7 ms·

To think that a company that has failed to create self-driving for nearly a decade could solve "content moderation", one of the hardest problem on the internet

To think that a company that has failed to create self-driving for nearly a decade could solve "content moderation", one of the hardest problem on the internet that every single of the big companies (who btw hire the best talent) have been struggling with, is truly a shocking statement to me. If all Google and Facebook couldn't solve it, why could a Tesla engineer of all people...

paisawalla 5y ago
We're not talking about the hardest content problems on the internet. Look at any of Elon Musk's (or anyone in crypto's) popular tweets and you'll see a flood of incredibly obvious bots that poorly mimic his account, spamming some shit coin or advertising a BTC address. Also, you're really understating how good Tesla self-drive is. It's not full self-drive but it's a real achievement.
- ehsankia 5y ago
  Yes, and Youtube and Facebook both have similar problems. There is survivalship bias at play here, you do see the bots that manage to get past the spam protection, but never see the ones that don't. For all you know, Twitter could be blocking 99% of the spam. Of course, as long as there's money in it (and there is), spammers will constantly be coming up with new and unexpected ways to bypass your systems. It's a neverending battle, and if you think it's trivial, you are the one understating how hard the problem is. I never claimed Tesla self-driving wasn't an achievement, but clearly they underestimated how hard that would be themselves, so I wouldn't be surprised if they'd do the same for content moderation.
  - paisawalla 4y ago
    There's a substantial qualitative difference between what gets through Youtube's and Facebook's filter, versus what gets past Twitter. They're not even practicing the current state of the art, as practiced by their peers.
  - LewisVerstappen 4y ago
    > Youtube and Facebook both have similar problems. There is survivalship bias at play here, you do see the bots that manage to get past the spam protection, but never see the ones that don't. For all you know, Twitter could be blocking 99% of the spam. If you spend 5 seconds on YouTube and Facebook and compare that to Twitter, you'll see a massive difference. The spam bots on Twitter are incredibly unsophisticated. It's literally bots with the same profile picture and name as the person they're trying to spam. On the other hand, YouTube spam bots are actually incredibly sophisticated. They're using GPT-3 or some language model to generate text and reply to each other. Like, sometimes I'll read a comment and not be sure if it was a spam bot or not. Twitter is leagues behind w/ their spam filters.
    - ehsankia 4y ago
      Obviously YMMV but in 2021, every tech video I commented on would get these replies, which all had 18+ profile pictures. Clicking their profile every account had the exact same "playlist" with a porn site ad inside, and the channel header had the exact same website link. I kept getting these for months.
    - ehsankia 4y ago
      https://www.youtube.com/watch?v=1Cw-vODp-8Y https://www.youtube.com/watch?v=1Cw-vODp-8Y
- itsoktocry 4y ago
  >Also, you're really understating how good Tesla self-drive is How good is it?
  - jeofken 4y ago
    It’s pretty good! Does the vast majority of my driving for me, except for roundabouts, gravel roads, and exceptionally snowy days (Scandinavia). I do long road trips and feel less exhausted than I would otherwise.
  - newaccount74 4y ago
    Good enough often enough that people stop paying attention (and die when the car hits an obstacle the AI doesn't recognise)
  - yellow_lead 4y ago
    Keep in mind Tesla self-driving users have a literal survivorship bias.
    - oneoff786 4y ago
      Not really. Deaths are too low to matter on a statistical level for “literal” survivor bias
  - andrepd 4y ago
    It's so good that Tesla sued people for posting videos where it fails miserably. Really screams "confidence", doesn't it?
    - diebeforei485 4y ago
      Did this actually happen? I know there was one instance when an engineer using an internal build got fired for vlogging it (which is not at all unusual, Apple would also fire you for showing off an internal build).
mavhc 5y ago
60,000 people have FSD beta, they've got further than everyone else.
- ehsankia 4y ago
  Tesla FSD is level 3. You still need to be fully aware and hands on the wheel. Level 4 is when no one needs to be at the driver seat, which is what Cruise and Waymo are doing. Quantity != Quality. Just because McDonalds serves billions of people doesn't magically make it good food.
  - mavhc 4y ago
    Does make it safe food though
syshum 4y ago
A large part of the problem for twitter Google and Facebook is their ever shifting goals for content moderation to appease their politically activist employees and media critics. Something tells me Elon would have different goals that should be less of a moving target and easier to accomplish. Less subjective target like "hate speech" and more objective target like true threats of violence, illegal activity, actual bot detection not "people who disagree are bots", etc
- athrowaway3z 4y ago
  No. Not relevant to the discussion. Gmail has no content moderation, but spam still gets through occasionally.
  - _benedict 4y ago
    Spam volumes are also down dramatically due to conventional means like making it illegal, and taking down bot farms that were sending it. Also the introduction of DKIM and DMARC making it harder across the board to be seen as a legitimate sender, and the fact that many-to-many emails are a huge red flag for spam filters, and not a concept for social media. Spam is a dramatically easier problem, and has many more mechanism to suppress it both legally and technologically.
  - irrational 4y ago
    Occasionally? Recently (maybe the last 6 months) I have been seeing a huge amount of spam getting through (comparatively - maybe 10 a day make it to my inbox). Previously I rarely saw any.
  - mywacaday 4y ago
    In the last two weeks my gdrive has been spammed with shared porn pdfs. It's not even being shared with the email address I use, it's being shared with my email without a full stop in the middle that Gmail ignores. Haven't looked into it much but apparently people have been asking for years for an option to only allow a doc to be shared from a know contact, Google has ignored this for a long time.
  - 4y ago
chippiewill 4y ago
The idea that an AI engineer for self driving could be trivially retasked to content moderation is fairly laughable as well. That's not to say there aren't overlapping skill sets, but the AI tech involved is quite distinct (although there is overlap, I believe some of the latest perception research is trying to adapt transformers from NLP to replace CNNs). Many of my colleagues are AI/ML engineers working on the self driving problems (perception, prediction, planning) and they're all super smart, but they'd still take a while to get up to speed in another area. I have no idea why a serious business would hire a non-NLP specialised engineer to solve content moderation (they'd maybe hire someone who was interested in changing sub-field, but not to build out the backbone of their teams).
- rgallagher27 4y ago
  Also to think that an engineer who’s intentionally joined a company to work on AI self driving tech wouldn’t immediately quit when they where forced to move and work on Twitter content moderation is laughable. They’d find another job faster than Musk changes moods.
- zo1 4y ago
  I also find it "laughable" how literally everyone is taking the comment. They weren't saying they'd just take some computer-vision expert and retool him to fix all of Twitter's spam problems.
bfz 4y ago
Spamming "$$$ TSLA to the moon! Get latest tips on my free Discord http://biy.ly/ocozxc http://biy.ly/ocozxc #TWTR #GOOG #AMC #GME $$$" thousands of times for days being suspended is hardly the hardest problem on the Internet
outsb 4y ago
Spamming "$$$ $TSLA to the moon! Get latest tips on my free Discord bit.ly/ocozxc #TWTR #GOOG #AMC #GME $$$" thousands of times for days before suspension is hardly the hardest problem on the Internet Heck, you could even just limit the number of hashtags in a single tweet (say, 2) and fix 95% of Twitter's spam problem without harming legitimate users in any way
- ljm 4y ago
  Supposing it really was that simple, 5% would still comprise a large number of tweets.
- aesthesia 4y ago
  Uh, wouldn't spammers just...stop using lots of hashtags? Part of what makes the problem difficult is that spammers are agents who respond to the techniques you use to stop them.
SergeAx 4y ago
1) I think Elon Musk's angle on Twitter spam is rather different than ours. Look at replies under his tweets. 2) I am not an ML engineer, but still a software engineer and in my opinion the problem of self driving cars is at least two orders of magnitude harder than the problem of Internet spam. Maybe even three orders of magnitude. 3) We actually know that the problem of spam is successfuly solved by Google. When was the last time you saw a spam message in Gmail? I am seeing them in my inbox approximately once a month. There are about 8 messages in my spam folder per day, so it is about 0.4% false negatives. This is a great result. 4) Social networks are not interested in fighting spam. It just will make their metrics worse. All they need is to keep it at moderately acceptable level, so users won't flee in bunches.

paisawalla 5y ago

We're not talking about the hardest content problems on the internet. Look at any of Elon Musk's (or anyone in crypto's) popular tweets and you'll see a flood of incredibly obvious bots that poorly mimic his account, spamming some shit coin or advertising a BTC address. Also, you're really understating how good Tesla self-drive is. It's not full self-drive but it's a real achievement.

ehsankia 5y ago
Yes, and Youtube and Facebook both have similar problems. There is survivalship bias at play here, you do see the bots that manage to get past the spam protection, but never see the ones that don't. For all you know, Twitter could be blocking 99% of the spam. Of course, as long as there's money in it (and there is), spammers will constantly be coming up with new and unexpected ways to bypass your systems. It's a neverending battle, and if you think it's trivial, you are the one understating how hard the problem is. I never claimed Tesla self-driving wasn't an achievement, but clearly they underestimated how hard that would be themselves, so I wouldn't be surprised if they'd do the same for content moderation.
- paisawalla 4y ago
  There's a substantial qualitative difference between what gets through Youtube's and Facebook's filter, versus what gets past Twitter. They're not even practicing the current state of the art, as practiced by their peers.
- LewisVerstappen 4y ago
  > Youtube and Facebook both have similar problems. There is survivalship bias at play here, you do see the bots that manage to get past the spam protection, but never see the ones that don't. For all you know, Twitter could be blocking 99% of the spam. If you spend 5 seconds on YouTube and Facebook and compare that to Twitter, you'll see a massive difference. The spam bots on Twitter are incredibly unsophisticated. It's literally bots with the same profile picture and name as the person they're trying to spam. On the other hand, YouTube spam bots are actually incredibly sophisticated. They're using GPT-3 or some language model to generate text and reply to each other. Like, sometimes I'll read a comment and not be sure if it was a spam bot or not. Twitter is leagues behind w/ their spam filters.
  - ehsankia 4y ago
    Obviously YMMV but in 2021, every tech video I commented on would get these replies, which all had 18+ profile pictures. Clicking their profile every account had the exact same "playlist" with a porn site ad inside, and the channel header had the exact same website link. I kept getting these for months.
  - ehsankia 4y ago
    https://www.youtube.com/watch?v=1Cw-vODp-8Y https://www.youtube.com/watch?v=1Cw-vODp-8Y
itsoktocry 4y ago
>Also, you're really understating how good Tesla self-drive is How good is it?
- jeofken 4y ago
  It’s pretty good! Does the vast majority of my driving for me, except for roundabouts, gravel roads, and exceptionally snowy days (Scandinavia). I do long road trips and feel less exhausted than I would otherwise.
- newaccount74 4y ago
  Good enough often enough that people stop paying attention (and die when the car hits an obstacle the AI doesn't recognise)
- yellow_lead 4y ago
  Keep in mind Tesla self-driving users have a literal survivorship bias.
  - oneoff786 4y ago
    Not really. Deaths are too low to matter on a statistical level for “literal” survivor bias
- andrepd 4y ago
  It's so good that Tesla sued people for posting videos where it fails miserably. Really screams "confidence", doesn't it?
  - diebeforei485 4y ago
    Did this actually happen? I know there was one instance when an engineer using an internal build got fired for vlogging it (which is not at all unusual, Apple would also fire you for showing off an internal build).

mavhc 5y ago

60,000 people have FSD beta, they've got further than everyone else.

ehsankia 4y ago
Tesla FSD is level 3. You still need to be fully aware and hands on the wheel. Level 4 is when no one needs to be at the driver seat, which is what Cruise and Waymo are doing. Quantity != Quality. Just because McDonalds serves billions of people doesn't magically make it good food.
- mavhc 4y ago
  Does make it safe food though

syshum 4y ago

A large part of the problem for twitter Google and Facebook is their ever shifting goals for content moderation to appease their politically activist employees and media critics. Something tells me Elon would have different goals that should be less of a moving target and easier to accomplish. Less subjective target like "hate speech" and more objective target like true threats of violence, illegal activity, actual bot detection not "people who disagree are bots", etc

athrowaway3z 4y ago
No. Not relevant to the discussion. Gmail has no content moderation, but spam still gets through occasionally.
- _benedict 4y ago
  Spam volumes are also down dramatically due to conventional means like making it illegal, and taking down bot farms that were sending it. Also the introduction of DKIM and DMARC making it harder across the board to be seen as a legitimate sender, and the fact that many-to-many emails are a huge red flag for spam filters, and not a concept for social media. Spam is a dramatically easier problem, and has many more mechanism to suppress it both legally and technologically.
- irrational 4y ago
  Occasionally? Recently (maybe the last 6 months) I have been seeing a huge amount of spam getting through (comparatively - maybe 10 a day make it to my inbox). Previously I rarely saw any.
- mywacaday 4y ago
  In the last two weeks my gdrive has been spammed with shared porn pdfs. It's not even being shared with the email address I use, it's being shared with my email without a full stop in the middle that Gmail ignores. Haven't looked into it much but apparently people have been asking for years for an option to only allow a doc to be shared from a know contact, Google has ignored this for a long time.
- 4y ago

chippiewill 4y ago

The idea that an AI engineer for self driving could be trivially retasked to content moderation is fairly laughable as well. That's not to say there aren't overlapping skill sets, but the AI tech involved is quite distinct (although there is overlap, I believe some of the latest perception research is trying to adapt transformers from NLP to replace CNNs). Many of my colleagues are AI/ML engineers working on the self driving problems (perception, prediction, planning) and they're all super smart, but they'd still take a while to get up to speed in another area. I have no idea why a serious business would hire a non-NLP specialised engineer to solve content moderation (they'd maybe hire someone who was interested in changing sub-field, but not to build out the backbone of their teams).

rgallagher27 4y ago
Also to think that an engineer who’s intentionally joined a company to work on AI self driving tech wouldn’t immediately quit when they where forced to move and work on Twitter content moderation is laughable. They’d find another job faster than Musk changes moods.
zo1 4y ago
I also find it "laughable" how literally everyone is taking the comment. They weren't saying they'd just take some computer-vision expert and retool him to fix all of Twitter's spam problems.

bfz 4y ago

Spamming "$$$ TSLA to the moon! Get latest tips on my free Discord http://biy.ly/ocozxc http://biy.ly/ocozxc #TWTR #GOOG #AMC #GME $$$" thousands of times for days being suspended is hardly the hardest problem on the Internet

outsb 4y ago

Spamming "$$$ $TSLA to the moon! Get latest tips on my free Discord bit.ly/ocozxc #TWTR #GOOG #AMC #GME $$$" thousands of times for days before suspension is hardly the hardest problem on the Internet Heck, you could even just limit the number of hashtags in a single tweet (say, 2) and fix 95% of Twitter's spam problem without harming legitimate users in any way

ljm 4y ago
Supposing it really was that simple, 5% would still comprise a large number of tweets.
aesthesia 4y ago
Uh, wouldn't spammers just...stop using lots of hashtags? Part of what makes the problem difficult is that spammers are agents who respond to the techniques you use to stop them.

SergeAx 4y ago

1) I think Elon Musk's angle on Twitter spam is rather different than ours. Look at replies under his tweets. 2) I am not an ML engineer, but still a software engineer and in my opinion the problem of self driving cars is at least two orders of magnitude harder than the problem of Internet spam. Maybe even three orders of magnitude. 3) We actually know that the problem of spam is successfuly solved by Google. When was the last time you saw a spam message in Gmail? I am seeing them in my inbox approximately once a month. There are about 8 messages in my spam folder per day, so it is about 0.4% false negatives. This is a great result. 4) Social networks are not interested in fighting spam. It just will make their metrics worse. All they need is to keep it at moderately acceptable level, so users won't flee in bunches.