software at the time hit about a 65% accuracy rate.
We invested years of time and a fair chunk of $$ to "solve" this problem with some very advanced code. Fancy AI with neural this and fuzzy that. So cutting edge, it hadn’t had time to bleed. With much work, we increased recognition rates by more than 30%, hitting as high as 85% in some cases.
Oh how we mocked out competitor with their backwards technology and measly 65% recognition rate. Oh how we praised our Deep-Thought-in-Lines-of-Code recognition engine.
But here’s the thing: If you hit 85%, 15% of the symbols in the drawing are still wrong, and you don’t know which ones are wrong until you check. Our competitor developed a snazzy UI that identified anything that *might* be a symbol in a drawing and gave you tools to very quickly bounce around all of them, manually converting the symbols when you found them.
With a typical drawing, allowing for processing, error checking and cleanup, our incredibly high tech solution was 2-5 times SLOWER than the stone-age-inspired technology from our competitors.
Their manual system optimized the contribution of both the CPU and the human, delivering a better result than either would have alone.
Stupid simple technology. Genius implementation. And they won.
So, like the title says, being the smartest isn’t always the best way to win. Sometimes a little stupid goes a long way…
Categories: Old Posts
No Comments »
