5 ms·

> I was disappointed to find out that AI ethicists at Google aren't LessWrong style AI ethicists. I've just read one Google AI ethics paper in it's entirety but

> I was disappointed to find out that AI ethicists at Google aren't LessWrong style AI ethicists. I've just read one Google AI ethics paper in it's entirety but it seemed pretty non-insightful and not actionable. This seems like a contradictory criticism to make. The Less Wrong folk, in my perception, mostly dabble in thought experiments (such as Roko’s basilisk) that aren’t terribly relevant to real world AI work. Timnit’s work on identifying racial and gendered inaccuracies in facial recognition [1] seems much more actionable. [1] https://www.technologyreview.com/2020/06/12/1003482/amazon-stopped-selling-police-face-recognition-fight/ https://www.technologyreview.com/2020/06/12/1003482/amazon-s...

ALittleLight 6y ago
Gebru's (and co-author's) work on racial error in facial recognition was done while they were academics at Stanford and MIT respectively and not while a Google AI ethicist to my knowledge. Roko's Basilisk is a banned comment from Less Wrong (though interesting) and doesn't really represent the totality of that site or the associated ideas. https://news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212 https://news.mit.edu/2018/study-finds-gender-skin-type-bias-...
- fossuser 6y ago
  Yeah - the basilisk thing is mostly a joke as far as I can tell and has little to do with the AGI goal alignment friendly AI work or the control problem. Miri is mostly focused on the AGI control problem and thinks that AGI is closer than others believe. If true, all other problems are pretty irrelevant. There is real work in ML bias and fairness to be done too, but it seems overrun by toxic personalities and partisan politics.
- cauthon 6y ago
  I’m not sure which Google AI paper you’re referencing. The recent paper at the center of the controversy did have actionable suggestions regarding model training costs (financial and environmental), appropriate training sets, and generally incorporating domain expertise to avoid misinformation or mistranslations. [1] That’s fair, I haven’t visited LW in a while. I did thoroughly enjoy HPMOR, but thought that most of Eliezer’s other work fell closer to speculation than practice. As other commenters have mentioned, these AI algorithms are already being implemented in the real world with real consequences. Their associated ethical concerns therefore seem more urgent, and they can be acted upon. [1] https://www.technologyreview.com/2020/12/04/1013294/google-ai-ethics-research-paper-forced-out-timnit-gebru/ https://www.technologyreview.com/2020/12/04/1013294/google-a...
  - visarga 6y ago
    Apparently not. Regarding energy usage, which has been discussed at length on HN and reddit, and is actively being researched in thousands of papers on efficient, scalable, edge-capable AI. > “It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources,” What? She should read previous work on the topic first - quantization, sparsification, distillation, fine-tuning pre-trained models, etc. I couldn't find her actionable ideas because she doesn't have any original ones. It's a hard topic which requires a complete rethinking of both algorithm and hardware. And she doesn't recognize that Google is investing in both to do that - TPUs on the hardware side, and improved algorithms such as the Reformer - a variant of the Transformer, reducing complexity from O(n^2) to O(n). Then regarding dataset bias - I couldn't find their actionable ideas. I mean, except telling people to be careful about how they select training data, but nothing about how to replace the model they criticize with an unbiased model that works just as well in the general case. It's easy to point out problems while not providing any solutions and making yourself the critic of those who are doing the hard work. See how she treated Yann LeCun on Twitter - she basically told him to go educate himself on her papers and fuck off because she doesn't have time for debating him, that after denouncing his tweets.

ALittleLight 6y ago

Gebru's (and co-author's) work on racial error in facial recognition was done while they were academics at Stanford and MIT respectively and not while a Google AI ethicist to my knowledge. Roko's Basilisk is a banned comment from Less Wrong (though interesting) and doesn't really represent the totality of that site or the associated ideas. https://news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212 https://news.mit.edu/2018/study-finds-gender-skin-type-bias-...

fossuser 6y ago
Yeah - the basilisk thing is mostly a joke as far as I can tell and has little to do with the AGI goal alignment friendly AI work or the control problem. Miri is mostly focused on the AGI control problem and thinks that AGI is closer than others believe. If true, all other problems are pretty irrelevant. There is real work in ML bias and fairness to be done too, but it seems overrun by toxic personalities and partisan politics.
cauthon 6y ago
I’m not sure which Google AI paper you’re referencing. The recent paper at the center of the controversy did have actionable suggestions regarding model training costs (financial and environmental), appropriate training sets, and generally incorporating domain expertise to avoid misinformation or mistranslations. [1] That’s fair, I haven’t visited LW in a while. I did thoroughly enjoy HPMOR, but thought that most of Eliezer’s other work fell closer to speculation than practice. As other commenters have mentioned, these AI algorithms are already being implemented in the real world with real consequences. Their associated ethical concerns therefore seem more urgent, and they can be acted upon. [1] https://www.technologyreview.com/2020/12/04/1013294/google-ai-ethics-research-paper-forced-out-timnit-gebru/ https://www.technologyreview.com/2020/12/04/1013294/google-a...
- visarga 6y ago
  Apparently not. Regarding energy usage, which has been discussed at length on HN and reddit, and is actively being researched in thousands of papers on efficient, scalable, edge-capable AI. > “It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources,” What? She should read previous work on the topic first - quantization, sparsification, distillation, fine-tuning pre-trained models, etc. I couldn't find her actionable ideas because she doesn't have any original ones. It's a hard topic which requires a complete rethinking of both algorithm and hardware. And she doesn't recognize that Google is investing in both to do that - TPUs on the hardware side, and improved algorithms such as the Reformer - a variant of the Transformer, reducing complexity from O(n^2) to O(n). Then regarding dataset bias - I couldn't find their actionable ideas. I mean, except telling people to be careful about how they select training data, but nothing about how to replace the model they criticize with an unbiased model that works just as well in the general case. It's easy to point out problems while not providing any solutions and making yourself the critic of those who are doing the hard work. See how she treated Yann LeCun on Twitter - she basically told him to go educate himself on her papers and fuck off because she doesn't have time for debating him, that after denouncing his tweets.