4 ms·

I lost the assumption of good faith already in her first tweet, in particular >You can’t just reduce harms caused by ML to dataset bias. She's already attacki

I lost the assumption of good faith already in her first tweet, in particular >You can’t just reduce harms caused by ML to dataset bias. She's already attacking a strawman right there. Yann did not deny any harms caused by ML.

joshuamorton 6y ago
Nor did Timnit claim that he was. Her disagreement was about the causes of harms, not the existence of resulting harms.
- dekhn 6y ago
  and I believe that Yann knows 100X about the causes of harms than Timnit does- so lecturing him is just wasting everybody's time.
  - joshuamorton 6y ago
    You've done a phd. Why on earth would you believe that someone, even an expert in the broad field, would know more about a particular topic than someone who specializes in that area is research? Do you think Yann knows 100X more about every area of ML then everyone else, or is it just fairness, accountability, and transparency that he happens to be more knowledgeable in than arguably a founder of the subfield?
    - dekhn 6y ago
      It's pretty simple: he did made ML work before almost anybody else did, and kept working on during the deep network explosion, and is now running Facebook AI, which has to deal with these sorts of problems with practical solutions on a daily basis, with billions of users. That sort of daily experience counts so much that I would I would place him in the "knows 100X about every area of ML" (excluding rare subfields). It's rare but I have encountered people outside my field who knew more than I did about my field, because of their daily experiences over decades, or their raw intelligence. Yann seems to have both.
      - joshuamorton 6y ago
        So are you suggesting that Yann has more expertise on what you're working on now (which I know, and would consider to be an ML subfield in the same vein as Timnit's), and therefore would defer to his expertise when he says things that show nothing more than an undergraduate level of the topic? Because I'm only a dilettante in the AI ethics space (and admittedly ML as a whole), and I can describe the flaws in Yann's reasoning. Blind deference of that level isn't rational.
        dekhn 6y ago
        I've argued with Yann about my field (we share a friend on facebook) and in that case, he did turn out to be technically correct. I don't agree with your assessment of his "Reasoning".
- rndgermandude 6y ago
  There was no disagreement. Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers). In particular Yann did not suggest in any way a) there are no harms or b) that harms are only related to biased training sets. Yann was commenting on the outcome of a particular research project and how they used a biased training set resulting in the outcome that was observed. Timnit brought up harms first, then pretended Yann did marginalize such harms and attributed them solely to a biased training sets. And then viciously attacked that strawman. That's a bad faith argument. I can appreciate that she might have been indeed generally sick and tired as she writes, and can appreciate that sick and tired people will not always manage to be nice or overcome their own biases and assume good faith all the time from the other party; we're all human after all. But that doesn't change anything about her argument being made in bad faith.
  - joshuamorton 6y ago
    > Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers) This feels like unreasonable semantics. The dangers are precisely the danger of causing harm. The harms therefore are concrete results of theoretical dangers manifesting. They aren't different. > In particular Yann did not suggest in any way a) there are no harms I agree, and I've said as much. > b) that harms are only related to biased training sets He did, insofar as he suggested that the dangers were due solely to bias in the training set, which is implied when he says that if you train the same model on different data, everyone looks African. Like yes, that is true, but it doesn't reduce the harms (or dangers, if you want to be precise). It just creates a different set of biases with a different set of potential harms (which again, are "dangers"). I'm not seeing a strawperson.

joshuamorton 6y ago

Nor did Timnit claim that he was. Her disagreement was about the causes of harms, not the existence of resulting harms.

dekhn 6y ago
and I believe that Yann knows 100X about the causes of harms than Timnit does- so lecturing him is just wasting everybody's time.
- joshuamorton 6y ago
  You've done a phd. Why on earth would you believe that someone, even an expert in the broad field, would know more about a particular topic than someone who specializes in that area is research? Do you think Yann knows 100X more about every area of ML then everyone else, or is it just fairness, accountability, and transparency that he happens to be more knowledgeable in than arguably a founder of the subfield?
  - dekhn 6y ago
    It's pretty simple: he did made ML work before almost anybody else did, and kept working on during the deep network explosion, and is now running Facebook AI, which has to deal with these sorts of problems with practical solutions on a daily basis, with billions of users. That sort of daily experience counts so much that I would I would place him in the "knows 100X about every area of ML" (excluding rare subfields). It's rare but I have encountered people outside my field who knew more than I did about my field, because of their daily experiences over decades, or their raw intelligence. Yann seems to have both.
    - joshuamorton 6y ago
      So are you suggesting that Yann has more expertise on what you're working on now (which I know, and would consider to be an ML subfield in the same vein as Timnit's), and therefore would defer to his expertise when he says things that show nothing more than an undergraduate level of the topic? Because I'm only a dilettante in the AI ethics space (and admittedly ML as a whole), and I can describe the flaws in Yann's reasoning. Blind deference of that level isn't rational.
      - dekhn 6y ago
        I've argued with Yann about my field (we share a friend on facebook) and in that case, he did turn out to be technically correct. I don't agree with your assessment of his "Reasoning".
rndgermandude 6y ago
There was no disagreement. Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers). In particular Yann did not suggest in any way a) there are no harms or b) that harms are only related to biased training sets. Yann was commenting on the outcome of a particular research project and how they used a biased training set resulting in the outcome that was observed. Timnit brought up harms first, then pretended Yann did marginalize such harms and attributed them solely to a biased training sets. And then viciously attacked that strawman. That's a bad faith argument. I can appreciate that she might have been indeed generally sick and tired as she writes, and can appreciate that sick and tired people will not always manage to be nice or overcome their own biases and assume good faith all the time from the other party; we're all human after all. But that doesn't change anything about her argument being made in bad faith.
- joshuamorton 6y ago
  > Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers) This feels like unreasonable semantics. The dangers are precisely the danger of causing harm. The harms therefore are concrete results of theoretical dangers manifesting. They aren't different. > In particular Yann did not suggest in any way a) there are no harms I agree, and I've said as much. > b) that harms are only related to biased training sets He did, insofar as he suggested that the dangers were due solely to bias in the training set, which is implied when he says that if you train the same model on different data, everyone looks African. Like yes, that is true, but it doesn't reduce the harms (or dangers, if you want to be precise). It just creates a different set of biases with a different set of potential harms (which again, are "dangers"). I'm not seeing a strawperson.