There actually is a science to follow on this, but given the left's skepticism against AI hype, it's doubtful we'll see any success trying to get them to accept ML evals as the science. Instead, look to cybernetics:
"Only variety [complexity] can absorb variety." -- Ashby's Law of Requisite Variety. In other words, a controlling system must be at least as complex as the controlled system in order to maintain regulation of the controlled system's behavior.
This law was penned >50 years ago (!!) but the idea that a law about information and complexity, regardless of substrate, is even possible, remains an incredibly niche position even to technologists and engineers. Cybernetics is nowadays (somewhat) diffused across cultures and scientific disciplines, but largely forgotten as a science in itself, despite its major geopolitical influence.
Those in management or tech have the correct instinct when they recognize that amplifying an individual person's capability to process/generate more data capable of interfacing with the world is literally world-defining. A big part of the ongoing "disruption" is also based on the realization that we have been using computers on the wrong side of the variety equation to primarily generate excess complexity and data, rather than amplifying regulative variety. The accelerationists' hope is that this will also lead to us breaking away from hierarchical organizational structures that can only take regulation so far -- but we've been through this with the initial introduction to computers and we haven't really been able to do that yet, so I'm not super jazzed about that as a natural inevitability.
I was recently introduced to Gebru and Torres' papers and I feel that while you didn't completely misrepresent their writings, particularly after reading Torres, I do feel like this does discard the parts of their work that are valuable. Other parts are not so valuable. Reading Gebru, I am astounded at how she behaves online and seems to promote denialism, yet simultaneously, I felt that she and Torres spend a great deal of time accurately describing the TESCREAL ideologies and correctly label them as a pro-extinction movement based on radical eugenics. Torres also likes to pair his writings with climate change (which seems like a shift in each article, but I respect it) He has also affirmatively stated that if someone is attempting to wipe out the human species he would oppose it. In light of his other writings, one could see the article you cited as a kind of provocation to get people to take this more seriously since anyone reading these arguments will have a fairly predictable emotional reaction of violent rejection.
So my feeling about these two is they are correct to name the radical eliminationism of TESCREAL ideologies that have captured Silicon Valley and a section of the billionaire class. As a software engineer, I am sympathetic to Gebru's idea that we cannot safety test ill defined models, though the cat is firmly out of the bag at this point. I am critical of Gebru in particular for selling denialism of the variety you are calling out here.
I am relatively new to this literature though, and I'm curious as to what you think comrade. I used to believe in at least some elements of the TESCREAL bundle before I rejected them and later became a socialist almost a decade ago, so when I read them, I had an immediate recognition that they were describing a real phenomenon.
I think it is almost without question true that while capability is an open maybe doubtful question, the objective of the tech CEOs like Page, Musk, Altman, Amodei, etc are the creation of a successor species. With the integration of longtermism and ordinary capitalist short-sightedness, they can excuse literally anything in their race to replace human labor and install an transhuman aristocracy. Regardless of whether its ultimate ends are practical, this is an extraordinarily dangerous ideology as there are many intermediate stages with horrible effects.
Gebru, Timnit, and Émile P. Torres. “The TESCREAL Bundle: Eugenics and the Promise of Utopia through Artificial General Intelligence.” First Monday, Apr. 2024. firstmonday.org, https://doi.org/10.5210/fm.v29i4.13636.
On the one hand, Timnit and Torres correctly identify the enemy: the AI companies. The problem is that they actively advocate against any policy that would slow down or pause the development of greater AI capabilities. See this DAIR letter from 2023 Timnit recently referenced, in which they actively advocated against a 6 month pause: https://dair-institute.org/blog/letter-statement-March2023/. They're too committed to denialism to even entertain the possibility that there is *some* chance the AI industry could succeed in their undemocratic ends, which stops them from actually wanting to regulate the development of frontier AI at all. While I agree with a lot of their critiques of the current harms of AI development, their advocacy against trying to slow the pace of AI capabilities growth is quite concerning to me.
This makes a lot of sense to me. I agree that their certainly about lack of capability against evidence is a massive weakness for them and bad for the movement. Thank you for your reply.
There actually is a science to follow on this, but given the left's skepticism against AI hype, it's doubtful we'll see any success trying to get them to accept ML evals as the science. Instead, look to cybernetics:
"Only variety [complexity] can absorb variety." -- Ashby's Law of Requisite Variety. In other words, a controlling system must be at least as complex as the controlled system in order to maintain regulation of the controlled system's behavior.
This law was penned >50 years ago (!!) but the idea that a law about information and complexity, regardless of substrate, is even possible, remains an incredibly niche position even to technologists and engineers. Cybernetics is nowadays (somewhat) diffused across cultures and scientific disciplines, but largely forgotten as a science in itself, despite its major geopolitical influence.
Those in management or tech have the correct instinct when they recognize that amplifying an individual person's capability to process/generate more data capable of interfacing with the world is literally world-defining. A big part of the ongoing "disruption" is also based on the realization that we have been using computers on the wrong side of the variety equation to primarily generate excess complexity and data, rather than amplifying regulative variety. The accelerationists' hope is that this will also lead to us breaking away from hierarchical organizational structures that can only take regulation so far -- but we've been through this with the initial introduction to computers and we haven't really been able to do that yet, so I'm not super jazzed about that as a natural inevitability.
I agree. We need more focus on the Left towards support a positive automated future.
Otherwise we cede this deliberation to others.
I was recently introduced to Gebru and Torres' papers and I feel that while you didn't completely misrepresent their writings, particularly after reading Torres, I do feel like this does discard the parts of their work that are valuable. Other parts are not so valuable. Reading Gebru, I am astounded at how she behaves online and seems to promote denialism, yet simultaneously, I felt that she and Torres spend a great deal of time accurately describing the TESCREAL ideologies and correctly label them as a pro-extinction movement based on radical eugenics. Torres also likes to pair his writings with climate change (which seems like a shift in each article, but I respect it) He has also affirmatively stated that if someone is attempting to wipe out the human species he would oppose it. In light of his other writings, one could see the article you cited as a kind of provocation to get people to take this more seriously since anyone reading these arguments will have a fairly predictable emotional reaction of violent rejection.
So my feeling about these two is they are correct to name the radical eliminationism of TESCREAL ideologies that have captured Silicon Valley and a section of the billionaire class. As a software engineer, I am sympathetic to Gebru's idea that we cannot safety test ill defined models, though the cat is firmly out of the bag at this point. I am critical of Gebru in particular for selling denialism of the variety you are calling out here.
I am relatively new to this literature though, and I'm curious as to what you think comrade. I used to believe in at least some elements of the TESCREAL bundle before I rejected them and later became a socialist almost a decade ago, so when I read them, I had an immediate recognition that they were describing a real phenomenon.
I think it is almost without question true that while capability is an open maybe doubtful question, the objective of the tech CEOs like Page, Musk, Altman, Amodei, etc are the creation of a successor species. With the integration of longtermism and ordinary capitalist short-sightedness, they can excuse literally anything in their race to replace human labor and install an transhuman aristocracy. Regardless of whether its ultimate ends are practical, this is an extraordinarily dangerous ideology as there are many intermediate stages with horrible effects.
Gebru, Timnit, and Émile P. Torres. “The TESCREAL Bundle: Eugenics and the Promise of Utopia through Artificial General Intelligence.” First Monday, Apr. 2024. firstmonday.org, https://doi.org/10.5210/fm.v29i4.13636.
https://www.realtimetechpocalypse.com/p/a-pro-human-manifesto-part-1
On the one hand, Timnit and Torres correctly identify the enemy: the AI companies. The problem is that they actively advocate against any policy that would slow down or pause the development of greater AI capabilities. See this DAIR letter from 2023 Timnit recently referenced, in which they actively advocated against a 6 month pause: https://dair-institute.org/blog/letter-statement-March2023/. They're too committed to denialism to even entertain the possibility that there is *some* chance the AI industry could succeed in their undemocratic ends, which stops them from actually wanting to regulate the development of frontier AI at all. While I agree with a lot of their critiques of the current harms of AI development, their advocacy against trying to slow the pace of AI capabilities growth is quite concerning to me.
This makes a lot of sense to me. I agree that their certainly about lack of capability against evidence is a massive weakness for them and bad for the movement. Thank you for your reply.