Dec. 5th, 2020

davidgillon: Text: You can take a heroic last stand against the forces of darkness. Or you can not die. It's entirely up to you" (Heroic Last Stand)

I've had several dreams in the last week, including at least one micro-dream in a micro-sleep, where I've been doing things like get on the train to head up to Durham for Christmas.

Sorry, dream-brain, but that ain't happening this year, get with the programme.

davidgillon: Text: You can take a heroic last stand against the forces of darkness. Or you can not die. It's entirely up to you" (Heroic Last Stand)

If you haven't been following the tech news, the big story this week is Google sacking Timnit Gebru, their star researcher on the ethics of AI, because she refused to take her name off a new paper (which apparently had six authors, four of them from Google). Google say she resigned, what actually happened was she said it was something she might have to resign over and they needed to talk when she got back from leave, but her manager, the head of Google's AI research team, promptly cut off her email and claimed she had resigned.

There's an analysis of the paper (not yet generally released) in MIT Technology Review https://www.technologyreview.com/2020/12/04/1013294/google-ai-ethics-research-paper-forced-out-timnit-gebru/ which is completely readable, no AI knowledge needed. And it turns out the paper looks into the ethics of AIs trained on huge amounts of written text ('large language models'), which Google is increasingly dependent on in its search business, and points out two major and several minor ethical problems, and where research might look to mitigate that.

The simplest is probably the sheer amount of energy these models consume, with a single run producing an amount of CO2 equivalent to the lifetime costs of five US cars. It's this Google has chosen to hang their public concern about, stating that the paper doesn't have enough references to papers discussing mitigation strategies already being developed. In fact it does address them and makes suggestions for further mitigation strategies. On top of which the paper has 128 references, apparently very large for the field, and multiple Google employees/ex-employees have said Google have never policed the amount of references in papers, simply reviewed them for propriety information.

The big one, and probably the real reason for Google's managers trying to force her out, which really worked well for keeping the story under wraps, is that Gebru and the other researchers looked at the ethics of teaching AIs from massive amounts of text, when written language is full of hidden biases. The examples people have been drawing has been how language looks at race, but I think an even better example might be how language looks at disability, with so many of our common negative terms having their origins in ableism. To an AI, it's easy to recognise from mass text that, for instance, the R-word is considered an unacceptable negative, and the R-word originated, and is commonly used, to describe disabled people. And the incorrect conclusion that disabled people are unacceptable flows out of that at least as easily as the correct assumption that there is a problem with the language we use to talk about disability.

There's also a related point that using massive amounts of text is subject to the inherent bias of the dominance of the English language in written text, and particularly the First World version of that, which may mean local, non-First World and non-Anglophone nuance is simply swamped.

The other problems raised are that large language models don't understand text, but are good at manipulating it, which means there's a disincentive to study AI that actually does understand it, and that AIs make good mimics, but mimics without any understanding, such as the AI which mistranslated a Palestinian man's "good morning" in Arabic into "attack them" in Hebrew and got him arrested.

So that's a major problem with the whole concept of large language model AI, and Google's search business, the company's cash cow, is increasingly dependent on precisely that kind of AI (Google search revenues were $23.6Bn in Q3 2020). No wonder Google management panicked.

Profile

davidgillon: A pair of crutches, hanging from coat hooks, reflected in a mirror (Default)
David Gillon

March 2025

S M T W T F S
      1
2345678
9101112131415
1617 18192021 22
2324 2526272829
3031     

Most Popular Tags

Style Credit

Expand Cut Tags

No cut tags
Page generated Jun. 9th, 2025 11:23 am
Powered by Dreamwidth Studios