Comment by AlbertCory | Hacker News Reader

Comment by AlbertCory | Hacker News Reader

Parent Full thread

> But around 15 years ago, before the proliferation of gig work, deep learning systems were considered merely an academic curiosity, confined to a few interested researchers.

Um, no. Google was using machine learning more than 15 years ago, notably in the SmartASS system that predicts whether someone will click on an ad [1]. I got my cousin Missy a gig (at $15 an hour) rating ads and search results.

This story is hardly a claim of humane treatment of gig workers, I should note. I intervened and got her treated fairly.

> Companies make sure to hire people from poor and underserved communities, such as refugees, incarcerated people and others with few job options, often hiring them through third party firms as contractors rather than as full time employees.

Citations, please. A rater who is many standard deviations away from the mean is extremely undesirable. You want their judgments to be representative of the population.

If you want to design something that appeals to Americans, you want American raters. Since the US is by far the largest consumer of Internet services, it stands to reason most raters of "quality" will need to be from the US.

If you just want to know if an image contains objectionable content, then yes, you could use Third World employees.

[1] https://albertcory50.substack.com/p/working-at-google-ads

These “journalists” see someone training image classifiers using Mechanical Turk, which awards a wage in excess of the local economy in the third world, and imagine to themselves this is how all models get trained because they don’t understand ML, they don’t understand market segmentation, and generally don’t understand technology or business.

I don't see what there is to understand.

Any 'ai' model right now requires classified data to train. It takes human work to classify data. Human work is cheaper in some places than others, so companies in expensive places exploit the situation of poorer places to lower their operating costs.

The wage size compared to the local economy doesn't really matter, since the exploitation comes from the difference in costs between the local economy and the local economy of the company.

In your reality, how is the data getting classified if not by people whose situation is being exploited?

> It takes human work to classify data. Human work is cheaper in some places than others, so companies in expensive places exploit the situation of poorer places to lower their operating costs.

1. Humans are not fungible in this way, as much as some people wish that they were. The ML models we built in my workplace were fed classified data that was built using some of the most expensive labor in the company, because of the highly specialized nature of the model. It would have been impossible to train using Mechanical Turk.

2. Taking money from wealthy countries and putting it into the economy of poor countries by paying an outsized local wage in the poor country that is cheap in the wealthy country, thus lifting up the poor country, is a very interesting way to classify "exploitation" (see: Asia and Eastern Europe).

> In your reality, how is the data getting classified if not by people whose situation is being exploited?

In your reality, apparently it's only acceptable for a company that would pay a First World worker $20/hr to do classification to also pay a Third World worker $20/hr to do classification, rather than $2/hr, which is an outsized compensation in a local market where people generally get $2/DAY, rather than per hour.

In my reality, BILLIONS of people have been lifted out of abject poverty by foreign investment dollars that arbitraged labor costs, to massively improve the economic situation and quality of life of those foreign workers.

Weirdly, my reality aligns with actual facts.

You seem to be arguing from abstract principles instead of considering the actual facts. What tristor has understood and you have not is:

"People" are not all the same. I just explained why US workers are preferred or even required for a lot of ML tasks. Did you read that?

As for exploitation, I'll limit myself to noting that certain people are always finding it whenever a rich person hires a poor person, despite the fact that nearly all jobs arise this way.

> Any 'ai' model right now requires classified data to train. It takes human work to classify data

Interestingly that isn’t actually the case. I would guess most data used in ML is not human labeled specifically because it is so expensive and error prone. For example, most work done in the natural sciences simply utilize sensor data. Want to forecast some aspect of the weather? Grab a couple decades of historical data collected by weather sensors. None of that needs to be human labeled, only sensor collected. Some applications certainly do rely on manually human labeled data but not all, and I would personally guess only a minority of applications because of the difficulty in collection.

Quite true. Although there is unsupervised learning, much ML is supervised, meaning there is some ground truth. In those cases, the ground truth comes from sensors, as you said.

Many ML models dealing with human interaction do need a human to provide the ground truth. More and more these days you can probably find corpuses of data already labelled, so you don't have to collect it afresh.

How is that exploitation?

Why don't US companies pay foreign workers US wages? The answer I come up with is that they chose to hire workers in disadvantaged situations(relative to opportunity and local labour laws) to reduce operating costs.

To me, using people in disadvantaged situations to reduce your operating costs is inherently exploitative.

Nobody is being forced to work for those companies. Is there any actual evidence of exploitation? If so, are the governments complicit in not acting against them?

This is silly. Rents and product costs are low too in those countries.

Are consumers exploiting business owners there by this logic?

> Why don't US companies pay foreign workers US wages?

Why doesn't money grow on trees? Why does inflation exist? Why are some nations wealthy and other nations poor? Why are some nations at different points in their economic, social, civil, and technological development? Why is the sky blue?

Are you trolling or do you seriously not understand how fancifully inane this question is?

>> they don’t understand ML

Please don't say that, it is upsetting that this false assumption thrown around so carelessly. At least check the authors' profiles, say on Wikipedia, if you want to know how much they understand.

If they have credentials but still say nonsensical things, then what is one to think? That they "understand" it but still somehow draw the wrong conclusions?

For example, I'd start by asking who says they draw the wrong conclusions. Are they "the wrong conclusions" or conclusions somebody disagrees with? Who is the person who disagrees with those conclusions and what is the expertise of that person? And so on.

Not everything is a 20-page dissertation replete with footnotes.

> A rater who is many standard deviations away from the mean is extremely undesirable. You want their judgments to be representative of the population.

An excellent point.

Reply on news.ycombinator.com