Weedpacket;11065671 wrote:The survey does suffer from a severe selection bias.
According to Douglas Hofstadter, the principal problems of AI are: What is "A"? And: What is I?
sneakyimp;11065676 wrote:I do believe I can rightly apprehend the confusion of ideas which might lead to such questions.
The first one ('A') might be harder in theory; the second is harder in practice. Our EzLink(tm) system categorizes products something over 90% correctly, the last time we checked it, on copiers, printers, etc. (Hardware).
However, throw parts for those, or toner cartridges or document feed trays into the mix and the score slips considerably. This became a real problem when we went from @10K products to well over 100K ... if we upload 1000 copiers and there are maybe 87 marked incorrectly as printers, that's possible to fix with human intervention and a little time (30 minutes or less? depends on the error, really.)
If we upload 50,000 parts and 5,000 of those are categorized incorrectly, that's perhaps a half day or more work, if anyone doesn't go blind/crazy trying to fix it. And the question for me is, with my limited time available, do I fix the errors that are now showing in public on the WWW site, or do I fix the system so it gets smarter about those errors and risk introducing bugs and reversing the success rate? I love to be able to do both, but most days only one or the other gets attention, and depending on the other projects in the pipe, maybe NEITHER.
Increase the system's failure rate to 20-25% and you've just shot a "man day", which probably isn't as costly in Joplin MO as in NYC, but still is something we can't afford.
And that's just money. What if we're talking about IBM's Watson and cancer diagnostics? I wouldn't want to be the guy where they say ... "well, it's just that the AI has a 4% failure rate, and you're one of those cases."
So the level of intelligence/accuracy is really the difficult thing about AI.
Of course that became deadly obvious just the other day. I wonder if Uber has an accuracy metric? ;-)