Hacker News Reader

Comment by tonymet | Hacker News Reader

Having worked inside a few big tech companies, this outcome was practically inevitable. It would have been impossible for Googlers to raise the concerns that the public did without serious reputational or occupational hazard.

The visible & spoken culture at Google, Meta, and most big tech companies is very much in favor of "anti-racism" and "ML-fairness", though a large minority harbors criticisms in private conversation.

Any criticism of content or bias made during the bard/Gemini dogfooding phase would have been met swiftly with ostracization and possible termination.

View on HN

If the results are as silly as the article suggests, I wonder if it could be deliberate sabotage. They know they can’t raise the issue normally, so go the other way, turn the dial to 11. Now it’s just ridiculous and is all over the media. The CEO himself is asking the dials to be turned back to 3.

it's about the only corrective measure that would have worked.

Why are you assuming that when the extreme leftist are literally in charge - have entire departments- have made it so that literally all managers must pretend they also believe in the same things.

The models are working as intended

When you are required to listen and obey and not to think or question the difference between loyal obedience and sabotage becomes entirely theoretical.

And yet this is exactly what management demands. I guess it's one aspect of the self-sabotage.

> dials to be turned back to 3

Let’s not kid ourselves, it’ll be dialled back to 9. This nonsense is evident in search and image search, and has been for years.

It’s unlikely to be sabotage though. What may have happened is that the teams working on RLHF may have only trained it on a dataset containing the strictest interpretation of liberal values without also training it on other data that grounds it in reality. As a result, the model tried to inject the liberal interpretation of values where it doesn’t make sense, such as Black Nazis, and so on.

It's difficult though.

A lot of societies up through the ages depicted themselves unrealistically, and our depiction of past societies can be rather questionable too. Cast a glance at https://www.moviestillsdb.com/movies/ben-hur-i52618 if you want.

Those pictures of nazis were clearly wrong, but describing something as wrong is easier than saying what's right. Should an AI match a society's prejudiced descriptions of itself (making the nazis blond), our later images, often also prejudiced, or the current best (yet fuzzy) historical notions of what reality was like?

> describing something as wrong is easier than saying what's right

This 1000x. This problem you also see with education of children. You have 2 problems:

1) there's a VERY large number of possible depictions. Let's say 10 million. Of those, a VERY small number are correct/acceptable. Let's say 10.

Then it's easy to see. A correct example has 1/10 = "10%" of all information so to speak. If 100% is what you need to be guaranteed a correct classification of all possible depictions.

Yet a negative example has 1/10,000,000= "0.000001%" of the information you need to classify all examples correctly. If you attempt to create correct behavior by giving wrong examples ... it'll be a long day.

A positive example, telling an AI (or a child/student) what to DO has ~10000 times more information than telling them what NOT to do, in this example. In practice, you'll find the difference is even more extreme.

2) BUT there is a problem, that's also always brought up with wikipedia. Giving positive examples requires that there are examples that everyone agrees with. And we all know there are rather serious problems with that. A positive example HAS to be in the intersection of what all groups find acceptable. And ... well easy examples are always wars: Who is right? Who started the war? Is it justified? Applied to Russia, Ukraine, Israel, India, Sudan, ... but in practice a lot of issues, some not at all that controversial (which word is correct "New York" english? Bupkis or bupkez?)

Replying to myself to say it better.

What should Gemini have done, what would be right? Should it match the historical record, which the Nazis twisted by choosing to photograph mostly tall blond soldiers that looked a bit Danish, or should it match fact, which then would look different from the historical record, and is difficult to establish accurately anyway?

Similar questions apply for other societies, both current and ancient. Current e.g. Russia, where the census asks people to select ethnicity, but any choice except "Russian" might as well be labelled "please discriminate against me". Whatever an AI does is going to be wrong, and calling anything right is extremely difficult. Ancient examples include India and Egypt, which both had long-lived dynasties that considered themselves somehow foreign to that land and were very selective about how their represented their realms.

It's really easy to find a horrible example and call it horrible, but it's low-effort thinking. Describing something that's right is so much harder and so much more valuable.

> Current e.g. Russia, where the census asks people to select ethnicity, but any choice except "Russian" might as well be labelled "please discriminate against me".

I’m going to be—well, not the, but among the—last to defend the current Russian political system, but in which sense is this true? There’s a hell of a lot of discrimination in Russia, but the targeting for it is usually along the lines of not having a Russian passport or a registered address of residence in a particular region, or having the “wrong” surname or appearance. I’m not aware of anybody ever caring what you wrote in the census papers—and the data from the latest census is so bogus basically everybody who does population statistics has disregarded it—but maybe it happens in smaller towns? That would be interesting.

The only place I’m aware of that actually associates your declared (not implied) ethnicity with your name is the ostensibly statistical slip you fill in when changing your registered address, but at least fifteen years ago quoting the relevant parts of the constitution resulted in surprise followed by begrudging acceptance. (Art. 26: Everyone can determine and specify their ethnicity. Noone shall be forced to determine or specify their ethnicity.—a direct attempt to preclude the “item 5” shenanigans common in Soviet times.)

Again, none of this is to say that there isn’t a lot of discrimination happening. (As one example, the disproportionately large military draft in some areas could arguably amount to ethnic cleansing, were it perpetrated with intent instead of arising through general administrative laziness seeking a softer target. Either way, some groups and possibly even languages are not going to survive.) I’m just very surprised to hear about the census, specifically, being important in that respect.

My source is gone from the web. It was by someone who grew up in a provincial city. Summarising from memory: Anyone who had a choice (e.g. one parent from Moscow and one from the frowned-upon region) would do well to ① pick a muscovite name/spelling and ② tick "russian" on an unnamed form he filled in when he became a student. And this situation is/was common in that kind of town.

I assumed it was a census form because of Wikipedia: "In the 2021 census, roughly 81% of the population were ethnic Russians, …" which seems to say that the census forms ask about ethnicity, doesn't it?

The census form[1] does ask about ethnicity, sure (item 14; while you’re free to leave the field blank or write “Martian” there if you wish, it’s true that most people won’t). It doesn’t ask for your name, however (and isn’t received or sent by post, either—census takers are supposed to go door to door). Not that it’d be difficult to recover if you intercepted the forms before they were collated (despite assurances of confidentiality printed on them), I’ve just never heard of anyone trying that or even being vaguely interested in it.

The story you’re telling sounds like something out of Kanevsky&Senderov[2]. I’m not denying it might be happening, but I’ve never heard of it and would be interested to hear more—at least a location and a year would be helpful. I’ve never filled in, seen, or heard of such a form in relation to higher education, but then I only have personal experience with Moscow and second-hand one with St Petersburg and Ekaterinburg, all large cities without much of a sharp divide between common nationalities of residents (as in Bashkortostan or Chechnya or a number of other places).

(To call Moscow ethnically homogeneous would be a huge stretch, mind you. I can name [former?] Moscow residents in my contact list with Armenian, Azerbaijani, Chechen, Jewish, Korean, Mordovian, Roma, and Ukrainian ancestry without even opening it—hell, I could nominate myself for a third of those slots. I can’t even imagine what I’d learn if I actually went around asking my acquaintances about their family history. I suppose most of the results would still count as “white” by US reckoning, but, well, meh to that.

Culturally, though, yeah, things are pretty uniform. I just wanted to warn you away from thinking that that 80% figure reflects the country’s people mostly coming from a single ancestral group as opposed to them being subjected to pressure to have “Russian” written in their ID for like half a century. There’s a reason why optimistic discussions of Russia’s medium-term future usually include the question of the number of states that would exist in its place.)

[1] See link to PDF file at http://government.ru/docs/38324/ for the version used in the botched 2021 census (labelled 2020, this being a document from back before Covid).

[2] https://www.worldscientific.com/doi/10.1142/9789812701169_00..., free at https://www-users.cse.umn.edu/~shifman/EinsteinBook.pdf

I looked at the map then and against tonight, can't find a city whose name rings a bell. Russia is large enough to overlook things ;) It was some way east of Moscow, a little north of due east perhaps, still west of the Urals, and from context it would be an industrial city in an industrial area. Sorry, but I just can't remember the name.

Are you suggesting that 81% is even close to reality? Almost 200 ethnic groups and one of them forms 81% sounded like cooked bookkeeping to me.

Or maybe if they hadn't fired a bunch of people whose job it was to say "I think you're over-correcting here" and with strong arguments and supporting data who cannot be so easily dismissed as cranks from the minority because it's an actual job function. This reads like the people blaming DNI for declines in software quality after a decade if eliminating quality assurance as a normal and expected participant in development. WTF? Eyes on the prize here. This is about getting it done how we expect it to be done, what ever that may be. Why are we shipping bugs? Because we don't value the function of identifying them and determining the scope of their impact and the relative urgency of addressing. We don't empower anyone to who knows what's going on to stop ship any more. What happened to that? Agree on the target and then hit it. Or don't agree on the target, but hit it none the less. That means understanding your stuff before you ship it and you pay people to do that for you. Boeing cuts the same corners. It's just plain negligence and it's completely unsustainable. Stop letting fear of falling behind lead to failed releases and empower the proper teams to evaluate so you know what's going out the door before your customers do.

Quality was not degraded across the board : the software was still performant, responsive , reliable, dynamic, useful etc.

the quality degraded along a very specific vector having to do with "ML fairness" and "anti racism".

You could have hired 50000 QA engineers and none of them would have been comfortable making comments in this area.

From my 8 years at Google:

s/large minority/supermajority/g

Yes. In other words, short sighted greed but really very simply: cowardice.

On the other side of the coin: it is often said that the difference between bravery and stupidity is whether you're successful.

There are 150,000+ employees at Google.

You must be pretty popular to be able to make such claims about what most of them believe in.

Dmitry is right. The vast majority of Googlers lean somewhere between against and apathetic to the DEI nonsense that's infected the company, but no one's willing to be the nail that sticks out and complain about it visibly.

Two anecdotes:

1. memegen had downvotes, and well-known instigators would continuously complain on plus about their shitty polarizing memes getting enormous numbers of downvotes, and get nothing but hugbox replies in the comments. The creator of memegen who was ideologically aligned ended up regretting adding downvotes at all, under the theory that people were conspiring to downvote instead of there being a silent majority of people that disagreed with the memes in question: https://www.mcmillen.dev/blog/20210721-downvotes-considered-...

2. Someone made an anonymous survey about whether people thought the Damore memo stated facts, had sound logic, was a good idea to post, deserved firing, etc. It was taken down shortly after because people didn't like the results and the implication that the majority of people that filled it out didn't think that the post was worthy of being immediately fired, etc.

More cynical me sees them as typical self-professed champions of tolerance that are quite selective of allowing diverging opinions. These people are interested in careers, not in technology or betterment in any form.

They believe themselves to know better than universal humanism. The reality is that they just want to have something that justifies their position, because it isn't their knowledge or ability. They will want to keep racism alive too because it fits their personal needs.

You're right. There's just a very loud minority of woke fools who are actively tanking a lot of the best companies. And a good chunk of them are not actually looking for solutions, they just like that they have a problem which fits their 'expertise'.

The "anti-racist" manifesto is likely the reason for a lot of this garbage where it's not just enough to report on a racist world, but you also have to modify things to fit a not-racist world.

That's what gave us a black Cleopatra irl and this is what's giving us AI imaggen of diverse Nazis in WW2.

I think Google fired the biggest "anti-racism" and "ML-fairness" proponent?

They fired the single one that was way too outspoken- Also she literally went against explicit rules about publishing I think

As I understand it, Timnit Gebru was fired for suggesting that the power usage of training ML/AI was significant.

Her other 'woke' views didn't contribute as much to her firing. (Though her firing did set off a firestorm about those issues)

I think from your tone, I disagree with you about her other views.

But it seems her firing was more about making Google look anti-environmental than anti-woke.

MITs summary of the paper: https://www.technologyreview.com/2020/12/04/1013294/google-a...

Only one section deals with already well known problem of bias.

I don't know of true reasons inside Google but the quality of the work alone would've justified termination for a high-ranking position.

If you read the famous "Stochastic Parrots" it's not a scientific work at all, it's a piece of journalism that just throws together as many unrelated AI-scares as the author could find. A good article for Vogue or Medium but unworthy of someone who claims to be a scientist.

She was mainly fired for not withdrawing a paper when it was requested by a leader in research (the leader who asked for this had no research background and asked for a retraction, which doesn't make sense because the paper hadn't been published yet). But there were other important factors; for example, she attacked other googlers (prominent ones working on LLMs) on internally public mailing lists (about woke things). Realistically, the world would have been better off if she hadn't been hired in the first place because she was always going to eventually have a conflict with leadership over publishing works like this. I think she would have done much better to become a professor at some liberal university where she would be free to publish her work.

I kind of wish the stochastic parrots paper had focused entirely on stochastic parrots, and not on energy consumption. In my opinion, Google has actually been a responsible steward of energy usage (way ahead of everybody else for at least a decade), and ML isn't really the source of most of the energy consumption in computing anyway.

She tweeted things like kill all white men. Its was nuts

Was that the situation where the tweet was from over a decade before when she was a teenager, and she said she learned to be more thoughtful as an adult?

That could be believable if that kind of lenience would be applied to other opinions aside for very selective ones. That would be real tolerance. I haven't seen anything worthwhile in that regard from Google yet. Or any corporation for that matter.

Teenagers are a big range from 13 to 19.

The difference in mental maturity and faculties at those ages are immense.

That's true, but between 19 and 22 also immense.

Regardless, the point is it's deciding that a person's moral value should be based on what they did in the past and not what they do now.

Unfortunately, the past is still the best indicator we have of current and future behavior.

It's a problem for many companies that hire young "intellectual" work force. Fresh graduates spend four years marinating in woke juices on campus and then try to bring that culture to the work place. The leaders often bend, and not to the benefit of the business.

I know I am sounding like Elon Musk but I am not really "anti" progressive. The holier-than-thou moral crusaders without any life experience or deep cultural understanding often get so "woke" that they accidentally come out racist on the other end. They don't realize it, however, because they are too busy cancelling someone for what they said 25 years ago.

I often equate this movement to the "family values" religious crowd of just a few years ago - both are equally self-righteous and insufferable.

the young learn from the old. their professors & TAs are the ones that indoctrinated them.

And the senior business leaders are all 40s-60s .

Who's to blame? It's on us to lead them

Indoctrinated? That's a common conservative talking point, except that many professors will just keep quiet in the fear of losing their jobs - no one wants to deal with a social media onslaught and sometimes even physical intimidation.

The young ones eventually grow out of it - but the "adults" need to have more spine.

That doesn’t mean it’s wrong

People are making this out to be a bigger issue than it is, making LLMs less woke will be doable and this will all be yesterday’s newspapers soon enough. I think releasing stuff and experimenting in public you’re always going to make a few mistakes, especially with something as complex as these AI systems. A genuine problem would be if Google were saying this wasn’t an issue. Let them fix it and we can all move on.

in my opinion the issue is much bigger. the subliminal influence has been there for about 10 years and it was the Gemini illustrations that made the bias so evident.

I think the anti-woke nonsense is just the old saying the young are idiots as every generation does. Nobody is arguing for characters from history to all be non-Caucasian, it’s just a mistake!

Exactly, they overshot the target. Bugs happen. The big mistake was not having team who actually knows what's going on to say "hey, you're over-correcting, hold the train. We got a bug here."

Reply on news.ycombinator.com