5 ms·

I suspect the long term plan is to create a training dataset labelled by the 3000 people and, when they have sufficient training data, let machine learning / AI

I suspect the long term plan is to create a training dataset labelled by the 3000 people and, when they have sufficient training data, let machine learning / AI take over.

partisan 9y ago
Do things that don't scale... Until you can scale. I think you are absolutely correct on this point.
itburnslikeice 9y ago
I assumed this too, those datasets would worth too many $$$
hetspookjee 9y ago
Imagine the horror that awaits the 3000.
tyingq 9y ago
100% automated ML doing this well seems far off to me. Google's ML for rich snippets thinks a quarter is worth 50 cents. http://imgur.com/1nNreR2 http://imgur.com/1nNreR2 Really parsing language for meaning is hard. Sentiment analysis, which seems simpler than this problem, still has relatively poor accuracy. I could see it perhaps thinning out how much has to be done manually, but I don't see it removing that need entirely.
- akvadrako 9y ago
  That information is so bad for such an easy question, I can only wonder if google is more interested in using their userbase as a training ground than actually providing a useful service.
  - tyingq 9y ago
    It's not an outlier either. I can find about 1 obviously bad answer a minute just randomly typing fact type questions that pop into my head.
- Spooky23 9y ago
  That's amazing. Nearly every fact presented there is completely wrong. Plus, the picture is a standing liberty quarter that is nearly 90 years old and pretty much completely out of general circulation.
  - fdsfsafasfdas 9y ago
    Could you specifically address his concerns for those not in-the-know about ML?
    - tyingq 9y ago
      I believe "Nearly every fact presented there is completely wrong" was referring to Google's snippet in the screenshot, not my comment.
    - Spooky23 9y ago
      Sorry, my comment was ambiguous -- I was referring to the Google result. Thanks for pointing it out. The commenter that I was responding to made a great point that I agree with and was highlighting as a particularly good example of ML gone awry in the wild. Other than the subjective statement that quarters are "useful", every fact in the google response is wrong. Quarters are: - Worth $0.25 (not $0.50) - Not the largest. (Dollar coins and half-dollars are larger) - Made of a blend of copper and nickel. (No silver content for circulating coins) The picture depicts a standing liberty quarter that is no longer a circulating coin. It's amazing that Google would put that in production.
- freyir 9y ago
  I thought this screenshot may have been from an early and outdated rollout of the Rich Snippets, but no, I just searched and Google still says a quarter is worth 50 cents.
  - Mtinie 9y ago
    Easily explained as an artifact of just how accurate Google's ML systems really are. What you're seeing is a hyper-accurate snapshot of the exchange rate between "a shave and a hair cut" and "two bits"[1]. Typically you hear about the value over a much longer period of time and the average value of "a U.S. Quarter Dollar" == $0.25. /snark ;) [1] "Two bits" -- Noun. Slang term used predominantly by Neo-Victorian communities living in California's Bay Area megalopolis. Edit: Fixin' of the grammar
- stefantalpalaru 9y ago
  > 100% automated ML doing this well seems far off to me. They do it for assassination targets: https://arstechnica.co.uk/security/2016/02/the-nsas-skynet-program-may-be-killing-thousands-of-innocent-people/ https://arstechnica.co.uk/security/2016/02/the-nsas-skynet-p... Bugs don't matter when there's no way for those affected to report them.
retox 9y ago
Last time facebook tried that they hired a bunch of lefties (collage kids are cheap labour) and they got pilloried for partisanship.
- celticninja 9y ago
  Are they kids that make artwork from newspaper and magazine clippings?
- tyleraldrich 9y ago
  Do you have a source for this?
  - retox 9y ago
    This report explicitly blacklisting and injecting news people weren't interested in. No mention of college students, so that is probably some bullshit I picked up somewhere, I wasn't following the story too closely. https://www.gizmodo.com.au/2016/05/former-facebook-workers-we-routinely-suppressed-conservative-news/ https://www.gizmodo.com.au/2016/05/former-facebook-workers-w...
erikpukinskis 9y ago
You are describing the past. That work is already done. All of these "community moderation" jobs at big companies are just checking the AI's work. If you have a billion posts an hour you're not hand triaging that with a team of 4500 people.
mrfrasha 9y ago
> These reviewers will also help us get better at removing things we don't allow on Facebook like hate speech and child exploitation. And we'll keep working with local community groups and law enforcement who are in the best position to help someone if they need it -- either because they're about to harm themselves, or because they're in danger from someone else. The workers aren't there to be a basis for machine learning (though that is a nice benefit). They are there to communicate with local community groups / law enforcement on an ongoing basis.
sametmax 9y ago
This way you will have even less way to find a solution when you are hurt because of a false positive. I mean, the GAFAs are already not known for their stellar customer service, but if we get rid of all humans in the process, it's going to be Brazil on steroid pretty soon.

I suspect the long term plan is to create a training dataset labelled by the 3000 people and, when they have sufficient training data, let machine learning / AI take over.

partisan 9y ago

Do things that don't scale... Until you can scale. I think you are absolutely correct on this point.

itburnslikeice 9y ago

I assumed this too, those datasets would worth too many $$$

hetspookjee 9y ago

Imagine the horror that awaits the 3000.

tyingq 9y ago

100% automated ML doing this well seems far off to me. Google's ML for rich snippets thinks a quarter is worth 50 cents. http://imgur.com/1nNreR2 http://imgur.com/1nNreR2 Really parsing language for meaning is hard. Sentiment analysis, which seems simpler than this problem, still has relatively poor accuracy. I could see it perhaps thinning out how much has to be done manually, but I don't see it removing that need entirely.

akvadrako 9y ago
That information is so bad for such an easy question, I can only wonder if google is more interested in using their userbase as a training ground than actually providing a useful service.
- tyingq 9y ago
  It's not an outlier either. I can find about 1 obviously bad answer a minute just randomly typing fact type questions that pop into my head.
Spooky23 9y ago
That's amazing. Nearly every fact presented there is completely wrong. Plus, the picture is a standing liberty quarter that is nearly 90 years old and pretty much completely out of general circulation.
- fdsfsafasfdas 9y ago
  Could you specifically address his concerns for those not in-the-know about ML?
  - tyingq 9y ago
    I believe "Nearly every fact presented there is completely wrong" was referring to Google's snippet in the screenshot, not my comment.
  - Spooky23 9y ago
    Sorry, my comment was ambiguous -- I was referring to the Google result. Thanks for pointing it out. The commenter that I was responding to made a great point that I agree with and was highlighting as a particularly good example of ML gone awry in the wild. Other than the subjective statement that quarters are "useful", every fact in the google response is wrong. Quarters are: - Worth $0.25 (not $0.50) - Not the largest. (Dollar coins and half-dollars are larger) - Made of a blend of copper and nickel. (No silver content for circulating coins) The picture depicts a standing liberty quarter that is no longer a circulating coin. It's amazing that Google would put that in production.
freyir 9y ago
I thought this screenshot may have been from an early and outdated rollout of the Rich Snippets, but no, I just searched and Google still says a quarter is worth 50 cents.
- Mtinie 9y ago
  Easily explained as an artifact of just how accurate Google's ML systems really are. What you're seeing is a hyper-accurate snapshot of the exchange rate between "a shave and a hair cut" and "two bits"[1]. Typically you hear about the value over a much longer period of time and the average value of "a U.S. Quarter Dollar" == $0.25. /snark ;) [1] "Two bits" -- Noun. Slang term used predominantly by Neo-Victorian communities living in California's Bay Area megalopolis. Edit: Fixin' of the grammar
stefantalpalaru 9y ago
> 100% automated ML doing this well seems far off to me. They do it for assassination targets: https://arstechnica.co.uk/security/2016/02/the-nsas-skynet-program-may-be-killing-thousands-of-innocent-people/ https://arstechnica.co.uk/security/2016/02/the-nsas-skynet-p... Bugs don't matter when there's no way for those affected to report them.

retox 9y ago

Last time facebook tried that they hired a bunch of lefties (collage kids are cheap labour) and they got pilloried for partisanship.

celticninja 9y ago
Are they kids that make artwork from newspaper and magazine clippings?
tyleraldrich 9y ago
Do you have a source for this?
- retox 9y ago
  This report explicitly blacklisting and injecting news people weren't interested in. No mention of college students, so that is probably some bullshit I picked up somewhere, I wasn't following the story too closely. https://www.gizmodo.com.au/2016/05/former-facebook-workers-we-routinely-suppressed-conservative-news/ https://www.gizmodo.com.au/2016/05/former-facebook-workers-w...

erikpukinskis 9y ago

You are describing the past. That work is already done. All of these "community moderation" jobs at big companies are just checking the AI's work. If you have a billion posts an hour you're not hand triaging that with a team of 4500 people.

mrfrasha 9y ago

> These reviewers will also help us get better at removing things we don't allow on Facebook like hate speech and child exploitation. And we'll keep working with local community groups and law enforcement who are in the best position to help someone if they need it -- either because they're about to harm themselves, or because they're in danger from someone else. The workers aren't there to be a basis for machine learning (though that is a nice benefit). They are there to communicate with local community groups / law enforcement on an ongoing basis.

sametmax 9y ago

This way you will have even less way to find a solution when you are hurt because of a false positive. I mean, the GAFAs are already not known for their stellar customer service, but if we get rid of all humans in the process, it's going to be Brazil on steroid pretty soon.