[RedDevil]

Do we learn from the past or not?
We'll never have infinite access to information that fits in our pockets.
We'll never fly to the moon.
Complex, noisy and expensive cars will never replace horses.
Flimsy houses are no comparison to a solid cave.

The learning behind FSD is progressing to another level. While the data set of traffic situations is growing rapidly, so is the success rate of the decisions the neural network makes. Once the latter increases faster than the former the quality of the decisions will accelerate and surpass human abilities.

For comparison, as a software engineer and amateur Go player I have been working on a Go playing program in the 90s. Our program never exceeded the beginner level while playing the small 9x9 board, but we had hope of challenging the strongest programs of the time which I could give a 5-7 stone handicap advantage and still beat, not being that strong myself.
Over time the programs slowly got better, approaching and finally surpassing my level in the early naughties, but they were still no match for the stronger amateur players, never mind professional Go players.
It was generally accepted that Go was too big a challenge for AI.

Then the AlphaGo team got rolling and quickly surpassed even the strongest Go players, each iteration completely destroying the previous one in playing strength.
The current version, AlphaZero, has a crazy high ELO rating. It is absolutely obvious that no human player will ever come near that level.

Once the learning gets more refined the power needed to get there diminishes, so instead of using a purpose-built supercomputer you can now have top professional strength from a single PC with Leela Zero.

I expect the same will happen with FSD. The neural net will keep refining and condensing to a limited set of really important rules, reducing the computational power needed for each level, allowing for raising the bar continuously.