I know this is ripe to use as a way to dunk on Amazon, but I sincerely doubt that it was as inane as it sounds. They likely had this workforce to train a system that they hoped would eventually operate autonomously. That being said, it sounds like it never panned out.
And this is why I think current “AI” technologies are a dead end. Once the rest of the market realizes that maybe we’ll finally get past this “AI” hype bubble.
What AI applications are in use where the error cost is imposed on the company using them, not the consumer? Reading paper checks and mail envelopes come to mind. Those are automated and work well. Waymo's automatic driving, but not Tesla's. What else?
This is a useful metric for gauging real progress.
If Tesla geofenced their FSD applications as aggressively as Waymo, I suspect the two would be near peers. You can hide an incredible lack of a capability by artificially limiting the scope of your application.
> This is a useful metric for gauging real progress.
Sure, cost / benefit analysis is always useful. I would propose a broader analysis for grading “AI” disruption though.
1. Are humans involved at any step?
2. What is the probability of success?
3. What is the consequence of failure? (Profit loss / property loss / personal injury, etc…)
4. How many variations on a domain (think winter driving, summer driving, driving in a tropical storm, driving dirt roads in a jungle, etc) can the application work in while maintaining consistent performance?
5. Can the model leverage its “learnings” to solve problems / reason about new domains that it hasn’t been trained on?
That’s just a few metrics I can think of off the top of my head that would help to distinguish a true disruptor from the hype.
Air Canada's website AI chatbot for one