Cheap, outsourced labour in Africa is shaping AI English
theguardian.com
theguardian.com
The thesis was that it won't matter if LLMs exhaust the available corpus of text that humanity has generated thus far, because it will always be easy to pay humans to create an inexhaustible supply of more text.
The article suggested that it was all going to be PhDs, lawyers, and poets creating this training data, but that all sounded like hand-wavy BS to me. I assume that, if more text is needed to train LLMs in the future, the vast majority of it is going to be pennies-on-the-page, mechanical turk-type workers. And the LLMs are going to reflect that kind of input.
[0] https://techcrunch.com/2023/06/14/mechanical-turk-workers-ar...
…that was generated by LLMs. So no you’re back to problem of getting data generated by humans.
But LLMs are not only used in corporate settings. Cutting the bottom line in medical professions, case law, academic journals, etc. may seem enticing, but their audiences will definitely bear the brunt of the loss of specialist diagnoses, case work and peer/lit reviews.
... how many (and to what extent) are today's students/trainees and future practicing pros already passing classes that they would have otherwise failed, undetected -for example- and undermining future professional services which we will all avail ourselves of?
Like why pay more for AWS when you can host things locallly. People just can't be bothered with the admin
https://static.googleusercontent.com/media/research.google.c...
Say what you will about the merits of this paper, it's hard to deny that it has been influential within the deep learning community.
(been offered contracts, have done some on the top end of the market, have access to data on some of the volume on the lower end)
Like the overseas human CSR disconnecting, and needing to go talk with the counselor (that the contracting firm reluctantly made available, under PR pressure), because this is the dozenth perv/psycho she got so far this morning?
Now along with all the things they get screamed at for already, they're also going to have people describing horrific/disgusting things to see if they're really human.
Actually, thinking about it, it's probably not even that boring.
Reference: https://ahdictionary.com/word/search.html?q=dribble&submit.x...
Or check out the scene in Burn After Reading where Frances McDormand confronts a Russian diplomat.
Is that where your billions come from? You comfortable with that?
That's not how I run my company.
You don't pay people less because you have them by the short and curlies of their own poverty. That is exploitive. You pay them a fair share of the value they're producing.
The problem is that if corporations did this, if they stopped exploiting, if they paid a fair share, then those billions in valuation magically vanish. It's because the valuation was predicated on exploitation to begin with.
I'm not seeing how this is a "shell game" in any meaningful sense. In contrast to the actual game[1], everyone knows and gets what they signed up for. There's no deception going on aside from a vague "but I thought corporations were supposed to give everyone a fair share (whatever that means) that idealists have".
>You don't pay people less because you have them by the short and curlies of their own poverty. That is exploitive. You pay them a fair share of the value they're producing.
What is a "fair share"?
I don't think invoking this debate in the context of lack of jobs in Africa, is useful. These are two completely different problems with different causes.
They could try to exploit Western workers as mechanical Turks but their strategy would then fail. The strategy succeeds because the rate of exploitation of workers in impoverished countries - in terms of actual share of the real dollars produced - is much higher in the global South than the global West.
Worse we're now all guilty and complicit, in using LLMs, and even in using open source llms train second hand on the unethically sourced llms. We will have to pay reparations.
By making a false equivalence between the two worker classes, you potentially support the continued existence of the globalist hyper exploitation strategy, and prevent Justice for AI workers. You can speak out against it instead, act against companies who use it, and move towards understanding what reparations might entail.
I think the exploitation of low-paid mechanical Turk style laborers is a real problem and we need to talk about it, rather than background it just because workers in the global West are exploited as well. True global worker solidarity means seeing this as a place where the iron is very hot and ready for striking.
Finally, the scarcity of jobs in Africa is not a problem being seriously discussed here. It's being used as a red herring to justify exploitive investment.
It's hard to imagine that you have anything constructive to contribute to a business conversation, when you see someone who acts out of ethics and assume they're acting performatively. That just means you don't believe ethics are real. That's a very dangerous attitude in a business person.
Curious about the company that you own, what its business model is. I bet you're in crypto.
And no I'm not in crypto. I've been saying it has no lawful utility so voluminously that mods have asked me to stop trolling crypto people. So we can establish your assumptions and fundamental intuition and instincts are very inaccurate. You have a broken compass and it has caused you to become disconnected from first principles of microeconomics.
What is your company name and who are the unlucky shareholders you're victimizing?
Last, I hope you can learn the difference between kindness and niceness. The former is legitimate while the latter is what you do - a manipulative social strategy, often employed by predators with subclinical psychopathology, to ingratiate yourself to others in leu of merit.
Now you're charging me with causing death of children. f** you. Laying the charge that I'm being nice to manipulate, while being manipulative.
Anyways if you want to argue that you need to reduce your workers to squalor and keep them there in order to be a virtuous business person... Guess what? I'm very comfortable with you calling me a bad person then. Because to you the good person is one who enslaves customers and workers alike. That's absurd to me.
Anyways, your argument is that a large mega national corporation should exploit the cheapest market because that saves lives in that market. I argue that they only use that market because it is exploited, and their strategy will keep it that way. So they are perpetuating the conditions that cause the deaths you are trying to prevent.
Anyways you keep making heroes of multi-billion dollar corporations, and I'll keep trying to dismantle them.
Not going to tell you my company and have you attack it, but I'll point out that I'm the sole investor, and every worker gets a full copy of our engineering stack because we run open source.
I'm not overpaying anyone. I'm paying everyone the fair share of the work that they produce. You want me to reduce them to serf's because that maximizes the companies utility and is good Management in your mind. Paying someone less because you can get away with it is an abuse.
You're going to keep projecting that I'm being manipulative and nice rather than simply having some ethics and having worked very hard to free myself from under the thumb of abusive bosses.
So f** you.
I'm going to f** you up. Logically I mean. In debate I mean. I don't think you have the balls to end this conversation with me. It will take more blood sweat and tears than you have in your body. And I will beat you effortlessly. Cuz I'm everything you think of when you think of a hateful human. A human that doesn't deserve to exist who continues to claw there body against the fabric of reality and exist. You hate me. And I will defeat you. Let's talk. Debate. Not afraid of you.
But that's if. And that's it. If you're not also on Marxist, arguing that AI companies need to be paying more than the market rate, and making the workers part of the company. Then I think you need to leave the conversation because I have won.
But I'm not scared of you.
How could we chat voice to voice? I'm curious to see how your abusive unethics, illogic, and harassment tactics will unfold in a voice chat. The benefit is that it will help me harden myself against counter-revolutionaries and pro-capitalists. Considered doing me the honor!
You would have loved him. Maybe you are him Chris is that you?
Too bad I buried him. Long ago in my heart.
The business is mine today. Operating like I wanted to. You hate me. It Ain't that I took it from you. It's that it was mine all along.
You want to call me crazy because I run a company and I don't exploit labor.
What's your name what's your name what's your name what's your name what's your name is your name Chris?
Talk to me
Talk to me
The more information u reveal the more information u talk to me.
So tired of dealing with people like you. So tired of people wanting to attack me just cuz wanting to do good business
Wanting to do decent business.
Wanting to do good business.
I still remember working retail, I'd love LLM work over having to stand up for 8 hours.
My retail job paid pretty well though, so that made up for a lot of the usual retail criticism.
For me, I’d say it depends heavily on what you have to sell.
I don’t think I’d be able to handle working a cheque-cashing place or a car dealership where you’re incentivized to sell what someone doesn’t need.
Are you being paid market rate for where you live, without coercion and without collusion? And does it meet local minimum wage laws? And are you fully informed of working conditions, and is physical safety properly regulated?
If so, it's hard to see what's exploitative about it.
And nothing in the article suggests that any of the things I described (coercion, collusion, etc.) are happening here.
EDIT: this is US-centric of course, but I am fairly certain that outside of the EU, the story is much the same.
1. I doubt the employees in question were being paid anywhere near minimum wage, so I don't see how this is relevant
2. That just moves the question around. If the minimum wage was adjusted for inflation, does it make it magically not exploitative? What makes the original minimum wage not exploitative?
>and continue to lobby Congress (at a state and federal level) for favorable laws to protect them. If that’s not collusion in spirit, I don’t know what is.
You're going to have to be more specific here. Not all lobbying is "collusion in spirit" (whatever that means). For instance arguing to weakening environmental regulations might not be great, but it's not exactly "collusion" either.
You implied that wages being above the local minimum wage was a mark against exploitation. I counter that if business leaders (via lobbying, campaign donations, etc.) have kept the minimum wage lower than it would otherwise be, this is de facto exploitation; they’re just able to hide behind the law (that they created).
As to where the magic line of exploitation is, I don’t know, but it seems to be along the lines of Justice Stewart’s definition of obscenity – I know when I see it. I don’t go so far as to say that paying anyone less than their output ultimately earns is exploitative, but my floor is substantially above minimum wage in any U.S. state.
> arguing to weakening environmental regulations… isn’t exactly collusion
If the EPA gets staffed by politicians eager to bend to the will of business, I don’t see how that isn’t collusion.
That's probably not the best thing to model your position off of, considering that was replaced with the Miller test in 1973, and he regrets saying it.
https://en.wikipedia.org/wiki/I_know_it_when_I_see_it
Moreover because it's opaque, arbitrary and inscrutable, it makes it impossible to have any constructive conversation about why exactly it's exploitative. For instance, you claim that the minimum wages in all US states are "exploitative", but provide no arguments in favor.
>If the EPA gets staffed by politicians eager to bend to the will of business, I don’t see how that isn’t collusion.
That's a subset of "lobbying" though. Recall my claim isn't that lobbying involves no collusion, it's that lobbying could involve no collusion.
OK then, _you_ define exploitation. And if you bust out the dictionary, "the action or fact of treating someone unfairly in order to benefit from their work," then I request that you define unfair. That's the problem – this is all highly subjective, and it's unlikely that we're going to agree.
> That's a subset of "lobbying" though. Recall my claim isn't that lobbying involves no collusion, it's that lobbying could involve no collusion.
Sure, and ISPs could treat customers better, but they're not going to. Lobbying as the U.S. has it is essentially all a form of collusion and/or bribery. Donate $10K to my campaign fund, and maybe your desired verbiage winds up in the bill. Treat my industry how we'd like, and there's a sweet post-government job waiting for you on K Street.
I don't deny that the ability for a person to go lobby their representatives is a critical part of a functional democratic republic. And, it makes sense to have a centralized voice – for example, it wouldn't be efficient for every single farmer to go speak to their reps independently. But at some point along the line, it stops being actual people discussing their issues with their elected officials, and becomes corporations showering government in money to buy the future they want.
Unless your a full blown Marxist you need to accept you'll never be paid the full value of your labor.
It's a matter of what your other options.
Would you rather lay bricks in the hot sun, or sit at a computer for 8 hours a day ?
Have done so for 5 years now.
We pay all workers their fair share of the socially necessary production time. If the organization needs surplus to invest in equipment, we requisition it from the production with consensus from the workers.
I had to kick a number of capitalists to the curb to achieve this. Why? Cuz they wanted to exploit workers, and they wanted me to do it for them.
Get lost!
You can get paid the full value of your work, if you seize control of the business. Just don't be a wimp, and out maneuver the pudding headed capitalists who know no actual economics or category theory, lol. It's f*** easy.
Even if only queer inclusivists join the movement, we can still win. But if anti-queers infiltrate the movement, then it's in trouble. So I aim to make anti-queers feel very unwelcome, and also see the movement as absurd. That's why I integrate the queer identify into my business theory and revolutionary theory.
What is a "fair share"? Suppose we're in a SaaS business. Does someone working in customer support earn more or less than a SWE? What makes the difference (or lack thereof) fair?
>If the organization needs surplus to invest in equipment, we requisition it from the production with consensus from the workers.
What is a "consensus"? What if 2/3rds of the people want to invest but the other 1/3rd doesn't? Would they forced to contribute? Are how are you going to "out maneuver the pudding headed capitalists" if you can only seek to raise capital from your employees?
Just calculate your opportunity costs, talk about it transparently, and negotiate socially necessary contribution factors for each department and each role. And sure, there will be departmental politics as each department tries to argue that it's much more critical to the business than another. But with openness, that's much better than that absolute f*ery we see in most companies. In most companies only the executives have access to that level of information, and everyone else is f*ed.
Consensus means everyone agrees or nothing happens. My team is four peeps, and we're pretty good at persuading each other into consensus, especially since we're open about our profit sharing and tend to make good cases for our initiatives. If we can't all agree then we don't do it, and just live with it.
And the pudding headed capitalists are gone. I took them to the curb with a number of maneuvers that I might want to write up one day, but may not have been strictly legal. And even with all of that, I tried very hard to accommodate them. But they burned me, and I went nuclear. Things were much better after then.
Once you no longer have any exploitative extractive sunnuvab**es around, you think they're going to make money just by having money, it's pretty easy to be open and honest, and reason through things.
>Just calculate your opportunity costs, talk about it transparently, and negotiate socially necessary contribution factors for each department and each role.
How does this work in practice? Going back to the SaaS company example, the company arguably wouldn't exist without SWEs, but could theoretically limp along by ignoring all customer support tickets (see: google). Does that mean the opportunity cost is 100% for SWEs and therefore they should get 100% of the revenue?
>Consensus means everyone agrees or nothing happens. My team is four peeps, and we're pretty good at persuading each other into consensus, especially since we're open about our profit sharing and tend to make good cases for our initiatives. If we can't all agree then we don't do it, and just live with it.
That works well with a small team where you're all friends, but how do you prevent bad actors from using their veto powers from sabotaging the group? eg. hungary and poland abusing their veto powers to block proposals that all other EU members support? As the group gets bigger the chance that you have a bad actor is going to increase.
>And the pudding headed capitalists are gone. I took them to the curb with a number of maneuvers that I might want to write up one day, but may not have been strictly legal.
That doesn't answer my question. You're at a capital disadvantage because you can only raise capital from employees and only if they all agree. That might be fine for a capital light companies like a consultancy, but you can't possibly expect the workers at a nuclear plant to stump up the entire construction cost of the plant.
I'll attempt to think of a model for
1) A fully open source saas that needs ethically sourced capital to build a data center
2) A community that needs capital for a nuclear power plant
I'm not entirely sure that we can't just obligate Capital holders to give the capital to the workers. So for example in the case of the nuclear power plant where the workers are to stump up the capital, they could visit the estates of the wealthy people living in the area, and requisition the capital. That's kind of like taxes. Or a protection racket.
In the interest of the conversation, I will try to think of strategies that don't involve violence or refactoring of the rule of law. But keep in mind that significantly limits things.
Yes, Delve has appeared to become more popular, and Even if its more common in business Nigerian, that's unlikely to be the cause of its popularity.
https://books.google.com/ngrams/graph?content=delve&year_sta... Its been getting more popular in british english, so perhaps its over represented in something like news paper archives?
Now, if we want to talk about working conditions, I'm all for it. But if we are going to make this poface and very guardian peal clutcher argument against something, can it not be for nigerians (I hope would be) getting a decent wage and conditions?
There are so many ethical issues with AI, but this isn't really one of them.
These are all bullshit business buzzwords. You will see them in any product description. I'm surprised "synergies" isn't in there as well though that may be too 1990s.
I can totally vouch for ChatGPT's total (ab)use of that word.-
(That and of "In conclusion" and of "In closing" ...)
Firstly, paying someone to do a job is not "exploiting". I mean, I guess it is if exploit is pared down to just mean "use for a benefit", but it is a pejorative in this context, meant to clutch some pearls and virtue signal (in the weird "I'd rather they don't get paid or be a part of this" way). I'm Canadian and to Americans getting skilled work done in Canada is usually cheaper, and I guarantee you we don't see it as exploitation.
But then it argues that "their" language in Nigeria/Kenya -- you know, English -- is being exploited by being "pushed back" on them. Bizarre. Not only does the contention make zero sense from an RLHF perspective, it's spurious and backed by nothing. Some random guy on YouTube said "Oh yeah this word must come from Kenya". Delve? I see this word regularly, and have for decades. The other words -- “Explore”, “tapestry”, “testament” and “leverage” -- are in no universe rare words. They may appear more frequently than random Facebook comments because OpenAI weighted the inputs and put more emphasis on high quality sources where the grade level is a bit higher, which makes infinitely more sense than it being some manifestation of Kenya or Nigeria.
Similarly words and phrases are viral, especially in niches. That words come into or out of favour in some realm like papers on PubMed by itself proves little and isn't particularly demonstrative. We all stand on the shoulders of giants, and we're all aping what we read and what we see, etc.
I would imagine that those other words (which seem very common to me) are being used in ways that would seem idiosyncratic to an American English speaker. Like I see tapestry all the time in very flowery language ("The country has a rich tapestry of influences ..."), but I would be amused if I read it being used to describe contents of a backlog at work.
I will disagree strongly with this:
>That words come into or out of favour in some realm like papers on PubMed by itself proves little and isn't particularly demonstrative.
This is ignoring what appears to be an overwhelming effect (an increase of around 1250%) that kicked off with ChatGPT's inception. I don't think you outlined adequate grounds for ignoring this as unrelated, given the timing, the magnitude, and the ready explanation of "delve" being used frequently in countries where the model was trained.
The argument is not that these words are being misused or abused. They're comparing the frequency on the web in general versus the frequency in GPT output, which is not useful or illuminating. For all those words, including delve, the usage is completely cromulent and grammatically correct, but someone needed to spin it into some egregious exploitation.
> an overwhelming effect
Delve was on an upswing for years. Again, it takes just one popular paper to use the phrase "delve into the results" and an army of new papers are going to use the same phrase because we're all stochastic parrots to some degree. Just as Attention Is All You Need has massively influenced countless subsequent papers.
Again, there is zero evidence that ChatGPT even overuses "delve". The guy's evidence that GPTs overuse delve is that it used it a few times and he thinks that's overuse. Okay. Similarly, the evidence that Kenya and/or Nigeria use delve more is just...well it just works with the whole tepid argument okay?
> "delve" being used frequently in countries where the model was trained
The model was trained on TBs of internet data. A small amount of, essentially, "tone" training is done via RLHF, but that wouldn't suddenly make local dialect surface.
Not that this exact phenomenon applies to “delve”, but it shows how word frequencies can vary across cultures regardless of “grade level”.
That's not just an issue of frequency but redefinition. I read about English-speaking Indians using the word "prepone" to mean "pushed forward" (as an inverse of "postpone"), and they were shocked when non-Indian English speakers didn't know what they were saying -- because it's not a word in British or American English, and AFAIK it's not a word in among any other native speaking English societies.
Neologisms and re-definitions aren't inherently bad, but it can make communication more difficult.
This list does not include the UK, the US, Australia, or New Zealand.
https://en.wikipedia.org/wiki/List_of_countries_and_territor...
during the dawn of transformers, i recall having to do manual labeling of a private image dataset as part of my internship "duties". funnily enough not much has changed in terms of how we still deal with data.
it would be nice to also have a peek inside the RLHF bit using user interactions that openai is benefiting from.
If robots halve in cost every 10 years, and at the same time 10x in functionality and accuracy, then in 10-15 years, they will be paid less than most people on the planet for the same amount of work.
Maybe not washing dishes, laying bricks or constructing houses, but there will always be work. Things will continue to get cheaper (and better), but they will always take _some_ human effort to make. (And we'll always find new things we want.)
There are several billions worth of investor capital that will need to be repaid first, plus whatever equity stake the founders choose to award themselves with.
See ants, which can figure out to dig holes and decompose everything. They know how apply minimum amount of pressure to cut leaves, they know how to choose the right angle to cut animal muscle and even wood, they know to cooperate and survive floods, and so on. It is no wonder they have the most brain-weight/body-weight than any other animal on the planet.
Excluding the processing unit, any robot part will be so cheap and available that no one will need permission of capitalist monsters to obtain.