The 1000 Indians story was completely incorrect though. They has 1000 people in India tagging the videos after the fact for improvements to the machine learning algorithms. They weren't watching in real time.

There are reports that 70% of the sales had to be manually reviewed by the team in India (https://www.businessinsider.com/amazons-just-walk-out-actual...), so no, they weren't just training algorithms.

That article in particular was the one that was debunked. It was based on a false rumor.

Debunked where? The only response is from Amazon (in the same article), who are disputing the number but still admitting that the team did intervene when needed.

And they did not give their own number, so even the dispute is very weak.

The 1000 Indians working to track all activities remotely was/is a false story.

See link below:

https://www.linkedin.com/feed/update/urn:li:activity:7186447...

Can you please produce any evidence that it was debunked? I can't find anything.

You don't have to cover for Amazon. It's a megacorp, it's going to be fine without you.

You make all these claims, "false rumor" this and "debunked" that, back them up with nothing, and then when someone posts a link in response you're like "See! Amazon debunked it."

???

No need to make this personal. This is about Amazon, not jedberg.

I'm certainly not defending Amazon, I'm defending AI. What they build was impressive and worked for the most part (I saw the behind the scenes). It was just really expensive because you needed high resolution full coverage of the entire store.

I just find it disingenuous for people to pull out this example as "why AI doesn't actually work" when it's not an example of it.

Why would it be disingenuous? They could be just misinformed or wrong.

Please provide evidence of your claim.

From TFA: "While Amazon insists these reports are “erroneous,” it doesn’t deny that humans aren’t involved with the process at all. Instead, Amazon says its workers are tasked with annotating AI-generated and real shopping data to improve the Just Walk Out system — not run the whole thing. “This is no different than any other AI system that places a high value on accuracy, where human reviewers are common,” Dilip Kumar, the vice president of AWS Applications, writes in the post."

Don't see how this refutes any claim. If anything, Amazon is confirming they have human intelligence involved, they're just leaving the quantity/portion ambiguous, as behooves them. Do you have Amazon equity?

Nevermind, I see from your profile, you do have Amazon equity.

"I used to run reddit.com's servers, then I ran Site Reliability at Netflix, then I worked on making Alexa faster and smarter."

Real astroturf hours here. Please keep your Amazon apologia to yourself.

I have a question: was the customer billed based on whatever the machine learning algorithm determined, or based on what humans decided after the fact? In other words, was the humans' job to just provide training data for future versions of the machine learning algorithm, or to manually fix the algorithm's errors and prevent over-billing?

IIRC, they could correct your bill after the fact.

I used to joke with others that shoplifted items should be retroactively free - as long as you self report after the fact. (Extra useful training data...)

From my experience you just weren’t actually billed until they were done, it could be many hours after you left the store.

Having not read the original story, this makes more sense.

"1000 Indians that were watching people shop at the Amazon Go stores" ...

Conjured up picture in my head of an Indian man inside an Amazon Go store sitting behind a registerless counter. What his purpose would be, I don't know.

Funny story. The amazon go stores do have a register at the counter now.

But as far as I can tell, they rotate positions. There's never really anyone standing at that counter. If you walk up to the counter someone will come from the sandwich counter or restocking the shelves or whatever.

For what it’s worth, I purchased from one of these Amazon stores every day for over a year. After leaving the store, it would typically take 2+ hours before my purchases were registered in the Amazon app. I agree that there probably was a large degree of automation but based on how the store worked, there didn’t need to be an army of “real time” reviewers for humans to be involved in each individual purchase… it’s totally plausible based on what I experienced that Amazon did have an army of people involved in the standard purchase path.

I think it was from a video I watched a day or two ago, describing how with those uber and lyft and other order delivery services that consumers apparently are placing those orders wtih the services, picking up the items themselves and walking out, and then when the actual delivery person arrives, order is gone, reported as missing, the consumer also reports similarly, even though they picked up the orders themselves, that this is apparently happening more widespread commonly that still is vulnerable with business infrastructures to not prevent the exploit from taking place, I am curious how this may affect industry standards with regards to this concept because I never witnessed it but I wouldn't be surprised if it is not even that widely realized about the issue yet thereby enabling more theft and loss and damages and whatnot.

Wonder if you could some how arbitrage your own food delivery...

How many times can one's order be 'stolen' like this before a pattern emerges and the account is banned? This is also fraud and because it requires using the internet, it's going to lead to felony charges.

If memory serves me well, I think some guys in Florida found out how many times they could do it and got busted for fraud against the delivery companies.

System caught the anomaly, humans investigated, law enforcement made the arrests.

Same here.

My theory is that purchase tracking is done automatically via AI, while humans are on the loop to check the feed of people taking each item and verifying that the amounts are correct before charging.

It makes sense. Even if your entire AI system is 99.9% accurate, you want a human to detect and remove that one-in-a-thousand mistake.

> remember the 1000 Indians that were watching people shop at the Amazon Go stores?

> Amazon told us it was all "sensors" because it fit their company narrative to do so.

I think you, like many others, fell into the trap of just looking at the headline. The 1k people were labeling training data, not watching you shop like a puppet master..

Per the articles I read, they did more than label training data. They also did other things to facilitate the checkout process. They didn't get into what those things were however.

> "Associates may also validate a small minority of shopping visits where our computer vision technology cannot determine with complete confidence an individual's purchases," the spokesperson said.

The whole discussion around that headline has always been so confusing. It's honestly unclear. What I'd like to know is, what % of actual costumer checkouts required a human being vs what % was 100% automated.

Some of the headlines and negative discussions make it sound like 100% of the checkouts required human approval. If that is true, then that's clearly terrible. If on the other hand it was anywhere below 5-10%, then I think that's a fair target for training the model in more difficult edge cases.

Something like Waymo also has humans constantly monitoring and helping, but I don't see much discussion about that being "controlled by humans".

> Something like Waymo also has humans constantly monitoring and helping, but I don't see much discussion about that being "controlled by humans".

Constantly doesn't really make sense, riders can call for support and get a near instant response, but it's pretty obvious from some of the mistakes that humans aren't watching every ride in real time. There's video from inside a Waymo which is stuck in a lot repeatedly looping, discovering the exit it wants to use is blocked and then just driving to another exit before deciding hey, perhaps the exit I want to use is open (nope, it isn't) and returning again, and again, and again for example.

All the humans involved can see what's wrong here, but the Waymo driver software doesn't get it and so I think eventually a human (Waymo employee) has to drive the car out.

And yeah, there's definitely in San Francisco for example some people who believe it's all a Mechanical Turk again, even though that doesn't make any sense. In their heads, driving is another of those "uniquely human" abilities and so the only way to have AI cars would be fully general AI, and the real explanation has to be off-shore remote driving.

Would it even be feasible to provide a taxi service with remote human drivers? I get the remote driver occasionally dealing with slow movement, but I think full time driving at speed would be impossible due to latency issues. It’s not a remote piloted drone that is in the air and so has much less of a chance of hitting something by accident.

Otherwise you think someone would have tried offering taxis driven by people in India or the Philippines already.

I believe there's a startup saying they'll deliver cars to you, so, cars with no people inside them driven remotely, once there's a human inside the human is driving. I don't recall the name.

Waymo itself is clear that it has no remote driving capability whatsoever. The Waymo driver is in your vehicle, so any technical issue (e.g. phone network drops out, satellite fails, whatever) won't affect the vehicle's self-driving. It may give up and pull over and let you out, but since it's local any situation where the driver is disabled would be similar impact to a human driver - if a block of concrete drops off an overbridge and smashes the Waymo driver, that could easily happen to your Uber driver the same.

Hence their choice to employ humans to attend in person and drive a Waymo which gets into trouble. In fact the humans sometimes have had to "chase" a Waymo car that, unless it has been specifically told to stop, assumes it should try to complete the journey once it can figure out how. Waymo's remote support can reach into the driver's model and tweak it e.g. to label a stray traffic cone as just trash, not actually off limits - but they aren't driving the car.

Claim was around 70%. Amazon disputed but didn't provide own numbers. https://arstechnica.com/gadgets/2024/04/amazon-ends-ai-power...

Maybe it was 69% XD

"may also" still doesn't suggest that the remote observers were doing all the work and the computer vision systems didn't work and it was all a face though. spot checking that the system works isn't the same thing as doing the job of the whole system

No one is suggesting that. The question is how automated was the system? According to the article 70% of all transactions were manually checked out because the algorithm didn't have enough confidence in its prediction. Amazon has disputed the number, but not the fact that the manual intervention was happening.

People keep saying this despite it being not true.

If they were labeling data the stores would still be open

All the Amazon Go stores that have been open near me for the past few years are still open. Went to one last Friday, in fact.

Agreed. These numbers are being cherry-picked. From Amazon quarterly filings[0][1], here's their employees count:

  - Q3 2019: 750,000 employees
  - Q4 2019: 798,000 employees
  - Q1 2020: 840,400 employees
  - Q2 2020: 876,800 employees
  - Q3 2020: 1,125,300 employees
  - Q4 2020: 1,298,000 employees
  - Q1 2021: 1,271,000 employees
  - Q2 2021: 1,335,000 employees
  - Q3 2021: 1,468,000 employees
  - Q4 2021: 1,608,000 employees
  - Q1 2022: 1,622,000 employees
  - Q2 2022: 1,523,000 employees
  - Q3 2022: 1,544,000 employees
  - Q4 2022: 1,541,000 employees
  - Q1 2023: 1,465,000 employees
  - Q2 2023: 1,461,000 employees
  - Q3 2023: 1,500,000 employees
  - Q4 2023: 1,525,000 employees

The 1.6M count was a clear seasonal/stay-at-home blip.

[0] https://ir.aboutamazon.com/quarterly-results/default.aspx [1] https://s2.q4cdn.com/299287126/files/doc_financials/2023/q4/...

Thank you for taking the time to break this out.

Wow, that shows a surge of about 800k employees from Q4 2019 to Q4 2021, just two years! Wild.

Hard to say how much is seasonality vs laying people off due to increases in robots based on historical numbers. Thanks for breaking this out.

> the height of at-home shopping

I wondered that, too, but I did see that Amazon revenue has continued to grow since 2021. Of course, it’s possible AWS and other divisions are compensating for a decline in physical goods sold. Unfortunately, I don’t have time to find revenue numbers for specific divisions.

That's from inflation. Look at CVS revenue growth and that's with a declining customer volume. Fewer customers buying fewer items, but fairly massive revenue growth.

Because of massive price increases due to a monopoly.

Can you share more on this? What monopoly?

CVS/Walgreens

They've also laid off a lot of office workers, who were not replaced with robots... so we can't correlate the robot number and all of the changes in head count.

Very true. Could be counting 50k IR sensors that count how many items pass on a belt. (Plus maybe a PIR motion sensor in the bathroom )