Whether or not I agree with the conclusions, that’s exactly the sort of paper one might expect to be written by someone whose job title includes the words “ethics” and “AI”. I’m not sure what Google expected.
The whole paper doesn't go beyond surface-level analysis. The main point is that language models reflect the same biases as its training data. Very little of it is technical. Most of it is about examples of how language models aren't as politically correct as the authors want it to be (e.g. calling female doctors "female doctors" when it's not relevant to the context). It's a legitimate concern, but anyone who remembered Microsoft Tay would know that.
the paper was not the problem, her attitude & reaction to peer review was the problem - giving ultimatums, demanding to disclose names of the reviewers, and broadcasting emails to entire team calling for sabotage.
Her behavior fits the definition of 'toxic employee', it was unprofessional by any standard.
The other curious thing is, behind the pink-boxed redactions are the names of four other Google employees. I'm wondering why they haven't been reprimanded in the same way. At least, not that we know of.
Well, she wasn't reprimanded for writing the paper nor for publishing it. In fact, she wasn't even reprimanded as far as I know for telling other employees to stop working on DEI. She told them she wanted confidental information or she will quit on a date she would like. They said they would not give her that information and that she is no longer welcome at the company. This isn't really reprimanding or even a normal firing. If I tell my boss do X or I quit and they say "Thanks for your work, we accept your resigination and the legal notice period applies." Fair play.
> As shown in the amount of compute used to train deep learning models has increased
300,000x in 6 years, increasing at a far higher pace than Moore’s Law which posits that the amount of computation that can be done per unit area would roughly double every two years. This means that power consumption per unit area is not staying constant as implied by Moore’s Law [108].
IIRC Moore's law doesn't make any predictions about power consumption at all, let alone ones meriting being cited in a scientific paper.
yeah, I've debunked this particular line as being technically nonsense. If I were a reviewer of the paper, reaching a sentence like that, I'd return to the editor with "revise and resubmit- this doesn't meet the basic technical level of expertise we require for publication."
if you build 2X the compute capacity with current nodes, you didn't violate moore's law.
She wasn't fired over the paper. She was fired because she demanded Google name the employees who gave (anonymous) negative feedback about the paper, and that anyone going forward who gave feedback on her work also be named, or she'd quit.