Where did you think the training data was coming from? | Hacker News Reader

Where did you think the training data was coming from?

idiallo.com

50 pointsspeckx14 comments

I thought it was common knowledge that we've been providing free training for machnine learning since the around the first introduction of captcha.

Not sure if the absolute first human verification systems were machine training, but it definitely became that quite quickly.

Like, did everyone just forget about the "provide feedback" button on your search pages? Or when google maps used to ask if its information was accurate?

And the fact that google/youtube/facebook/etc have almost always used your interactions to train algorithms to tune the machine learning not just for you but for everyone?

Why should it be any surprise that every new offering from these companies tracks user data in order to improve the economic efficiency of their models just as every single prior offering from these companies has always done?

This is why I like the kill switch that Framework laptops have to disable the webcam and mic physically. I wish it was standard... especially with some of the more creepy articles of school issued laptop admins being able to "observe" remotely.

I've said it for a long time: the main reason Meta wanted to created glasses or VR/AR is to circumvent the privacy rules that Google and Apple slowly introduced in their App Stores.

More specifically, it's because Meta views the lack of ownership of their own hardware platform as an existential threat. They see AR/VR as the next revolutionary platform, so they're betting the farm on being first movers in a mass-market AR/VR space that they anticipate to exist in the future.

Assuming AR/VR is actually going to be the next thing, how are they going to fight the lock-in due to Google and Apple ecosystem?

Being first isn't going to be enough if Google and Apple manage to release something soon enough and with a better integration with all their other devices

If they manage to offer enough people a good enough product, that'll be a win. But you're right, it could be another fiasco like the HTC First.

If Meta really wanted training data from the glasses, they would give them away.

They could easily afford it, and they would get 1000x the uptake.

meta spies on you, news at 11/s

But if you run any proprietary system, yes they are spying on you, maybe even if you disable what setting you can find. You can tape the camara, but the mic ? I think the only option is to disable the mic in the BIOS.

If you really care about spying, you really should be on Linux or a BSD. IMO, tinfoil hat people should be on OpenBSD, they disable the cam and mic at the OS level and does not have bluetooth at all. FWIW, I still believe bluetooth is a cesspool of security issues.

> You can tape the camara, but the mic ? I think the only option is to disable the mic in the BIOS.

From my experimentation, filling the little microphone hole with blu tack is pretty effective. Not as effective as opening the machine up and cutting out the microphone (my preferred solution), but close.

I really would love an "AR" web. It'd be so cool to walk around a city and get fed real-time information about landmarks, history, cultural events, and shopping destinations—whether through smart glasses or simply a smartphone.

Unfortunately my confidence that any modern tech company could build such an AR web without it being based on horribly-invasive privacy-invading AI-training fascism-enabling user-hostile software is roughly zero. Thus the dream of AR will have to remain just that…a dream. Perhaps the sci-fi of it all should remain "fi".

I think the closest you might expect would the the likes of Pokemon go... though it would be a cool option from the same company/org... though you're right, would likely be a privacy invasion trap that slurped up all the travel data to offer sales to nearby hotels/restaurants to pop up offers...

"Hey, it's getting close to lunch time, would you like some restaurant recommendations nearby?"

Imagine your goal is to create superhuman robots that work like humans

Now also assume that you need data to train the robots and that you have a perspective that visual data will eventually be good enough to capture fine grained movement

So how do you go about capturing that data at the largest scale as fast as possible?

You give people an incentive to wear egocentric video cameras all day so that you can capture 100% of the data and actions from humans as they go through their life in order to transfer it to policy rollouts

Ultimately artificial intelligence is about transferring action policies from humans to machines and the fastest way to do that is to have 100% surveillance of a person

if people actively sign up for that that’s about as good as it gets

"It's people!!! Soylent green is people!!!"

I'm skeptical of this both because this post was largely speculation, and because I would argue Meta has no need to record every private moment for training data.

After all, as the owner of massive social networks, they already have plenty of cleaner, public data that has way more associated labels, like captions, comments, viewer engagement data, and that way they would avoid another data scandal. It's like how people think the Facebook app is secretly recording their microphone, even though their behavior is already obvious through just their and their friends' app engagement. The point of the glasses is to share your footage online, so the recordings would end up public anyways without Meta needing to do anything sneaky.

Reply on news.ycombinator.com