I wish more stores would just do something like Scan and Go at Sam's Club. By far the smoothest checkout experience I've ever used.

https://tech.walmart.com/content/walmart-global-tech/en_us/b...

I wonder if that only really works becuase Sam's club can just revoke your membership if you steal

Apple stores work the same way, except that it is truly “scan and go”, whereas Walmart makes you show a digital receipt. What happens if you steal? I dunno, without checking I’m pretty sure you need an Apple account to use the app, so maybe that gets revoked. Or maybe Apple’s stuff simply works well enough.

At least at Sam's, cameras glimpse in your cart and seems to apply some kind of trust score. Usually, the person at the exit just waves me by, sometimes if the cart is really loaded with odd items or I've already bagged some things they'll want to take a peek.

My understanding is this turned out not to be true. People were used to label stuff for new stores, but the actual implementation did not depend on some sort of fakery.

I imagine it's possible the truth was somewhere in between. But if it worked, why did they stop using it in their grocery stores after putting so much money into it?

In the UK at least the grocery stores are completely empty. I've barely seen anyone in them. Bizarrely they are shutting loads down in London but opening new ones at the same time. Absolutely no idea what the strategy is, they must be throwing out the majority of the fresh food stock they have.

That was the original idea, and that's what Amazon claimed, but IIRC they never got over 70% automated.

Phrased differently, 30% of all transactions were still entered by a human overseas watching cameras at the time they decided to pull the plug, years after the initial launch.

Interesting that they can automate some of the transactions and not others... Wonder what was special about those other 30%.

There was a 2 year period in which I bought lunch at an Amazon Go daily. I was naive to the magic so I thought it was the greatest innovation ever.

I remember being 13 years old and stepping into an Amazon Go store in Seattle. Little me lost my mind. I think I walked in and out of the store like 5 times just to see the amazon charge. Sucks that half of the magic was a lie.

Shame to see another project fall to the strategy of AI = "actually indians". I wonder how many other companies have engaged in this stuff.

See: https://www.businesswire.com/news/home/20240611122778/en/Bui...

Depends on how you define real. I would argue that GPT-2 was a real LLM and it almost certainly cost a lot less than a billion. I'm sure there are much better examples.

Can you imagine Builder.ai using a model that argues with their clients or discriminates against them? I don't think so. GPT-2 is like bringing a knife to a gunfight in 2025.

If you want to compete with the likes of GPT-4, Claude, or Gemini today, you're looking at billions, just for training, not counting infra, data pipelines, evals, red teaming, and everything else that comes with it.

Builder.ai wasn't able to use GenAI to actually build software. And when the money ran out and no model was ever announced, investors lost trust and clients lost patience.

[Disclaimer: Former Amazon employee and not involved with Go since 2016.]

I worked on the first iteration of Amazon Go in 2015/16 and can provide some context on the human oversight aspects.

The system incorporated human review in two primary capacities:

1. Low-confidence event resolution: A subset of customer interactions resulted in low-confidence classifications that were routed to human reviewers for verification. These events typically involved edge cases that were challenging for the automated systems to resolve definitively. The proportion of these events was expected to decrease over time as the models improved. This was my experience during my time with Go.

2. Training data generation: Human annotators played a significant role in labeling interactions for model training-- particularly when introducing new store fixtures or customer behaviors. For instance, when new equipment like coffee machines were added, the system would initially flag all related interactions for human annotation to build training datasets for those specific use cases. Of course, that results in a surge of humans needed for annotation while the data is collected.

Scaling from smaller grab-and-go formats to larger retail environments (Fresh, Whole Foods) would require expanded annotation efforts due to the increased complexity and variety of customer interactions in those settings.

This approach represents a fairly standard machine learning deployment pattern where human oversight serves both quality assurance and continuous improvement.

The news story is entertaining but it implies there was no working tech behind Amazon Go which just isn't true.

That's some fascinating background, thanks! Probably explains why they keep operating it in stadiums but not grocery stores. Works pretty well with a small handful of items, does not scale up reliably to shopping carts full of stuff.