Elon: FSD Beta tweets

[JupiterMan]

Must be employee only Notes. NO WAY this:

- Added voice drive-notes. After an intervention, you can now send Tesla an anonymous voice message describing your experience to help improve Autopilot.

Makes it into the consumer version. Can you imagine Tesla hiring >50k people to sit and listen to a >million audio recordings every day?

If wide release I would imagine voice notes would be transcribed automatically and then each clip filed according to key words/phrases.

It is possible that for many clips having such a description could speed up a resolution, as saves Tesla having to work out why the button was pressed.

 

[Supcom]

If wide release I would imagine voice notes would be transcribed automatically and then each clip filed according to key words/phrases.

It is possible that for many clips having such a description could speed up a resolution, as saves Tesla having to work out why the button was pressed.

Exactly. If the voice comments were for employees only, then why make them anonymous?

 

[qdeathstar]

Exactly. If the voice comments were for employees only, then why make them anonymous?

So musk don’t fire them…

 

[JulienW]

Exactly. If the voice comments were for employees only, then why make them anonymous?

But if Tesla was interested in our feedback then why was the snapshot button removed in the first place? Makes no sense to remove data collection but then add it back with even more data collected.

 

[willow_hiller]

But if Tesla was interested in our feedback then why was the snapshot button removed in the first place? Makes no sense to remove data collection but then add it back with even more data collected.

To be honest, the snapshot button was a little dangerous. I could see them deciding to disable it while they worked on the voice activated version for this reason.

Right after your car does something potentially dangerous, you were expected to take your eyes off the road, one hand off the wheel, and lean over to press a tiny button at the top of the screen.

 

[JulienW]

To be honest, the snapshot button was a little dangerous. I could see them deciding to disable it while they worked on the voice activated version for this reason.

Right after your car does something potentially dangerous, you were expected to take your eyes off the road, one hand off the wheel, and lean over to press a tiny button at the top of the screen.

...and how do you access Voice Drive-Notes?

 

[2101Guy]

To be honest, the snapshot button was a little dangerous. I could see them deciding to disable it while they worked on the voice activated version for this reason.

Right after your car does something potentially dangerous, you were expected to take your eyes off the road, one hand off the wheel, and lean over to press a tiny button at the top of the screen.

yep. All the while being told in writing "keep both hands on the wheel/yoke at all times".

 

[DrChaos]

Must be employee only Notes. NO WAY this:

- Added voice drive-notes. After an intervention, you can now send Tesla an anonymous voice message describing your experience to help improve Autopilot.

Makes it into the consumer version. Can you imagine Tesla hiring >50k people to sit and listen to a >million audio recordings every day?

Whisper AI voice transcription from OpenAI is amazingly good now. It will surely be all automated transcription. And most of it will be ignored.

Introducing Whisper

We’ve trained and are open-sourcing a neural net called Whisper that approaches human level robustness and accuracy on English speech recognition. Read Paper View Code View Model Card Whisper examples: Reveal Transcript Whisper is an automatic speech...

openai.com

 

[Knightshade]

The transcription concern seems like a weird red herring?

Voice transcription being good or not isn't the problem.

Voice or text, you still need someone (a LOT of someones) on the receiving end of the data in whatever form to actually DO something with feedback from a fleet of millions of cars.

 

[sleepydoc]

If wide release I would imagine voice notes would be transcribed automatically and then each clip filed according to key words/phrases.

It is possible that for many clips having such a description could speed up a resolution, as saves Tesla having to work out why the button was pressed.

I’d imagine that text transcriptions of comments may actually be more useful than a bunch of video clips. They’d certainly be easier to screen and categorize.

 

[Supcom]

But if Tesla was interested in our feedback then why was the snapshot button removed in the first place? Makes no sense to remove data collection but then add it back with even more data collected.

Because the car had gotten the capability to self-report any disengagement, so the only use the report button had was to flag a non-disengagement where the driver didn't like what the car did. But that button had no means for the driver to indicate what the issue was. A voicd comment allows a driver to state what the issue was that caused her to disengage. Most of these are likely not to be directly reviewed by Tesla engineers, but, through voice recognition, could be stored and searched by keywords or phrases to assist identifying problems that might not be apparent.

 

[willow_hiller]

...and how do you access Voice Drive-Notes?

It's pressing physical thumb-button on the steering wheel, followed by speaking. No hands off the wheel or eyes off the road.

How did you think it worked?

 

[aronth5]

But if Tesla was interested in our feedback then why was the snapshot button removed in the first place? Makes no sense to remove data collection but then add it back with even more data collected.

It made perfect sense to remove the button if Tesla determined they weren't providing adequate value. Many people posted they wish they had a way to add context to the button so adding voice notation does that.

 

[qdeathstar]

The transcription concern seems like a weird red herring?

Voice transcription being good or not isn't the problem.

Voice or text, you still need someone (a LOT of someones) on the receiving end of the data in whatever form to actually DO something with feedback from a fleet of millions of cars.

That’s what India exists for…

 

[JulienW]

It's pressing physical thumb-button on the steering wheel, followed by speaking. No hands off the wheel or eyes off the road.

How did you think it worked?

Likely not how it will works. If you just start talking it will think it is a Command it doesn't understand. You will probably have to say a specific phrase or key word since it needs to know if this is a Command, Bug Report or Voice Drive Notes. For instance if logging a bug you must say "Bug Report" first and then say what it is.

The release Notes are completely ambiguous on how this would work. You can't just say "added Voice Drive Notes" and expect people to know how it works. Before they had instructions on how to use the Report Button.

 

[qdeathstar]

More than likely this is some sort of nhtsa requirement.

 

[Supcom]

More than likely this is some sort of nhtsa requirement.

I very much doubt that.

 

[qdeathstar]

I very much doubt that.

Why?

 

[DarkForest]

Must be employee only Notes. NO WAY this:

- Added voice drive-notes. After an intervention, you can now send Tesla an anonymous voice message describing your experience to help improve Autopilot.

Makes it into the consumer version. Can you imagine Tesla hiring >50k people to sit and listen to a >million audio recordings every day?

I could see this being a part of the auto-labeling.

It could be looking for keywords like "phantom braking" or "wrong lane choice," and that could be fed into the auto-labeler or another mechanism to try to automatically figure out what the NNs should have done. The car itself could throw out a report if none of the keywords are present.

Even if 90% of the reports are thrown out, those 10% that are kept can still be useful to the NN training while also making drivers feel like they're being heard (even if they aren't).

It might not even trigger saving the clip to begin with but could be added to any clip the car was going to save anyway. It would basically work how it is today but with the added benefit of the clips having user-submitted keywords attached.

 

[kabin]

Of course audio/audio to text is much less memory intensive and will give the average FSD driver a sense of value and hope that the team cares enough to listen. But the team isn't required to listen and likely won't.

I agree with others here. There would need to be specific communication instructions to be effective. Possible keywords, max duration, max frequency/drive, .... Most likely events would be excessively emotional with a few curse words sprinkled in which would discourage the team unless they want a good laugh.