About Google's approach to research publication – Jeff Dean • Hacker News Board

Re: About Google's approach to research publication – Jeff Dean

#471

Post by ergocoder » Sat, Dec 05, 2020, 1:55 AM UTC

dragonwriter wrote: post #254

wpietri wrote: post #248

Earlier quoted context omitted.

No, she did not want to part ways. If she did, she would have quit. Instead, she said she needed certain things to keep working there. That pretty obviously indicates a desire to keep working.

> Instead, she said she needed certain things to keep working there. And, from what I understand, that she wanted to discuss the issue in person when she returned from vacation (which she was on at the time.) EDIT: I point this out because I think that this potentially recasts the whole communication from an non-negotiable ultimatum to something more like a fair warning to avoid blindsiding anyone going into an in-pe…

> And, from what I understand, that she wanted to discuss the issue in person when she returned from vacation (which she was on at the time.)

Then, she should have said that.

It would be ridiculous for Google to let her hang out at the company.

"Oh we cannot meet your meet ultimatum for sure. But please feel free to hang around and finish your vacation. We'll wait for you, a disgruntled employee, to choose when to leave."

Re: About Google's approach to research publication – Jeff Dean

#472

Post by jariel » Sat, Dec 05, 2020, 1:57 AM UTC

jolux wrote: post #452

kps wrote: post #367

Earlier quoted context omitted.

> at the last second It was only ‘at the last second’ because Gebru chose not to follow the normal procedure. If the paper genuinely can't be ready until one day before the external deadline, the right thing to do is engage with the reviewers in advance, explain the problem, and provide them with drafts and work in progress, so that they can complete their work a few hours after yours. What Gebru did is the equivalen…

According to Jeff Dean's account of events, yes. According to nearly everyone else's, this process was unusual in that it even involved reviewing for content and not just IP, and Gebru herself says that the website about the process says to submit at least 1 week before publication, not two.

"According to nearly everyone else's, this process was unusual in that it even involved reviewing for content and not just IP"

This is ethics for Google's AI+Search which is currently undergoing global scrutiny, particularly by Congress and specific politicians who are considering anti-trust measures against Google - and who believe that 'their political party is being treated unfairly'.

It's existential concern for them right now, relating to the possible breakup of the company.

Every public communication on 'ethics' or search results etc. at Google is obviously going to have to be reviewed.

If you're publishing the latest thing on 'AI Random Number Generation' obviously nobody cares about anything other than IP.

The fact is, she must have known this and submitted anyhow - which is in and of itself not so bad, but that there was calamity afterwards ... there is no excuse.

Google was absolutely reasonable - they did not ask to change the nature of the research, but wanted to make sure that information about new, better processes were included.

It's beyond gracious for Google to do this, when really their starting point is 'silence' and they really don't have to do anything at all.

A request for a fairly short review with very basic and reasonable concerns blew up.

This is not a public university, you don't get perfectly tenured academic freedom, if Google wants to put a reasonable subnote in there - and take 2 weeks to do it, it's perfectly fine.

Obviously Google would have kept her if they wanted to, but it's clear they were both looking for a way to part ways and it's probably for the better.

Re: About Google's approach to research publication – Jeff Dean

#473

Post by ergocoder » Sat, Dec 05, 2020, 2:00 AM UTC

ryanobjc wrote: post #316

SpicyLemonZest wrote: post #102

Earlier quoted context omitted.

We don't know the exact language, but by Timnit's account she made a numbered list of demands, which is pretty inherently direct.

And she said her email contained language like “and if my demands aren’t met, then I will find an exit date.” It sounded from her account a straightforward threat to leave.

"We can't meet your demands for sure. But please hang around the company and choose your exit date at your liking" -- no reasonable person would respond like this.

In any case, making this an issue sounds way overkill.

Both want to part way. Either side can choose an earlier date of the termination.

She doesn't need Google's permission to leave earlier. Google also doesn't need her permission to make her leave earlier.

Re: About Google's approach to research publication – Jeff Dean

#474

Post by jariel » Sat, Dec 05, 2020, 2:03 AM UTC

cmsj wrote: post #321

nickff wrote: post #224

Earlier quoted context omitted.

As a manager, if someone gives an ultimatum, you basically have to fire them. There's no real option; Ben Horowitz covered it somewhere, but the bottom line is that if you yield, you've given up all control.

I understand some of the reasons why some managers think that way, but you can't have such simplistic rules. An ultimatum like this is an opportunity for a responsible manager to talk and rethink, but it seems like Google jumped at the opportunity to double-down on their mistake and then send out cowardly emails claiming the employee had actually resigned. If I were to apply a simplistic rule here, I would actually i…

"an opportunity for a responsible manager to talk and rethink"

Mostly it's an opportunity to let the staffer know that such ultimates are unacceptable, and that taken literally by her own terms - she could be called out and let go. Which is what happened.

It's very doubtful that if they wanted to keep her, that they couldn't have found terms.

Surely the manager would have bent, indicated the wording was a little bit strong, and found a way forward.

It seems clear they were wavering, she crossed a line and offered them the path out and they took it.

If there were material issues being covered up, there was material suppression of information, this story would look completely different - but there wasn't.

This was the right thing to do by Google in a tricky situation.

Re: About Google's approach to research publication – Jeff Dean

#475

Post by joshuamorton » Sat, Dec 05, 2020, 2:03 AM UTC

xiphias2 wrote: post #470

joshuamorton wrote: post #397

Earlier quoted context omitted.

> A manager in ethics shouldn’t ask Google to break the law by not providing confidientality that was requested. It is not implied anywhere that the reviewers requested confidentiality. It is certainly not implied that Google would be violating the law to rescind that confidentiality (and, to be clear: they almost certainly would not be).

You are right, it's not in the article. It was published by Timnit: https://www.platformer.news/p/the-withering-email-that-got-a... But you are not right about privileged and confidential : it limits what Google can do legally. ,,Privileged and confidential communication is the interaction between two parties having a legally protected, private relationship.''

Timnit was presented with a privileged and confidential document. That binds her from sharing the information in the document with others. It has nothing to do with the authors of the document or the confidentiality of their identities.

And importantly, in the context of a company like Google, an executive (like jeff Dean) likely has the authority to "unprivilge" said contents, whatever they are.

Re: About Google's approach to research publication – Jeff Dean

#476

Post by xiphias2 » Sat, Dec 05, 2020, 2:06 AM UTC

joshuamorton wrote: post #475

xiphias2 wrote: post #470

Earlier quoted context omitted.

You are right, it's not in the article. It was published by Timnit: https://www.platformer.news/p/the-withering-email-that-got-a... But you are not right about privileged and confidential : it limits what Google can do legally. ,,Privileged and confidential communication is the interaction between two parties having a legally protected, private relationship.''

Timnit was presented with a privileged and confidential document. That binds her from sharing the information in the document with others. It has nothing to do with the authors of the document or the confidentiality of their identities. And importantly, in the context of a company like Google, an executive (like jeff Dean) likely has the authority to "unprivilge" said contents, whatever they are.

,,Have you ever heard of someone getting “feedback” on a paper through a privileged and confidential document to HR? ''

To me it sounds like the person who submitted the feedback to HR was using privileged and confidential communication.

And yes, Jeff Dean has the authority, but the HR who got the feedback doesn't.

Re: About Google's approach to research publication – Jeff Dean

#477

Post by tylerl » Sat, Dec 05, 2020, 2:10 AM UTC

theobeers wrote: post #33

gcr wrote: post #2

Maybe different teams are different, but on my previous team within Google AI, we thought the goal of google's pubapproval process was to ensure that internal company IP (eg. details about datasets, details about google compute infra) does not leak to the public, and maybe to shield Google from liability. Nothing more. In all of my time at Google AI, I never heard of pubapproval being used for peer review or to criti…

And the examples of issues flagged in review that Jeff keeps highlighting—like Timnit’s alleged failure to mention recent work to reduce the environmental impact of large models—are themselves a bit worrisome. Jeff gives the impression that they demanded retraction (!) because they wanted Timnit and her coauthors to soften their critique. The more I read about this, the worse it looks.

The communication doesn't give that impression; instead it says that the paper makes claims that ignore significant and credible challenges to those claims. Dean said that these factors would need to be addressed, not agreed with.

Publishing a transparently one-sided paper in Google's name would be a problem, not because of the side it picks, but because it suggests the researchers are too ideologically motivated to see the see the problem clearly.

Ironically, it indicates systemic bias on the part of the researchers who are explicitly trying to eliminate systemic bias. That's just a bit too relevant to ignore.

Re: About Google's approach to research publication – Jeff Dean

#478

Post by sidlls » Sat, Dec 05, 2020, 2:11 AM UTC

fatbird wrote: post #362

lostmsu wrote: post #162

Earlier quoted context omitted.

This is incorrect, she did receive the feedback, then demanded the source to be deanonimized.

A week before you go out on vacation, you see a meeting pop up at 4:30pm PST on your calendar (this popped up at around 2pm). No one would tell you what the meeting was about in advance. Then in that meeting your manager’s manager tells you “it has been decided” that you need to retract this paper by next week... And you are told after a while, that your manager can read you a privileged and confidential document and…

Sure that’s a fair point but asking for the reviewers to be identified is way out of bounds. What possible legitimate reason could she have for that request?

Re: About Google's approach to research publication – Jeff Dean

#479

Post by dodobirdlord » Sat, Dec 05, 2020, 2:15 AM UTC

aphyr wrote: post #111

Dean writes: Our aim is to rival peer-reviewed journals in terms of the rigor and thoughtfulness in how we review research before publication. But Gebru writes that HR and her management chain delivered her feedback in a surprise meeting where she was not allowed to read the actual feedback, understand the process which generated it, or engage in a dialogue about it: Have you ever heard of someone getting “feedback”…

It does feel off, and it’s easy to read as Google trying to suppress her findings. But there’s a simple charitable reading as well. Gebru has recently and very publicly developed a reputation for behaving with hostility toward colleagues. Google is a company that prizes (or at least claims to prize) psychological safety of employees. I can see all of this being the chain reaction caused by a number of her coworkers expressing that they were unwilling to give an honest opinion about her work if she would be able to tie it back to them. If an ordinary employee caused their coworkers so much concern they would probably already have been dismissed, but Gebru is especially talented and was high-ranking.

I imagine some reviewers extracted agreement from management before giving negative feedback on this paper that the “anonymous” in “anonymous feedback” was a promise. This explains the unusualness of the situation, why the feedback flowed though a special HR channel, why specifically management was unwilling to let her have a written copy of the feedback (that could be closely analyzed to de-blind the reviewers), and why management accepted her resignation and the resulting fallout rather than agree to de-blind the reviewers.

Re: About Google's approach to research publication – Jeff Dean

#480

Post by joshuamorton » Sat, Dec 05, 2020, 2:16 AM UTC

xiphias2 wrote: post #476

joshuamorton wrote: post #475

Earlier quoted context omitted.

Timnit was presented with a privileged and confidential document. That binds her from sharing the information in the document with others. It has nothing to do with the authors of the document or the confidentiality of their identities. And importantly, in the context of a company like Google, an executive (like jeff Dean) likely has the authority to "unprivilge" said contents, whatever they are.

,,Have you ever heard of someone getting “feedback” on a paper through a privileged and confidential document to HR? '' To me it sounds like the person who submitted the feedback to HR was using privileged and confidential communication. And yes, Jeff Dean has the authority, but the HR who got the feedback doesn't.

Given that Timnit ostensibly does not know who the people who gave the feedback, nor how they did so, and elsewhere notes that she was able to see the feedback, but was told she couldn't share it because it was privileged, can you explain how you reach this conclusion instead of that "to" was a typo in a tweet and instead should have been "via" or similar.