Google told its scientists to ‘strike a positive tone’ in AI research • Hacker News Board

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#221

Post by benlivengood » Wed, Dec 23, 2020, 5:41 PM UTC

izacus wrote: post #69

chaboud wrote: post #33

Earlier quoted context omitted.

When I was young and (more) idealistic, my department chair pulled me aside after I’d been dogmatic about something and said “everything is political... The sooner you realize that, the better off you’ll be.” It took 10 years for that to really sink in, but I knew that every paper has an angle well before that. There’s a position of inevitability about AI around Google, and, to an extent, I agree. Unlike tobacco, alc…

The thing is - I don't think the bias in AI is fixable. The way we fix it is to regulate (not self-regulate, Google, Apple, Facebook and other AI companies aren't trustworthy to do it right) where AI is allowed to be used and how much of a negative impact of a persons livelihood can it impact. In other words, we need to use all the power we have to avoid "The computer said NO, you're homeless/jailed/poor forever now"…

Unfortunately a legal solution is weaker because it has to enumerate the precise conditions that are unacceptable. Do you ban the logistic function? Tanh? Statistics?

Actuarial tables have been a part of decision-making for a very long time and it's not possible to form a comprehensive legal statement of "all decisions must be made in good faith with respect to inherent biases and to the most rigorous epistemological standards".

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#222

Post by ska » Wed, Dec 23, 2020, 5:47 PM UTC

WaitWaitWha wrote: post #201

There are many comments that identify the work by corporate sponsored researchers as fraud and not research. I have the unpopular opinion, and I vehemently disagree. There is no altruistic, sponsored research, no matter of the sponsor, be it corporate, government, non profit, or the Holy See. There is always a specific outcome the sponsor hopes for. The question is, how much pressure the sponsor puts on the researche…

I think there is a useful distinction to be made between the impact that financial support has just by existing, and the leverage that can give the sponsor if they choose to use it.

If I offer a 1mm grant the use of X drug on Y disease, I am being opinionated about where time should be spent and what diseases are "important". This is not neutral, and can be very political.

However, this is in a different universe than if I offer the grant with strings attached like editorial control on the paper(s), or a veto on publishing.

As you note, most failed research doesn't get published but vanishingly small amounts of that are due to outside pressure, it's mostly more mundane reasons.

Research is sponsored by, and done by people, hence it is flawed. But the best of is places value on the rigor of the research regardless of outcome, and certainly does not attempt to silence results that are not favourable to the sponsors goals. Commercially sponsored work should aspire to this in the same way that public research does - the specific outcome the sponsor hopes for is progress in our understanding. In both cases, sometimes the results are damaging to other projects you have (business goals, policy programs, etc.) which is the real test.

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#223

Post by _Nat_ » Wed, Dec 23, 2020, 5:49 PM UTC

cmiles74 wrote: post #207

_Nat_ wrote: post #193

Earlier quoted context omitted.

I'm having trouble getting a clear perspective on what's going on. So many of the descriptions are vague and based on apparently informal descriptions, leaving much to the reader's imagination. > It's not theoretical, researchers are quoted in the article and papers have been altered. In grad school, PhD students' advisors typically insist on various revisions before publishing a paper, as publications reflect on the…

When an graduate school advisor provides feedback on a paper, the goal is to improve the quality of the paper. The peer review process also has the same goal in mind: produce a better paper. According to this Reuters article, Google's new process happens after peer review and Google's other processes have completed. "The “sensitive topics” process adds a round of scrutiny to Google’s standard review of papers for pit…

> I think this is very different from the advising process in graduate school.

All of that sounds normal to me. Including filtering out trade-secrets, which is completely normal when working with trade-secrets in grad school too. Additionally, it's completely normal to filter out intellectual property you might plan to patent; confidential information; proprietary industrial information; information protected by law; dangerous findings (e.g., hackers often omit details of an exploit until the relevant vendor has had time to fix); and a few other categories.

Maintaining a positive, constructive tone is also completely normal. For example, failed experiments are typically described as progressive steps toward an ultimate success; unforeseen problems are discoveries; and major issues are seen as research challenges to be overcome. Or, ya know, stuff like that.

I mean, is that all this story's about? Because if that's it, then it seems like nothing substantial. But if that's the case, why is this in the news?

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#224

Post by shadowgovt » Wed, Dec 23, 2020, 5:50 PM UTC

WaitWaitWha wrote: post #201

There are many comments that identify the work by corporate sponsored researchers as fraud and not research. I have the unpopular opinion, and I vehemently disagree. There is no altruistic, sponsored research, no matter of the sponsor, be it corporate, government, non profit, or the Holy See. There is always a specific outcome the sponsor hopes for. The question is, how much pressure the sponsor puts on the researche…

I've definitely seen this in academia. Some institutions will bias in favor of supporting high-visibility research over "meat and potatoes" base-line research (such as reproduction of other experimenters' results, or chasing down un-sexy problems of elaboration or exploration of an information space). Their incentives are to chase grant money and get published, even if the published results aren't thorough or rigorous but will capture column inches (leaving someone else to do the un-sexy work of verifying the claims from a single experiment run).

It leaves gaps in the robustness of human knowledge, a bit of a Swiss-cheese at the edges that isn't necessary, but for the way we reward people for cutting the edge instead of filling the back.

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#225

Post by throwaway2048 » Wed, Dec 23, 2020, 5:50 PM UTC

Laarlf wrote: post #55

nisuni wrote: post #13

And that’s why I was happier when basic research was mostly being carried out in public institutions.

Oh yeah. Those are surely independent and don’t have aims at all.

This sounds like the typical false equivocation about "all media being lies" when somebody brings up the danger of outright, constant propaganda like OANN.

Just because bias is impossible to avoid, does not mean that all bias is the same.

The goal isn't outright elimination of bias (which is completely impossible), its ensuring incentives that are aligned so bias doesn't become a self-reinforcing feedback loop.

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#226

Post by jldugger » Wed, Dec 23, 2020, 5:51 PM UTC

titzer wrote: post #204

I published several papers while working at Google on Chrome. There is a publication review process that at that time was not onerous at all. In fact, all of my publication submissions were reviewed post facto and basically rubber-stamped. All, with the exception of one, our publication on Spectre, which was carefully vetted to not disclose any sensitive information about further vulnerabilities or violate NDAs with…

> I published several papers while working at Google on Chrome.

I'm curious -- did any of your papers touch on the ethics of browsers or call on specific caucuses of Congress to take action that your employer will not?

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#227

Post by shadowgovt » Wed, Dec 23, 2020, 5:51 PM UTC

ska wrote: post #222

WaitWaitWha wrote: post #201

There are many comments that identify the work by corporate sponsored researchers as fraud and not research. I have the unpopular opinion, and I vehemently disagree. There is no altruistic, sponsored research, no matter of the sponsor, be it corporate, government, non profit, or the Holy See. There is always a specific outcome the sponsor hopes for. The question is, how much pressure the sponsor puts on the researche…

I think there is a useful distinction to be made between the impact that financial support has just by existing, and the leverage that can give the sponsor if they choose to use it. If I offer a 1mm grant the use of X drug on Y disease, I am being opinionated about where time should be spent and what diseases are "important". This is not neutral, and can be very political. However, this is in a different universe tha…

Offering a grant on Y disease at all also biases research in favor of pursuing that over other diseases, possibly diseases of more pressing urgency to cure in the realm of improving human quality of life.

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#228

Post by mlthoughts2018 » Wed, Dec 23, 2020, 5:53 PM UTC

joshuamorton wrote: post #213

mlthoughts2018 wrote: post #187

Earlier quoted context omitted.

You appear to be mistaken on nearly all the facts about the Gebru situation. Given this it seems clear we would draw very different conclusions. Anyone claiming Jeff Dean lied must be so disconnected from the actual facts and so prejudiced to blame Google and view Gebru’s actions as acceptable that their opinion amounts to nothing but partisan gainsaying. To be clear, it’s ok to disagree with Jeff Dean or to dispute…

It really isn't. He lied. Normalize calling out corporate executives when they lie. I wish he hadn't. I'm disappointed that he did. But I don't know how else to describe sharing statements one knows (or should know) to be false.

No, he really, really didn’t lie. I wish people predisposed to only see Google as being on the wrong side of it would stop saying stuff like this.

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#229

Post by jonahrd » Wed, Dec 23, 2020, 5:53 PM UTC

WaitWaitWha wrote: post #201

There are many comments that identify the work by corporate sponsored researchers as fraud and not research. I have the unpopular opinion, and I vehemently disagree. There is no altruistic, sponsored research, no matter of the sponsor, be it corporate, government, non profit, or the Holy See. There is always a specific outcome the sponsor hopes for. The question is, how much pressure the sponsor puts on the researche…

> The only time I have seen altruistic, "clean" research was performed by individuals as hobby. Once there are more than one person involved, there will be some preferred outcome. Nature of humanity.

I also think it's worth noting that even though "clean" research does exist coming from individuals, there's also a significant chunk of individuals whose research is even _more_ predetermined for a specific outcome.

Re: Google told its scientists to ‘strike a positive tone’ in AI research

#230

Post by shadowgovt » Wed, Dec 23, 2020, 5:54 PM UTC

googlethrow3134 wrote: post #206

titzer wrote: post #204

I published several papers while working at Google on Chrome. There is a publication review process that at that time was not onerous at all. In fact, all of my publication submissions were reviewed post facto and basically rubber-stamped. All, with the exception of one, our publication on Spectre, which was carefully vetted to not disclose any sensitive information about further vulnerabilities or violate NDAs with…

When have trillion dollar companies not been authoritarian in protecting and furthering their own corporate interests?

All too true. The biggest challenge Google is struggling through right now is that it's not just impractical, but dangerous for it to continue to behave like a small and scrappy startup.

Behavior patterns that are beneficial for a small company are harmful to a big company. To give a concrete example, "move fast and break things" is fine when you have 100 people relying on you, and should probably be sanctionable (though, in our industry, it is often not outside the context of contracts and pre-arranged agreements) when you have 2 billion people using your system several dozen times a day for critical information, where their lives are materially negatively impacted if that information is false or out-dated.