Hacker News Reader

Hacker News

Comment by hkmurakami | Hacker News Reader

Parent Full thread

hkmurakami

Well for GM, 80k cars doesn't move the needle.

GM sold 320k cars in December 2016 alone.

https://www.wsj.com/articles/auto-industry-poised-to-set-ann...

View on HN

And that's only in the US -- GM sells something like 10 million cars per year.

[Edit: Over 10 million cars last year: http://money.cnn.com/2017/02/07/news/companies/gm-record-sal... ]

None of those 320k cars GM sold have autopilot hardware sending data back to the manufacturer for self-driving algorithm refinement. All of those Tesla vehicles do (Autopilot 2 hardware). New Teslas are delivered every day that accelerate the rate at which Tesla collects autopilot data.

Hence the comparison.

EDIT: https://electrek.co/2016/11/13/tesla-autopilot-billion-miles...

"Tesla has now 1.3 billion miles of Autopilot data going into its new self-driving program"

There's a reason Tesla picks up the tab for your vehicle's machine to machine cellular connection. And why it connects to wifi when you're at home.

I think you're missing the point that a car manufacturer does not care about autopilot for autopilot's sake. They care about the feature in so far as it can drive sales and revenue.

Think about it this way: Companies like Google or Amazon don't launch products just because the tech is cool or innovative. They only launch products that can become $1B+ drivers of revenue for them.

(fwiw I am not the one downvoting you)

Having a better autopilot than your competitors can strongly drive sales and revenue if that's what the market wants.

>Google or Amazon don't launch products just because the tech is cool or innovative. //

That is a thing though, have cool stuff that serves as advertising to showoff your company as being cool & cutting-edge.

Amazon gets a lot of mindshare out of echo beyond the raw value accrued through sales IMO.

The only thing Tesla vehicles have the bandwidth and storage to send back is "oops i'm crashed".

It's possibly a way to test your implementation by comparing it to what the human does, but that doesn't apply for a million basic things like sign recognition and you can't teach machine learning with that.

Tesla fans seem to believe this very very strongly. Do you get a massive mobile bill from your car? Does your car have a huge storage device in the trunk? Does it upload 50GB of data over your wifi every time you park it in your garage? If not, the amount of data that Tesla collects for this purpose has probably been dramatically overstated.

They're not necessarily uploading continuous video of the entire time you're driving to the mothership - that would indeed be wasteful. More likely, they're monitoring diagnostic values from their autopilot algorithm (like "error in predicted steering angle") and using that to trigger recording of video and sensor values for a short period, which could then be used as "test cases" for future algorithm updates.

When is the autopilot data "good enough" with any further improvements being marginal, i.e. how much of the data being sent back to Tesla is repetitive?

This problem lends itself to horizontal scaling very well, I'd say.

GM sold a lot more cars, but in my line of work 2% definitely counts as "moving the needle."

That figure was comparing Tesla worldwide sales to GM's US sales. GM sold over 10 million cars worldwide in 2016, it's more like 0.75%.

Reply on news.ycombinator.com