patrick.net

I guess that stuff works until Toto comes along:

Isn’t Amazon the one you should be denigrating?

GNL says

Isn’t Amazon the one you should be denigrating?

Amazon may have been using this for AI training purposes.

The way AI works is that you have this mass of neurons, and then you have inputs and outputs. You have to identify both the input and output CORRECTLY, and the neurons setup their weights so that they can also match inputs to outputs correctly.

The Indians may have just been, without their knowledge, being used to train the AI. I guarantee what they did was all recorded.

If you show an AI system a cat, a dog, a horse, a butterfly, an eagle, and show 100 of them in various positions and in different scenes, you can train it so that when that AI sees a cat it's never seen before, it's VERY likely to be able to identify it correctly as a cat. If you show it a lizard though, it will categorize it as either a cat, a dog, a horse..

Lizard image will not have features needed to identify as cat dog or horse. If feature match is below threshold probability it will mark is as unidentified. If algorithm is forced to pick only between three then yes it will pick. But most of these program will have 90% or more match to identify.

Who amazon hires to do work is is irrelevant. But you need people to verify ai specially during initial usage, n testing.

Image comparing is done by two means, edge detection, and pixel matching. If you have a classic bottle of Heinz ketchup, AI can detect it by edge detection with 90% certainty, and use pixel matching to identify the label. But if the bottle is placed on a weird angle where the bottle contours are not pronounced and the label is not in view, then it will fail. Scanning labels and using OCR to extract the text and query the product would be most efficient. But then that isn't AI, that's just using plain old Enterprise technology that was around before the AI buzzword. AI is matching the product based on what it thinks the shape and label of the product looks like.
But at the end of the day, it's probably just easier to hire an army of third world humans to work for $20 or less a day, to help the bugs along. They will always be needed, this is beyond testing phases. Every technology has its limitations, and the customer then has to decide if augmenting the software's failures with humans is feasible. The biggest failures of enterprise software is the executives don't want to concede decisions back to the employees that they are trying to take decision making from.

It seems to me this can be done a lot simpler.

Place OCR codes on the bottom of all products, and require the item to be placed upright to checkout. Verification can be done by weight and shape as well, to be certain that labels aren't swapped.

I know that with producers they are extremely particular about what goes on the face of the box, and sometimes even the top, but I don't think they care about the bottom. The real problem is how do you stop shoplifting? There's got to be somebody there, it's not reasonable to hold the customer to ensure they've paid for everything before they leave if, say for example, somebody swapped the bottom labels are there's just a software fault, or if indeed they attempted to steal.