Thread by @lizadixon

Self-driving is hard.

We have different ideas about how to “solve” for L5, and various teams are all taking shots at it. In recent years, two schools of thought have emerged about how to approach solving this problem.

🧵👇

For some it is either:

1) a fundamental AI problem which needs a new approach
2) a data problem, which can be solved by more data & more simulation

Some see the greatest challenge as developing the right AI approach.

Others believe that they already have the right approach, and therefore the challenge is acquiring more (and the right) data and doing more training.

Imo, there is some truth in both schools of thought.

All of this is all further complicated by a lack of oversight/ability of 3rd parties to robustly test these systems. In this application, standardized tests are easily gamed. We do not have the deep technical knowledge at the governing/policy level to support this technology 🐔🥚

Self-driving is "the mother of AI problems" bc we're not only trying to solve judgment under uncertainty (~AGI), but we are doing so in a safety-critical environment that requires instantaneous perception, sensing, & acting.

#autonowashing this tech was a major industry blunder.

GPT-3 is a good point of reference. It's impressive, especially @ 1st sight. Still, GPT-3 is nothing without human input––it’s not going to write anything up on its own and it needs clear instruction in its required input syntax. In contrast, its applications are safe & contained

GPT-3 can improve with a larger dataset, dramatically so. But no matter how much data it is fed it, it's still just spitting it back––the current AI approach does not allow it to solve new problems without human input & under uncertainty.

How good is "good enough"?
And who gets to decide?

These are the questions we will have to answer in the future. Our safety & wellbeing is on the table, and we are dependent on the tech literacy of people in power to protect us from those seeking profits before all else.