I don't really see the point of getting deep into this, the logic is very clear in the peer reviewed paper. The energy consumption to train new models is marginal extra electricity consumption and so they've cited the EPA's numbers for CO2 emissions per Watt-hour of electricity. It's a perfectly fair comparison. Yes, you can caveat all sorts of "But we could do this with green energy" but that's not what the paper was about.
>none of her work seems particularly interested in technology.
Honestly! Her PhD in computer vision seems pretty technological to me.
Fine, if the position you want to have is all universities are race ideology centres great, again, it reflects more on you than on anyone else.
The logic is clear but incorrect. It's not a correct comparison because it assumes a watt-hour used by Google is the same as an average watt-hour. They actually buy enough renewable power for all their DCs (just not always locally). However, admitting that the CO2 impact of Google training LLMs is zero would reduce the impact of her paper, so she just makes this deceptive comparison. This is one of the reasons her paper was squashed, right? It effectively pretends all the efforts made by her colleagues on greening their DCs doesn't exist. Not very nice.
In her own words it "pertains to using large scale publicly available images to gain sociological insight" and is about "visual computational sociology". The primary reported finding is that you can predict things like racial demographics based on what brands of cars are being driven around. It's consistent with a primary interest in society and not technology.
> none of her work seems particularly interested in technology
Here is a very representative quote from her PhD thesis which you clearly did not really look up
> Augmenting Existing Adaptation Algorithms with At-
tribute Loss
> We can augment any existing adaptation algorithm with our attribute based losses to perform adaptation at the attribute as well as the class level. Here, we describe
how we apply our method to [116]. To use our method with [116], we add the
domain confusion and...
It's clear you will talk any amount of dishonest nonsense to make your point
The thread here started with the claim that "she has a Masters in Electrical Engineering, a PhD in Computer Vision, and has had a fantastic career".
It's reasonable to expect given that description for her PhD to be actually in computer vision. You know, a new SOTA on some well known benchmark or something. Some new general technique useful to anyone who deploys computer vision. Even she doesn't claim that's the case, the dissertation title claims it to be visual computational sociology which would be a different field. Her findings are pure sociology, there's no outcome here which is about computer vision (beyond "hey CV seems to work").
The original point was she doesn't seem very interested in technology. I don't see how a PhD that is at least 50% sociology disproves this point, it clearly supports it. And that's before dealing with the general dishonesty that pervades woke universities; who knows if the thesis can even be taken at face value? A place like Stanford wouldn't hesitate to misrepresent things for equity purposes.
Honestly, this is just silly. I'm sure you've got some very interesting thoughts about the way that Google manages it's data centre energy demand. It's not relevant to the paper. In fact if you read the paper you'd know your whole schtick about Google is irrelevant because their models ran on TPUs whose power characteristics aren't public information and therefore aren't in the paper.
You're quite literally making nonsensical complaints.
No, it's the other way around. The complaint being made by you and Gebru is nonsensical.
The exact power characteristics of TPUs don't matter because as I keep repeating and you guys keep desperately trying to ignore, Google years ago committed to buying renewable power 1:1 for the consumption of their datacenters. It doesn't matter how that splits between CPU, TPU and GPU, it's all covered by them.
Therefore, Google's computations net out to zero carbon emissions.
Most organizations don't do this. On average, a kilowatt hour of consumption is not from renewable sources. Therefore, Gebru's use of average emissions would be valid at other companies but not at Google, which means she was lying in order to gain impact.
Honestly, this point isn't hard to understand. It's difficult to escape the feeling that if she was a white man, it would all be clearly received.
>none of her work seems particularly interested in technology.
Honestly! Her PhD in computer vision seems pretty technological to me.
Fine, if the position you want to have is all universities are race ideology centres great, again, it reflects more on you than on anyone else.