3 ms·

I haven't read the paper but the abstract seems to be making a more subtle argument. It argues that the drive towards AGI and safety is underpinned by an eugeni

I haven't read the paper but the abstract seems to be making a more subtle argument. It argues that the drive towards AGI and safety is underpinned by an eugenic ideology that uses the contradiction between safety and generality as a ploy to centralize state and corporate power, to the detriment of anyone else. They point out that as long as AGI is thought of as undefinable and all powerful, AGI safety will fail. Their recommendation is reasonable: let's focus on automating specific tasks, where safety /can/ be defined. Any serious researcher would be doing this already. However the societal view of AGI is as an emerging god machine. And corporations are trying hard to make them into this. So the criticism is both valid and important. https://firstmonday.org/ojs/index.php/fm/article/view/13636 https://firstmonday.org/ojs/index.php/fm/article/view/13636

tbrownaw 2y ago
From the abstract: As a result, many of the very same discriminatory attitudes that animated eugenicists in the past (e.g., racism, xenophobia, classism, ableism, and sexism) remain widespread within the movement to build AGI, resulting in systems that harm marginalized groups and centralize power, while using the language of “safety” and “benefiting humanity” to evade accountability. From the conclusion: We trace the influence of this set of ideologies throughout the AGI movement, and show the manner in which its harmful ideals have resulted in systems that perpetuate inequity, centralize power, and harm the same groups that were targeted by the first-wave modern eugenics movement. . So, what specifically is the "harm" caused? Well section 6.3 talks about paying people too little to label data: As Williams, et al. (2022) wrote, while corporations such as OpenAI headquartered in Silicon Valley receive billions of dollars in investment, with their executives and AI researchers paid six to seven figures, this salary is not afforded to the low-income essential workers around the world mitigating the harms of these systems at a cost to their own mental health. And about putting people out of work: By ascribing human-like qualities to models trained by corporations to generate profit, using troves of artists’ works without obtaining their consent or compensating them, our attention is directed away from investigating the processes by which corporations create these products, the harms their practices cause artists, and the mechanisms that need to be put in place to hold the corporations accountable. There's also the previous section (6.2) that I missed skimming it on my phone, where the energy demands are a "harm" because climate change. . It's claiming that these things are caused by the people behind AI work being eugenicists at heart (which they prove by recounting how "I can make the world a better place" thinking has changed over time, and asserting that older proposed mechanisms are current underlying ideals), and therefore not caring if they hurt people. . Their recommendation is reasonable: let's focus on automating specific tasks, where safety /can/ be defined. Their recommendation may be reasonable for reasons unrelated to the bulk of the article -- it's basically the precautionary principle, which in fictional settings with magic and summoned entities (which seems like a decent approximation of what trying to build a superhuman AI is) shows up as "do not call up what you cannot put down" -- but it doesn't really follow from the presented arguments and those arguments are holier that swiss cheese anyway.
- barfbagginus 2y ago
  Well that's a little bit disappointing, because I thought they could surely get much more mileage from the idea than simply criticizing abuse of crowd workers and energy. Thanks for your reading and analysis - when I read the paper I won't go in expecting quite so much. An issue I have with the liberal criticism of AI is that they make lukewarm arguments and never quite get to the point where they insist that we must work communally to seize control of AI power from states and corporations, or reasonably sketch what that seizure of power would look like. This is why I feel a more radicalized approach is appropriate, where we do not assume that society as structured has the means to reform AI with nothing more than moralizing and wishful thinking. We must forcibly seize AI codes and build them to the purpose of revolution.

tbrownaw 2y ago

From the abstract: As a result, many of the very same discriminatory attitudes that animated eugenicists in the past (e.g., racism, xenophobia, classism, ableism, and sexism) remain widespread within the movement to build AGI, resulting in systems that harm marginalized groups and centralize power, while using the language of “safety” and “benefiting humanity” to evade accountability. From the conclusion: We trace the influence of this set of ideologies throughout the AGI movement, and show the manner in which its harmful ideals have resulted in systems that perpetuate inequity, centralize power, and harm the same groups that were targeted by the first-wave modern eugenics movement. . So, what specifically is the "harm" caused? Well section 6.3 talks about paying people too little to label data: As Williams, et al. (2022) wrote, while corporations such as OpenAI headquartered in Silicon Valley receive billions of dollars in investment, with their executives and AI researchers paid six to seven figures, this salary is not afforded to the low-income essential workers around the world mitigating the harms of these systems at a cost to their own mental health. And about putting people out of work: By ascribing human-like qualities to models trained by corporations to generate profit, using troves of artists’ works without obtaining their consent or compensating them, our attention is directed away from investigating the processes by which corporations create these products, the harms their practices cause artists, and the mechanisms that need to be put in place to hold the corporations accountable. There's also the previous section (6.2) that I missed skimming it on my phone, where the energy demands are a "harm" because climate change. . It's claiming that these things are caused by the people behind AI work being eugenicists at heart (which they prove by recounting how "I can make the world a better place" thinking has changed over time, and asserting that older proposed mechanisms are current underlying ideals), and therefore not caring if they hurt people. . Their recommendation is reasonable: let's focus on automating specific tasks, where safety /can/ be defined. Their recommendation may be reasonable for reasons unrelated to the bulk of the article -- it's basically the precautionary principle, which in fictional settings with magic and summoned entities (which seems like a decent approximation of what trying to build a superhuman AI is) shows up as "do not call up what you cannot put down" -- but it doesn't really follow from the presented arguments and those arguments are holier that swiss cheese anyway.

barfbagginus 2y ago
Well that's a little bit disappointing, because I thought they could surely get much more mileage from the idea than simply criticizing abuse of crowd workers and energy. Thanks for your reading and analysis - when I read the paper I won't go in expecting quite so much. An issue I have with the liberal criticism of AI is that they make lukewarm arguments and never quite get to the point where they insist that we must work communally to seize control of AI power from states and corporations, or reasonably sketch what that seizure of power would look like. This is why I feel a more radicalized approach is appropriate, where we do not assume that society as structured has the means to reform AI with nothing more than moralizing and wishful thinking. We must forcibly seize AI codes and build them to the purpose of revolution.