Minimum wage ‘ghosts’ keep AI arms race from becoming a nightmare
latimes.com
latimes.com
All that nightmare data (which we generate, look I'm doing it now), is a reflection of the joys and traumas of people writing it, and if the majority of the NLP chat data is anti-social then there is no possible other outcome.
You cannot fix this with more "filtering" as the data is embedded in almost all written text. You have to change the sentiment.
[1]https://snap.stanford.edu/class/cs224w-2016/projects/cs224w-...
I have worked as a ML engineer close with internal labelling teams and let me tell you, not everyone could do it. It is a hard job, and often requires deep thinking to solve. Many ML engineers themselves are actually bad at labelling.
And the contribution of the labelling team can be equal sized or more important than that of the ML engineers. In normal conditions you need at least one labelling person / engineer, all projects need labelling, there's never a gap, labelling is never finished. Finding a good labeller is also hard - not everyone on the street could do it, contrary to public expectations.
Although I’m starting to think I ought to finish it soon just in case this current pace of advancement in LLMs doesn't stall out for once.
This is maybe caused by precisely the nature and the origins of the training data.
1984 was a big critic of communism, where facts dont mater and history can be rewritten. Are we really far from that?
* Those poorly paid raters label things - and a lot of things depends on quality of their work. Do those companies even cross-check their work, when the idea is to get the lowest bidder? (on a side note, why FAANG companies dont understand the "garbage in, garbage out" concept? They can pay 300k usd to some programmer, yet they skimp on getting classifiers from USA or Europe who know the cultural context)
* Then there are probably censors, who can censor things based on their own ideas. Probably censorship can happen at two levels: first by using labels, then by deciding what to show based on those labels.
* Some websites have "fact checkers", who are necessary (with the amount of lies written in troll-farms and soon lies written by AI), but you cant really appeal about their process in any way - and those are some faceless people from third world countries.
*Often you cannot ever contact with a real person, Google and Facebook are the worst here. Microsoft is a bit better, because if you pay them money for support, you can get human support.
* Wikipedia is a constant war between vandals, but also actors that are much smarter (seem state sponsored even). When we speak about Wikipedia it is incredibly sad that it never introduced any systems to review the work of their admins (I saw a situation where some articles are "guarded" by 2-3 admins who share their own >version of facts< and nobody can do anything about that). And this could be solved by setting up panels of judges that judge edits that are anonymous. Or setting a panel of judges who judge actions of admins (yes, there is that "Arbitration Committee" that barely works). Of course then there is the question "who watches the watchmen", but still - Wikipedia Foundation has millions of dollars that they waste on various things, instead of spending it to improve their core product.
On a side note, I always thought about starting such a data labeling business; I think a lot of work done by current labelers is still low quality. I saw some start-ups that were checking things (e.g. medical data) and they would pay peanuts to the people who labeled their data. Then they would get bad labels.. and build wrong conclusions on that with their classifiers.
Imagine the balls it takes to pretend like the United States isn't the biggest icon of "facts don't matter".
These jobs are not horrible, not for humanity's historical standards. They can be done from home, without commute, without physical and demanding labor. A medieval peasant, a 1800s factory worker and probably many of the 1900s low skilled laborers would take this over the alternatives.
Moreover, these people have every right and option to not do it, in fact right now the labor market is such that they have valuable alternatives!
We keep thinking about how this or that job is not acceptable for _our personal standards_, but until we get Fully Automated Luxury Communism we have to be honest that progress looks this: slightly less shitty jobs in exchange for higher quality of life.
Anyone of average IQ and education could do this work, and there’s no shortage of people willing to do it.
This is not a job that requires rare or difficult to acquire skills, nor is there a shortage of people willing to do the work.
The main subject — who appears to work from home and set his own hours — complains that his hourly wage is $2 less than his daughter working in fast food.
I’d argue she has a much more demanding job.
His first contribution, besides joining with his credentials was to cut down a few cents per hour from what we’ve been paying to a sweatshop some place in India. That was annotating images for us. From somewhat strange biases in the annotated data, it was likely that this sweatshop was employing children.
Incidentally, Google had acquihired that startup.
Do other countries label a caste of “prestigious” universities like the U.S. does?
> Stackhouse has a serious heart condition requiring medical management, but his employer, Appen — whose sole client is Google — caps his hours at 26 per week, keeping him part-time, and ineligible for benefits.
What a cruel system.
Benefits should not be yes/no at some hours-per-week cutoff.
Or really, they shouldn't be a thing. Fix any tax bs that makes it cheaper for employers to pay for them rather than employees and remove any rules that say employees have to provide them. ... err, I guess time off might count as a benefit, so make sure any rules scale linearly all the way to zero.
How much an employee costs shouldn't have sudden jumps at particular numbers of hours. Employees shouldn't depend on their employees for anything more than a (fungible, because it's just money) paycheck.
sounds like what you're saying is that without this employer, this employee would have a worse or no job; if you start from that and then give them this job then it sounds pretty positive
> they won't even let someone survive anymore.
who's they?
Some jobs will pay the lowest wage... that it is these jobs? I actually would rate and expect these jobs pay lower than the typical lowest pay jobs which require physical labor like the typical low pay jobs in fast food, or building cleaning?
Or the specific amount given current circumstances? True.. on the other hand still well above the minimum wage by law in most European countries.. maybe not comparable well, but still not much out of range.
That rich companies pay low wages? Yeah, absurd and unfair, but that unfortunately never hold a rich company back where it could just pay that (I'd like different regulation here).
So the problem is really independent of these specific jobs or companies? And the answer to this is better regulation and worker protection laws, that many frown upon?
Much in the same way that you can’t pay an industry supplying you with critical parts forever because then the parts will eventually stop being built, companies shouldn’t be allowed to pay for less than the necessary amount to support the workers living. If this makes the job they are doing not valuable enough then the companies shouldn’t be engaged in the work that can’t be profitable without subsidies, ask the government for specific earmarked subsidizes if it generates a positive externality, or actually do some innovation and figure out a way to make the work profitable
That was my point, or "really?" Again, I'd like it to be more that way, but given current laws and what most people agree to: Do they even have to? No!