Hacker News Reader

Comment by 1sembiyan | Hacker News Reader

> Teams at the San Mateo office were tasked with evaluating customer vehicle data related to the Autopilot driver-assistance features and performing so-called data labeling.

My sympathies with the folks laid off, but do American firms still do data labeling in-house? I assume most of it is out-sourced via Scale or other vendors to eventually India, Vietnam etc.

View on HN

So Tesla's $750 million dollar state funded factory in Buffalo, which was supposed to produce solar panels as Solar City, then solar roofs as Tesla, instead employs unskilled data labeling workers.

Recently NY state sold off all the solar panel equipment, and Panasonic pulled out of the partnership. Tesla hired minimum wage data taggers to avoid fines under the job creation agreement.

One of the worst public subsidy scams in state history if you ask me, alongside Scott Walker's Foxconn giveaway in Wisconsin. I wish state governments would collectively agree to stop falling for this.

https://www.investigativepost.org/2021/12/28/tesla-reaches-i...

If I recall correctly, Karpathy during one of the Tesla AI talks addresses this and said that keeping data labeling in-house was an advantage because it could be integrated more tightly into their processes with better customization & turnaround etc.

Firing most of their data-cleaning & labeling capacity isn't going to be good for their car efforts, but I suppose if things are that bad at Tesla now, it's better to fire to cut expenses to the bone, and rehire much later once the storm is past...

The Tesla bull will say that Tesla’s ML has gotten so good that these top-shelf human curators are now unnecessary, and can be outsourced or dispensed with entirely. The Tesla bear will counter that this is merely to keep up appearances; it looks a hell of a lot better to fire data curators over machine learning engineers, even if the former are still just as (if not more!) necessary as the latter.

They still have a lot of human curators (Elon said they have over 1500 in a recent interview). I imagine that these cuts are for some category of labeling that has become redundant.

This is really the most realistic answer. As they get closer and closer to FSD (and it is quite close now), they will need fewer and fewer people.

How do you know its close?

You can watch user videos of it in action on YouTube. It improves with (nearly) every patch and often requires minimal or no user interventions. It still sometimes makes very bad mistakes though so I don’t think it’s as close as the previous poster does. But I think in the next year or two it could be good enough for many cities during clear weather.

As someone using FSD Beta, I concur that it is getting significantly better with every new over the air update.

> The Tesla bear will counter that this is merely to keep up appearances

I would be very surprised if Tesla optimized for optics here.

Not to mention the release of Tesla bots due to September event.

Quality data is one of the most important parts of building ml models. It’s probably a competitive advantage

It doesn’t have to be all-in or all-out. They may outsource the bulk of it and bring in house borderline, not clear cut, or critical data points.

Tesla insourced all of this. They have their own labeling tools and they are labeling in 360% video.

Reply on news.ycombinator.com