Hacker News Reader

Comment by devilmoon | Hacker News Reader

It's interesting that they picked Amazon Go as an example of the power of AI (or ML) when in actuality most of the grunt work in those stores is done by "dumb" sensors AFAIK. The store "knows" that an item has been picked up from a certain shelf and thus it assumes that the item is what is usually on that shelf, it has little to do with either CV or AI in general, just careful placement of the items and weight sensing. I think that if you let a few kids roam free in there picking stuff up and setting it down in an unexpected place the magic of AI would dissolve into thin air.

View on HN

Have you actually visited one?

I've been a couple times to one of the SF locations with coworkers. We definitely tried to trick system by passing items between us, and putting items down in random locations and then picking them back up later. When we left the store everyone got charged for the items they had on them.

I think you are too dismissive in saying its "just careful placement of the items and weight sensing".

How do you know there aren't humans on the other side of the cameras doing the item-customer matching?

Go has relatively few stores. Seems like the sort of thing you can still do manually, especially if the model expresses doubt.

I went to one of these stores in Seattle on a weekday afternoon with a couple friends. It was literally packed with people. We tried breaking the system by leaving items in random places and exchanging them between us. It still charged me accurately for a couple bottled waters, a sandwich, and some other snacks.

Again, how do you know there aren't humans on the other side of the cameras/sensors facilitating this?

At the scale they're operating at and the accuracy you point out, to me it's more plausible they're having humans assist.

I think they're saying that humans behind cameras likely wouldn't be able to do this accurately.

Except that cameras in stores have been used for anti-shoplifting for decades. They've been used for more subtle things like monitoring tables at casinos.

I mean, I'm just asking the question how people know for sure Go isn't (at least partly) a Turk. It seems like an obvious thing to ask? But I first came to wonder if there's manual component involved after shopping a few times at Go. When it was busy, it took sometimes well over an hour to get a receipt. When it was empty, it was immediate. How come?

Also, the accuracy is completely anecdotal both ways... at least one person I visited the store with never got billed.

This is the fun kind of Turing test: tasks where only a computer/mechanized system can succeed.

Customer "4201, tracer". Troublemaker. Enable customer-specific pricing increases until resolution.

It's also RFID tags, which some other stores use to track product locations.

I can't help but think -- I'll bet if things get tricky they pass the footage to a human to sort out.

Did you forget about the aspect of the store that knows who picked up what? You focused completely on the irrelevant aspect. The store tracks every customer to those dumb sensors and then keeps a running tally of their cart. The dumb sensors are used as double confirmation to handle occlusion.

Reply on news.ycombinator.com