Hacker News new | past | comments | ask | show | jobs | submit login

> Though it seemed completely automated, Just Walk Out relied on more than 1,000 people in India watching and labeling videos to ensure accurate checkouts.

That line was unintentionally hilarious to me. Like looking inside a TV and finding a bunch of fast-working little people drawing images.




> Though it seemed completely automated, Just Walk Out relied on more than 1,000 people in India watching and labeling videos to ensure accurate checkouts.

Yup.

I got the chance to go shopping at one in 2018 (so it's been a while). You could tell there was some "reconciliation" happening, because your receipt didn't show up until 20-30 minutes after you left. My guess is that this was some person via mechanical turk crawling thru the data and indexing what you -really- bought.

Of course, my colleague and I (who were working on computer vision at the time) did stuff like take our bags off, put them in the middle of the floor, and roll a can of soda into the bag, just to see what would happen.


"Of course, my colleague and I (who were working on computer vision at the time) did stuff like take our bags off, put them in the middle of the floor, and roll a can of soda into the bag, just to see what would happen."

I'm many jurisdictions this could be enough to get you prosecuted for shoplifting if it evaded payment.


After advertising "Just Walk Out"?


Yeah, the author is trying to see what would happen if they roll a can into the bag. This is meant to conceal the act from the computer. The act of concealing merchandise is all that is required in many jurisdictions. If they walked out without the concealing act, then I would think the error should be on the store.

There have been some cases of self checkout prosecution too. They don't need any proof of intent. Even regular errors can be prosecuted.


> They don't need any proof of intent. Even regular errors can be prosecuted.

In Common Law jurisdictions, there's no intent required for prosecution of alleged theft, but there absolutely is intent required for conviction! Anything else is a miscarriage of justice.


Except statute and common law can define intent using prima facia definitions. In PA it's considered prima facia evidence as defined in statute that you're hunting if you're walking in the woods with a gun. Doesn't matter your true intent. If you removed an item from a store without paying for it, that's prima facia evidence of shoplifting defined in common law. Again, true intent doesn't matter. The courts will even create president that goes against statute just to make conviction easier - the statute for dog control says it must be "under reasonable control". Instead the courts have ruled that it is a strict liability offense and not subject to the reasonable standard.

So yeah, there's a lot of miscarriage of justice, or at least potential for it. Nobody really cares until it happens to them.


> Even regular errors can be prosecuted.

Volunteer to work for free, with bonus legal liability, as a self-service cashier.

So convenient!


https://en.wikipedia.org/wiki/Mechanical_Turk, on the off chance you've never heard of it.

(This article is not the AWS service by the same name, though by the time you conclude reading the Wikipedia page, it should be obvious where the AWS service got its name, and what it does.)


Thanks for the link. I knew of the concept and the Amazon service, but wasn't aware of the connection between the two. Interesting stuff.


It isn't that bad an idea.

Imagine that next time you are drunk, you can hire a driver who will drive you back home remotely (along with some AI to stop the car in case connection goes away).


I disagree - at a basic level if people are able to tell which products you're picking up the resolution is necessarily high enough that there's a huge creep factor. I am sometimes amazed at what companies will pursue without a glance towards common sense.


I'm a huge privacy advocate and don't think stores should track what you buy at all, so don't confuse what I say next with the is/ought fallacy.

They already have perfect resolution and data retention of everything you buy at checkout time when it's scanned, plus they can verify your identity rather than have to rely on facial recognition or other things. I don't think this is any creepier than what they already do so from their perspective it is "common sense."


The scandal where Target data scientists bragged to reporters about knowing when teenage girls are pregnant before their fathers broke in 2012, and they said they were doing it since 2002. It was based entirely on data mining purchase histories with rewards cards at the register. No fancy AI or facial recognition needed.


https://www.nytimes.com/2012/02/19/magazine/shopping-habits....

Imho more like an anecdote than a sourced story, but a good one nonetheless.


That's fair as far as tracking purchases go, but high resolution cameras while You're in the store could also read your phone screen or other things on your person right? That seems more concerning than normal cashiers.


Yes very fair point. I just had a dystopic thought of dynamic prices (electronic price tags?) that update in real time based on information. Is the person undecided? Brand A bids to lower their price slightly. Is the person looking at the same item on Amazon? Read the price and beat it by 10 cents or whatever. Those are situations where it might benefit the shopper, but I could easily see it going the other direction also. not looking at the item on Amazon? Now you're gonna pay too much. And of course, store the complete history of what strategy works with which people so "the algorithm" can tune for your individual weaknesses.


Someone will reply to you saying "I have nothing to hide"


Is high enough resolution to identify products more creepy than simple ubiquitous surveillance?


One creepy thing existing doesn't mean that another thing can't also be creepy. Getting rid of one of many creepy things is still nice even if it's not the only one.


Except all these trillion dollar valuations in the AI bubble are based on the belief that AI is replacing humans, not just outsourcing to cheaper humans.

Of course outsourcing to cheaper humans can be great. But that's not what tech is shilling to the world as "AI" right now.


The grocery industry generally has pretty low margins. Wal-Mart's profit margin is 2.39%.

If you go into a grocery shop to grab lunch, spending $8 on 4 items, and they make 5 cents of profit per item? They need to run an extremely lean operation.

Or target price-insensitive customers, I suppose.


>They need to run an extremely lean operation.

they simply need to run a good ordering/inventory system. If they sell every item (on average) in each store every week, that's 2.39% return on the value of the inventory investment each week, or 52*.0239 or over 100% annual return on money they borrow for free because the store pays its grocery bills to suppliers in arrears, net 10 days, etc


What I mean is: If you're making 5 cents of profit per item, and your workers cost $10/hour then the difference between a profit and a loss is 18 employee seconds per item.

Which is not much.


This actually exists - testing in Berlin, running in Las Vegas: https://vay.io

Maybe not a bad idea - it is a step towards autonomous driving, and will probably ensure drivers have less unintentional down time.


So instead of hiring an uber you're purchasing a car specialicially designed to be controlled remotely by a third party..


Like a taxi or bus?


> Imagine that next time you are drunk, you can hire a driver who will drive you back home remotely (along with some AI to stop the car in case connection goes away).

Is this satire? It doesn't seem like a fantastic idea to allow someone to remotely pilot a car over a transoceanic Internet link.


Imagine you hire a taxi, but the taxi driver follows commands of someone in another country who receives video stream from the car.

That’s what it’s actually like. It’s just strictly more work than simply having a cashier.


The article they link for that isn't readable without subscription. Were they actually labeling data live? That sounds like effectively a secret concierge following you around and writing down what you're getting & like it would be pretty difficult to keep up with. Otherwise, labeled data and machine learning kind of go hand and hand. It's not wild that Amazon would dogfood its own product while creating training data to improve it.


It does sound like a person was watching the shopper live:

> Though it seemed completely automated, Just Walk Out relied on more than 1,000 people in India watching and labeling videos to ensure accurate checkouts. The cashiers were simply moved off-site, and they watched you as you shopped.

But who knows, tech articles are so inaccurate that it might as well have been some misunderstanding regarding how the labeling/training was done.


Yeah, that's line was what got me trying to look at the article they cite, which is unfortunately paywalled. It'd be one thing to do that at like... several stores to generate the training data. But scaling it to dozens while it still requires that? Baffling.


It sounds like the receipts weren't necessarily "live", sometimes taking hours to appear, so I'm guessing they did it as live as possible and when they got backed up they would just revisit recorded video. Something tells me the expectations for throughput were also pretty overwhelming to the individuals who worked there.


I often wonder how much human intervention goes into self-driving car fleets.


2-4% of the time according to the former Cruise CEO Kyle Vogt in a Hacker News comment.

https://news.ycombinator.com/item?id=38145997


That reminds me of this, https://xkcd.com/1897/ perhaps it was based on something after all.


those grocery delivery robots you see in LA, turns out are being monitored/operated by people in a central location with all their camera feeds.


Kind of reminds me of how back when speech-to-text started to get good enough to use (in the form of Google Assistant/Siri), my parents wondered if there was someone on the other side doing the transcription.


So what we thought was automation was really just outsourcing.


It's me, aws Mechanical Turk again!


i've written about this multiple times on HN and always got downvoted

I work in the CV industry and this was a very badly kept secret. There is absolutely no financially feasible way to run something like this with current CV


i'm curious to know what the limitations of the technology are. Are the machine learning/CV algorithms not accurate enough to run it at scale?


I work with ML CV industrial systems, and they can certainly do accurate and detailed analysis and identification very quickly. The systems that do this are also necessarily very, very expensive -- much more expensive than any grocery store could possibly justify.

Of course, costs change with time, but right now I don't see how this sort of application could approach being financially feasible.


Mechanical Turk




Consider applying for YC's W25 batch! Applications are open till Nov 12.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: