I live in West London, not San Francisco, so I never expected to enjoy any Artificial Intelligence firsts. But the first Amazon Fresh opened in Ealing in 2021. When I visited late one night, it was largely devoid of customers. There were quite a few staff stacking shelves and milling around, but none at the checkout; because of course, there were no checkouts. You had to confirm you were an Amazon customer via an app when you came in, but after that, you were tracked and your shopping was automatically registered to your Amazon account, and you could “just walk out”. The story at the time was the large number of hidden cameras spying on customers — about a thousand — and we knew that AI was behind the process of logging our shopping habits. The store, which was hemmed in by bigger and more popular supermarkets, shut last summer.
There was some disappointment in the tech community on hearing about Amazon ditching the technology altogether at the start of this month. Some colleagues working with sensors have been somewhat surprised at such a large company giving up on its own technology goals, and there has been the continued whisper of AI still not finding ways to contribute substantially to company profitability.
In his Guardian article last week, the highly respected tech critic James Bridle covered the social dimension — that is, using AI to hide low-paid employment. He reported that about 70% of sales were “reviewed” by a thousand-strong remote team based in India. Apparently, Amazon will be pivoting to smart shopping carts, but this loses the “just walk out” proposition completely. It also proves that there has been no decision to leave the bricks and mortar retail sector; so there has been no radical direction change. It is likely a specific problem with the project.
Fail Faster and Other Software Lessons
From a software developer viewpoint, how did this project from one of the richest businesses on the planet fail? It didn’t fail fast either, clearly. Deciding to discontinue the idea before deployment would have just been chalked down to experience. But this type of public failure is share affecting.
It could be that there were too few project milestones, or some were cut. This could have led to some of the smaller staged rollouts being omitted, leaving insufficient data available to spot upcoming hurdles. When testing, changes to the environment need to be carefully managed, otherwise results can be much harder to interpret. For example, if there were surges of customers at the same time as more theft attempts, it would be much harder to adapt to one of these issues independently.
It could be the team was hoping for generative AI to produce a magic bullet. While research and testing should always improve efficiency for a specific task, maybe an exponential breakthrough was needed in one area. Even with AI improving in leaps and bounds, there still needs to be a stable working position to jump up from. Amazon is big enough to wait and then quickly exploit a change, but that needs careful explanation to stakeholders. This can be trivially disrupted by outside factors.
Scaling back the technology by using a backend human team to override decisions was technically a good idea, but if it was not announced then there were obvious political ramifications. This is an important point for both seniors and product managers to take into consideration — the risks to the company of negative impact from hiding information. Using “mechanical turks” for a short period, and reducing the need for them over time against agreed thresholds, feels like a sane tactic — but it needs transparency and humility. Stating upfront that Amazon Fresh had a human backstop would surely have been sensible. We know that pilots of autonomous car schemes have to have a human ready to take over. However good the AI driving was, this will remain.
There are other ways to lighten the cognitive load for the AI, for example by changing the store layout to make it easier for the sensors. The shelf design in stores is clearly based on known optimal stacking standards, but this makes shoplifting fairly straightforward. While developing the technology, a less combative setup might have helped.
It could be that the feedback elements that agile needs were simply not available by the time the project was installed. Working with staff used to a conventional physical store may have vastly reduced the feedback loop or altered expectations. Projects that cannot improve with feedback are obviously problematic.
When to release and deploy a product has to be a business decision, but it should not be a “fire and forget” affair — especially with novel products. The development team (not just the “operations” part) must be involved because they can spot where assumptions at development can no longer be relied upon in a live environment. They can also bring up forgotten ideas given new circumstances.
AI and Agile
We know that AI still cannot fully appreciate context, nuances, and subtle cues in communication. If one looks at Yann LeCun’s definition of Objective-driven AI systems, what Amazon Fresh was doing seemed to fit in — working towards limited goals and objectives, learning through sensors, and training on video data. But it is possible that doing everything at once was too ambitious. Tracking both an individual’s entrance and exit and tracking their shopping may have been better treated as separate problems for a longer period of time.
Does any of this mean that there won’t be any use of AI in the retail sector? Quite the reverse. There are known projects in retail concerning shelf filling (Morrisons in the UK) and checkout theft (Target, in the US). These projects prove that successful AI-based projects can be designed, developed and rolled out on the shop floor. Because the public has been exposed to the wonders of ChatGPT, there probably still exists a small reservoir of trust in its abilities.
In summary, there is no particular issue with machine learning (other than it might have been easy to fool the leadership team initially regarding any project with the term “AI” in it). Rather, there’s clearly a need to treat AI as a tool like any other and run projects in an agile manner — like any other. Use increasingly larger-scale trials and observe new issues as they arise. Fail fast if some basic assumptions are wrong. Scale back and fix showstopper issues before moving up.
But I’m sure another company is busy continuing where Amazon left off.
The post Applying Agile Techniques to AI: Lessons from Amazon Fresh appeared first on The New Stack.
We look at Amazon's controversial AI-powered grocery store for lessons in agile software development — particularly as it relates to AI dev.