What’s funny is how this shows how bad poor algorithms can be at scale. Even with human intervention. These cases have lots of exposure and people who notice.

What’s funny is how this shows how bad poor algorithms can be at scale. Even with human intervention. These cases have lots of exposure and people who notice. So the real QA is in these edge cases where customers find bugs. I worry about how not so visible algorithms. I had a company market to me about sentiment analysis. I asked their accuracy and they answered that they can’t actually test but their training model has x%. So to make decisions accordingly. But then I had coworkers doing stuff like what percent had what sentiment vs the other with no assumptions. I think the problem is people are kind of bad with probabilities so something either is or isn’t. And I seem to witness a gulf where there are people who kind of understand and then people who don’t but trust the output.