Setting a bar is the mistake. We need to reframe the entire narrative.
Safety implementation is never objective. You can only implement a system by subjecting it to context. Traffic safety is a world of edge cases, and each driving implementation will engage with those edge cases from a different subjective context.
We are used to framing computation as a system of rules: explicit logic that is predictably followed. Tesla is using the other approach to "AI": statistical models. A statistical model replaces binary logic with a system of bias. A model that is built out of good example data will behave as if it is the thing creating that data. This works well when the context that model is situated in is similar to the example. It works poorly when there is a mismatch of context. The important thing to know here is that in both cases, it "works". A statistical model never fails: that's a feature of binary logic. Instead, it behaves in a way we don't like. The only way to accommodate this is to build a model out of examples that incorporate every edge case. Those examples can't conflict with each other, either. The model must be biased to make the objectively correct decision for every unique context it could possibly encounter in the future; or it will be biased to make the wrong decision.
The only real solution to traffic safety is to replace it with a fail-safe system: a system whose participants can't collide with each other or their surrounding environment. Today, the best implementation of this goal is trains.
Humans have the same problems that statistical models have. There are two key differences, though:
1. Humans are reliably capable of logical deduction.
2. Humans can be held directly accountable for their mistakes.
Tesla would very much like us to be ignorant of #1, and to insulate their platform from #2.
Safety implementation is never objective. You can only implement a system by subjecting it to context. Traffic safety is a world of edge cases, and each driving implementation will engage with those edge cases from a different subjective context.
We are used to framing computation as a system of rules: explicit logic that is predictably followed. Tesla is using the other approach to "AI": statistical models. A statistical model replaces binary logic with a system of bias. A model that is built out of good example data will behave as if it is the thing creating that data. This works well when the context that model is situated in is similar to the example. It works poorly when there is a mismatch of context. The important thing to know here is that in both cases, it "works". A statistical model never fails: that's a feature of binary logic. Instead, it behaves in a way we don't like. The only way to accommodate this is to build a model out of examples that incorporate every edge case. Those examples can't conflict with each other, either. The model must be biased to make the objectively correct decision for every unique context it could possibly encounter in the future; or it will be biased to make the wrong decision.
The only real solution to traffic safety is to replace it with a fail-safe system: a system whose participants can't collide with each other or their surrounding environment. Today, the best implementation of this goal is trains.
Humans have the same problems that statistical models have. There are two key differences, though:
1. Humans are reliably capable of logical deduction.
2. Humans can be held directly accountable for their mistakes.
Tesla would very much like us to be ignorant of #1, and to insulate their platform from #2.