[icanhackit]

>> Being able to test your code out in the real world and dark launch a feature gives Tesla a ridiculous advantage.

> Is the comment I'm responding to.

And that dark launching includes sending the latest neural net builds to the fleet to safely shadow test the changes, collecting millions of hours of real-world telemetry and rapidly iterating without end users even needing to pick their nose. You either intentionally or through ignorance reduced OP's statement to being an OTA ECU update that you "could literally put that together with scraps in my garage".

But feel free to keep shifting the goal posts.

[BoorishBears]

I hate it when people say "keep shifting goal posts" in the most random contexts to sound pithy.

I said no OEM wants their engineers to have a capability and you're fixating on the uses of the capability.

My comment is saying it doesn't matter if they wanted to use it for something as trivial as ECU updates, They. Do. Not. Want. It.

If you want a case study in why, AP regressions are a fine one.

They don't want the liability when an updates goes wrong and kills people. They do not see people's lives as prod vs dev. They do not want to hotfix safety critical systems. They want to get it right the first time, and that's why they're conservative and by god, they get it right a lot more than they get it wrong.

Because no matter how much spin TSLA fanboys put on these things, they are liabilities as much as, if not more, than they are strengths, especially combined with the flippant approach to valuing human life the company that called Adaptive Cruise Control + Lane Centering "Autopilot"

[icanhackit]

> My comment is saying it doesn't matter if they wanted to use it for something as trivial as ECU updates

What I'm saying is that you've completely missed the OP's point: The advantage Tesla has in the autopilot space is that they are able to have hundreds of thousands of real-world test units for their autopilot systems by running the neural net in parallel with the driver inputs (i.e. shadow mode) versus companies that at best have a few hundred test units thus limited material to improve/iterate on their deep learning models.

A metaphor for the situation is Apple playing catchup in the AI and voice recognition space to Google because Google had years of material and huge sums of it to build their deep learning models while Apple had comparatively little. The underlying technologies used by either company might be very similar, but you can't throw hardware at the problem to play catch up -- you need that real-world input/learning material for the deep-learning model to be robust.

[BoorishBears]

I haven't completely missed any point, you just seem to be insistent on missing mine.

My point is the way that Tesla has gained this, using consumer vehicles as testbeds, not just for collection, but for actual running code in charge of managing peoples lives is utter nonsense and traditional automakers are choosing to stay away from it.

I mean, this isn't even a hypothetical, we've literally seen it happen, AP regressions where a lane transition your car was taking fine one day suddenly sends your car aimed at a concrete barrier!

You can call them slow, or whatever you want, but thank god the real auto manufacturers are not so flippant about the value of human life.