I'm not sure what you're looking at here. There's no "tech demo" going on here. It's a new beta of technology replacing an old and quite well functioning technology that is a SAE level 2 system. This is not frozen in stone. In fact Tesla is updating it every ~5 days right now. There's many people that post daily videos of how Tesla handles on a specific road near them and you can see it improve update to update.

Ironically, deep learning might finally be the solution to this, in a convoluted way.

Previous AI systems generally fell over when they hit the final 2% of real world scenarios, unable to cope with the explosion of complexity.

In contrast, modern AI finally provide a toolkit for "last mile AI" (proverbially speaking), by allowing incorporation of those scenarios without breaking the entire previous model.

All deep learning implementations to date fail on the last 2% of real world cases. Why would deep learning be the solution for its own shortcomings?

Failing on the last 2% (or whatever small number we want to set) isn't a shortcoming of deep learning: it's a shortcoming of all prior knowledge systems. Humans included.

And the barrier to surpassing that problem has been hardware (now approaching solving) and calendar- & person-time.

The problem with 1970s AI (ingnoring hardware) isn't that we couldn't build useful AI, but that we couldn't afford to build useful AI in most domains.

Calendar time costs money. People time costs money.

Deep learning and derivates finally provide a solution where we can substitute machine time and data volume for calendar and people time. As a result, we can afford to do things in domains we couldn't before.

> Failing on the last 2% (or whatever small number we want to set) isn't a shortcoming of deep learning

Citation needed. Deep learning is used to power some of the biggest money making machines in the world (like google or facebook ads) and they're still bad. It's also used to power products like voice recognition and it'll still add dog to my shopping list instead of soap.

I would counter that those systems are as good as they need to be. Google and Facebook have realized advertisers are dumb. So why waste a ton of people time optimizing past what moves the revenue needle?

Where I'd expect to see interesting developments (hopefully published openly) is in sentiment analysis / disinformation identification, as companies begin to see it as an existential threat to the unregulated nature of their businesses and fund accordingly.

Forget this 2% business. How does it do in snowy conditions at night? Because in some places, that's around 12.5%.