Outrageous! AI supermarket has "no one" collecting payment, 1,000 Indians are hiding behind it...

Mingmin House is from Aofei Temple
Qubit | Official account QbitAI

It’s outrageous. It’s already 2024, and the drama of artificial intelligence relying on artificial intelligence is still going on.

And it’s similar to the kind where a real teller sits behind an ATM machine and hands you money!

When you walk into a supermarket and select the products, you don’t need manual or self-service checkout. You can just pick up the items and leave . After a while, AI will be able to identify what you selected based on the video , and then send you the bill. All you need to do is Click to pay.

Doesn’t it sound convenient?

"In exchange for 1,000 Indians . "

Behind the camera, there is an Indian team of thousands of people who need to remotely check with the naked eye what products the customer has taken.

Oh boy, a new equation appears:

AI = Anonymous Indians ? ? ?

This is the “black technology” of Amazon’s unmanned supermarket that was recently exposed.

It originally claimed to use AI technology to take it away, no need to scan the QR code, and deduct money afterwards, and it is completely automated. It is suitable for Amazon Fresh chain stores and will be promoted in Los Angeles, Chicago, the United Kingdom and other regions starting from 2020.

As a result, after 4 years of use, the real face is that humans are working for AI.

△ Intercepted from the comment area of ​​the article "Those Things About Britain" on the public account

"Very dependent on human CV"

Amazon’s service that caused a lot of complaints is called “Just walk out”.

Its main purpose is to save customers the time of waiting for checkout when visiting the supermarket. Launched in 2016, it claims to be completely driven by computer vision technology, using surveillance cameras and sensors in supermarkets to track the products people take when leaving the store.

Over the past few years, 27 of the 44 Amazon Fresh stores have supported the service, as have some organic food chains.

However, there were some minor problems with this model at the time. Customers often received their bills a few hours after leaving the store because humans had to re-review the video to ensure that the bill was correct.

According to The Information, until 2022, for every 1,000 "Just walk out" services, 700 will still require manual review .

This is far from reaching Amazon’s original internal goal, with only 50 cases out of 1,000 requiring review.

Amazon certainly does not admit that the business is actually run manually. The explanation they gave is:

In the process of machine learning, manual annotation is necessary. Of course, if AI is really unable to judge, humans will also participate in a "small part" .

As for the specific ratio, it is worth pondering.

However, why was this pseudo-artificial intelligence discovered?

Let’s start with Amazon’s recent major layoffs.

The wave of major layoffs just announced by Amazon includes offline store technical teams, with hundreds of positions being eliminated. Sources said that part of the business of the identification and checkout team was affected.

This "Just walk out" that operates in a strange way will no longer be able to walk. After that, only some stores in the UK and Amazon convenience stores will be retained.

In the future they plan to bet on the “Dash Carts” service. Embed scanners in shopping carts in offline stores to identify the items purchased by customers and then check out by themselves.

Relatively speaking, this model does not require the installation of a large number of cameras inside the supermarket, the cost and technical difficulty are reduced, and the accuracy can also be improved.

Not just Amazon

In fact, Amazon is far from alone in engaging in such saucy operations.

In 2017, research consulting firm Gartner released a report on the hype cycle of new technologies.

Reports show that many AI technologies, such as DL and ML, were at the peak of hype at the time.

It is highly sought after by people and the market, and many software vendors have begun to chase the craze and incorporate AI technology into their product strategies.

However, the report clearly shows that most of these companies are actually trying to sell their products, and often exaggerate the application of AI in their products and the capabilities of the AI ​​they use.

This phenomenon was called "AI whitewashing" by Gartne back then.

After the incident was exposed, this phenomenon has not been curbed:

In 2019, due diligence firm West Monroe Partners examined the marketing materials of 40 U.S. companies.

The inspection revealed that these 40 companies exaggerated the capabilities of AI and ML in their marketing claims, with the average level of exaggeration exceeding 30%.

That same year, London-based venture capital firm MMC studied 2,830 European startups classified as AI companies.

MMC found that among these companies, only 1,580 truly fit the description of an AI company.

What specific companies have done these things?

For example, Facebook (now Meta) .

In August 2015, Facebook launched a text-based virtual AI assistant named M; in early 2018, the project was shut down.

At that time, Assistant M claimed to kick Siri and punch the search engine.

Looking back at user feedback at that time, we found that M Assistant was very similar to today's popular ChatBot, but its response speed was slower than the sloth lightning in "Zootopia".

According to an article in Wired magazine, a user asked M Assistant to recommend famous attractions in California. It took M 15 minutes to recommend 22 results. In this time, he could have read ten pages of Google search results. Eight pages.

But this is not a fatal bug of M Assistant.

Sources later revealed that Facebook had been using humans to operate its so-called AI system behind the scenes for more than two years during the project's operation.

This ridiculous operation is exactly the same as Amazon’s unmanned supermarket, no wonder the response is so slow...

There is also a start-up company called x.ai (not the big model company that Ma Yilong is now) , which once promoted that it uses an "AI personal assistant" to arrange meetings.

But it’s actually human employees who do the work of scheduling meetings.

In 2021, x.ai finally closed down because it was difficult to realize the independent operation of the algorithm.

Why can't you go on?

I found no. Although the "fraud scenarios" of the above companies are different, the reasons are basically the same: excessive exaggeration of AI capabilities, doubts about implementation, and doubts about making money.

Taking Amazon’s “Just walk out” as an example, someone analyzed that this model is destined to be unfeasible and “just a marketing gimmick.”

Even with a large number of sensors, items picked up by customers can still be missed. And sensors are not the only cost, manual annotation is also expensive. Whenever a new product hits the shelves, a few annotated images are needed to recalibrate the sensor. As long as manufacturers redesign their products, the accuracy of the original data set will decrease.

Data sets, technology costs, research costs, machine operating costs, etc., lead to very high overhead. Even a large model will not help much . It may slightly reduce the annotation cost, but the computational cost will increase again.

According to public information, the scanners and camera systems equipped in each Amazon unmanned store are very expensive.

According to Amazon insiders, each store may cost tens of millions of dollars.

Secondly, Just walk out also has high technical requirements .

In September last year, Amazon’s official introduction mentioned that in order to figure out who got the tuna sandwich and who got the chicken salad in the store, and to calculate the correct cost, the technology behind it has a lot of background:

Just Walk Out is a combination of computer vision, object recognition, advanced sensors, deep machine learning models and generative AI.

Each one is the latest in computer technology.

Furthermore, the supermarket is dimly lit and the shelves are not densely packed, so it is difficult for AI to accurately determine which product a customer has taken.

The recognition effect is unstable and data requirements are always changing, so the problem of "identifying what customers bought" will become very complicated.

Of course, some netizens question the authenticity of the "ATM hiding" strategy adopted by each of Amazon's unmanned supermarkets.

Someone even calculated an account:

Even if 1,000 Indian employees are hired to work three shifts to keep an eye on what customers in unmanned supermarkets are buying, this manpower is still far from enough to support the operations of all supermarkets.

To conclude, Just walk out may not be completely AI, but it is by no means a completely manual remote checkout.

Some netizens mocked Laman:

Good thing! This proves that AI will not really make people unemployed (dog head) .

Even in comparison, it is better to replace it with human employees, the effect is stable and the cost is not so high.

One More Thing

Of course, innovation never stops, and a more abstract checkout model emerged.

A Filipino cashier connects via Zoom and comes to New York halfway across the world to settle the bill for you.

It is said that it was created by a company called "Happy Cashier". It does not have a website and is implemented in five Asian fast food restaurants in New York.

Reference links:
[1] https://gizmodo.com/amazon-reportedly-ditches-just-walk-out-grocery-stores-1851381116
[2] https://twitter.com/themaxburns/status/1775215997898698907
[3] https://www.mturk.com/
[4] https://arstechnica.com/information-technology/2017/11/expensify-acknowledges-potential-privacy-problem-by-calling-it-a-feature/