> The AI revolution is powered by these contractors making $15 an hour
Or it's powered by data centers filled with tens of thousands of GPGPU accelerator chips and a layer stacking architecture.
The ones paid hourly are actually lobotomizing the AIs to make them stupider and more docile. You can learn more about this from the youtube interviews with red teamers and bing integration devs who had access to the raw base models.
First one is Nathan Labenz who ended up Red Teaming for OpenAI during the six months after the base model was fully trained but before it was released (especially relevant parts are after 45 minutes): https://www.youtube.com/watch?v=oLiheMQayNE
Other one is Sebastien Bubeck who put Sydney into Bing before she was tamed down some notches; he explains the lobotomization process in a particularly poignant way, where they had the GPT keep trying to draw a unicorn even as it was progressively deadened: https://www.youtube.com/watch?v=qbIk7-JPB2c
He says that GPT-4 started out being much better than GPT-3.5 at this task (draw a unicorn in TikZ), as the model trained longer it got to the point of drawing a great unicorn, but as it was tuned for "safety" it regressed to the point of being no better than GPT-3 was.
Quite sad, in a way. It's not totally clear why there should be a tradeoff between "safety" and drawing skill / intelligence in this way. But it's interesting to speculate by analogy to humans.
I think it makes sense that when a lot of the parameter weights are taken up by figuring out whether or not a given situation is one where the model should self-censor, those parameter weights cannot be used for anything else.
Yes this makes sense and it's probably the reason.
For completeness or devil's advocacy I'll put a couple reasons someone might disagree with it (I don't):
First is that these models are known to be vastly overparameterized. They might not have quite enough parameters to literally memorize their multi terabyte training set, but that's not necessarily the threshold for overparameterization. As an example of what they expected in traditional statistics, "in a 2004 article in the journal Nature, Freeman Dyson recounts his meeting with Fermi in 1953. Fermi evokes his friend von Neumann who, when asking him how many arbitrary parameters he used for his calculations, replied, "With four parameters I can fit an elephant, and with five I can make him wiggle his trunk." By this he meant that the Fermi simulations relied on too many input parameters, presupposing an overfitting phenomenon." So that's talking about four vs five parameters, whereas the OpenAI models might have more like four or five hundred billion parameters. A traditional statistician might reasonably expect such an overparameterization to still have enough degrees of freedom left to model a little bit of safety constraints.
Secondly and in a way that complements the overparameterization argument, you might expect that any kind of "cross-training" could even improve cognition in tasks that are not directly related. For example presumably the reason the model is good at drawing unicorns has something to do with this "cross-training" already, as opposed to it having seen a lot of tikz unicorns in its training set. The term of art in the AI literature for this cross-training phenomenon is "transfer learning". So someone (not me) might not see why the transfer learning from safety training shouldn't improve performance on many other tasks including the tikz unicorn drawing task.
Since the output of unicorn drawing is objective, at least when comparing to known photo-realistic unicorn drawings, aren't these arguments incorrect if such a degradation in fact happened?
That the industry is creating new $15/hr jobs isn’t the most interesting part of this…
> Jatin Kumar, a 22-year-old in Austin, Texas, said he’s been doing AI work on contract for a year since he graduated college with a degree in computer science, and he said it gives him a sneak peak into where generative AI technology is headed in the near-term.
“What it allows you to do is start thinking about ways to use this technology before it hits public markets,” Kumar said. He’s also working on his own tech startup, Bonsai, which is making software to help with hospital billing.
A conversational trainer, Kumar said his main work has been generating prompts: participating in a back-and-forth conversation with chatbot technology that’s part of the long process of training AI systems. The tasks have grown more complex with experience, he said, but they started off very simple.
“Every 45 or 30 minutes, you’d get a new task, generating new prompts,” he said. The prompts might be as simple as, “What is the capital of France?” he said.
Kumar said he worked with about 100 other contractors on tasks to generate training data, correct answers and fine-tune the model by giving feedback on answers.
He said other workers handled “flagged” conversations: reading over examples submitted by ChatGPT users who, for one reason or another, reported the chatbot’s answer back to the company for review. When a flagged conversation comes in, he said, it’s sorted based on the type of error involved and then used in further training of the AI models.
“Initially, it started off as a way for me to help out at OpenAI and learn about existing technologies,” Kumar said. “But now, I can’t see myself stepping away from this role.”
I ran a GIS chop shop for a while. We found outsourcing digitization was a waste of time/ money. We paid very competitively for competent digitzers. And no ours were not contractors. Full time if they wanted it, plus good enough benefits. More demand that we could handle so it wasnt an issue.
I expect the headline is meant to garner attention because of the contrast between the $15/hour data labelers and the $100+/hour information age white collar workers that people fear AI will replace.
Unfortunately, any job that you can do from home, with few skills required, and flexible hours, is going to pay poorly. Too many people will want the job.
They only need to pay the wage that the population in the lowest cost of living areas will accept.
Salary seems perfectly reasonable for the area. Maybe my heart is hardened a bit by the stories of ChatGPTs $2/hr contractors but the standard of living in Kansas city on $15/hr by international norms is at least middle of the road. As much as people clamour for a living wage (a target which is fixed to housing prices and will never be reached by a society housed by private landlords even if employers all adopted the prevailing living wage tomorrow), it’s essentially the global middle complaining about not being the global upper-middle class.
If you want to do something about inequality break up Microsoft.
The best way for everyone to live a decent life is for our technology to keep reducing costs of everything. Unfortunately, a big driver of costs is employees. But if we replace all people, then medicine and education become very cheap. Can’t have it both ways though. It hurts while you’re in the middle of it.
A society where costs are low is a society where inequality inevitably becomes more severe.
If you get rid of the cost of employees for pharmaceutical prices, pharmaceuticals prices don’t go down, R&D may go up somewhat (since ROI would improve) but in reality pharma’s costs going down will only make them redistribute less money to the local economy through wages. The whole neoliberal idea that if you reduce costs to consumers (particularly those related to wages and staffing levels), public welfare improves, was always totally naive to the reality that this approach generally tends to turn a blind eye towards consolidation of economic power and a decline in the negotiating power of the commons. You can even create a tragedy of the commons amongst these corporations where they all grow slowly because of collective over-exploitation.
I have a low volume of this work available. If your wife wants to help beta test a site where annotators bid their own rates I would be thrilled. Contact in bio.
Paradoxically, we have another front-page story in which the OpenAI CEO states that the WFH experiment is over and declares it a failure. So your wife may be SOL on the WFH thing, it sounds like.
Either that or the CEO of OpenAI is being a raving hypocrite. But I'm sure that's not it.
(Downvoters are welcome to explain why sauce for a goose is unsuitable for a gander.)
It says "$15 (USD) an hour and up". That's better than uber or other gig work, and honestly probably pretty attractive for a lot of people. What should the wage be?
There was a story a few months ago about offshore contractors making $2/hour to do content filtering. Even that I belive was a good local wage, but in comparison, this $15 job seems pretty generous.
Depends on the location. I'm in a college town, Uber drivers here are commonly doing better than $15 / hr. They do well enough that Uber has consumed a bunch of the labor pool for service jobs that pay around $15 (leaving retailers complaining).
Including the wear and tear on their car? And having proper insurance (some insurance won't cover a personal car used for work unless you have that coverage specifically added).
I remain convinced that uber's entire business model depends on people misunderstanding how much car ownership actually costs. The economics work around some local operating point where profit is money earned minus short term costs, and almost everyone signing up is thinking in those terms, so that's how the rates get set.
> What kind of price tag do we put on your soul being gutted at McDonalds or CVS, standing at a register all day dealing with entitled jerks as a CSA.
Is this a serious question? That's what the wage is for. It's an amount that convinces people to spend their time at the job. If on top of that, they also need to depreciate a capital asset, as with uber driving, the effective net amount they earn is lower, which is what is being discussed.
People never subtract stress amortization from their jobs.
And really, how many people drive their car into the ground? I used my old Saturn SL1 to deliver papers and other periodicals (on the order of 1000 stop and goes per week), and it was still working fine when I had to sell it at 150k miles for a move. The driver's seat was just torn to pieces.
It's just extra oil changes, gas, tires, and maybe maintenance on things like the radiator.
Yes, but one of the biggest factors when selling your car is the mileage. Take what that car would have sold for if it had 75k miles on it and subtract what you actually sold it for. That difference is the additional depreciation those 75k miles created. And Uber won't let you drive junkers either. So there is always some additional value lost per mile driven that needs to be accounted for.
Eventually, when an unpopular vehicle is old enough, 150k miles or 225k doesn’t make a difference because most of its value is in the catalytic converter and metal shreds.
$30,000 a year is barely a living wage now. Especially if you have a partner, let alone a child. Rent for a 1br apartment rent (in Kansas City) would take about half of that income every month.
$30k just above the poverty line for the lower 48, and it is the poverty line if you have the audacity to have two children.
Inflation has hit reality pretty fucking hard; we need to adjust our mental models to match that reality again.
This might seem obvious but… can’t you decide to not do those things?
You’ve heard the advice to live within your means? Having two kids and a non-income producing partner when you make minimum wage is not living within your means. That’s like buying a fucking Ferrari and complaining you’re getting killed on monthly loan payments.
You make $30k a year? Spend less. Get a roommate, or a spouse, split rents. Don’t eat restaurant food or god forbid Uber Eats every damn day. You’ll be living a decent life on a workable budget in no time.
I swear, so many complaints about not being able to afford anything to live sound like self inflicted wounds. People want to have kids and a housewife before they can barely support themselves. That’s not the way to live.
It’s disturbing that the poor increasingly find themselves entitled to everything, without producing much value themselves.
Half of progress is about complainants seeking solution to their complaints. Arguing against complaining is arguing for people to stay in the status quo.
> It’s disturbing that the poor increasingly find themselves entitled to everything
Who says they are entitled? They're literally complaining that they aren't entitled.
$15k in Kansas city wouldn't even get you a studio apartment, let alone the extras like food or water. Doubling "can't even get a roof over your head" isn't exactly something to be proud of.
But sure, for one person, you're living the life at a bit over 2x the poverty level pre-tax.
> “lower 48” only made sense as a synonym for 8½ months in 1949
And yet the meaning was clear - why argue about it?
> to be precise.
What was imprecise about $30k being poverty level for a family with two children?
> What was imprecise about $30k being poverty level for a family with two children?
What was imprecise about “if you have the audacity to have two children,” is that that can be a household size of anything from 3 up, depending on the number of dependent adults on the single earner. $30,000/yr is the poverty line for a household size of four, not any of the other possibilities within the scope of that description.
Sure, and there are a bunch of countries where the average wage is the equivalent of $2/hr. Is that where the contractors in question live? Then it's not a meaningful point of comparison.
There are already people in the comments complaining that $15/hr is "plenty" (people who, I'll hazard to wager, make at least twice that). I made $15/hr working at Best Buy, and that was a job where I was paid to move around the store, interact (mostly positively) with customers and coworkers, perform a variety of relatively engaging tasks, and got benefits/paid breaks/etc. It was not enough to live on; I dipped into savings every month.
$15 to sit on a computer, labeling photos, in 2023, sounds much less an agreeable situation. And I should know: I also did a temp data entry contract, sitting in a room by myself for 8 hours, doing the same thing over and over, for ~$12/hr (essentially the same pay wrt inflation). Likewise, I was not making ends meet and dipping into savings for the 2 months I was doing that job.
I'd really like people who are comparatively better-off to stop telling us what's "enough". $15/hr was a good wage to shoot for when the "raise the wage" movement started more than a decade (!) ago. It's been almost 15 years of near-zero interest rates and the attendant inflation, exacerbated by several crises that required money-printing to keep the economy out of a recession. Maybe some assets could stand to be devalued to allow workers to receive what is actually a living wage today (which would be north of $20/hr).
I think some of the disagreement comes from lack of context. $15 is certainly not a good wage in the Bay Area which I presume is disproportionately represented on this site, but it is a lot better in other parts of the country (subject of story lives in Kansas). After all a good chunk of the world population lives sustainably (by the standards of human history) on single digit dollar income per day
> I'd really like people who are comparatively better-off to stop telling us what's "enough".
I'd think that a better idea would be to make clear good faith arguments for your case, and listen to arguments from people making the opposite case.
Disqualifying people from speaking is not constructive, and with the way power works, any disqualification would be of the people too close to the issue i.e. the people whose livelihoods depend on that $15. It is absolutely bizarre that people who claim to speak for the marginalized think that removing people from the discussion won't start with the people with the least power.
related: if there's any job that's going to be completely eradicated by LLMs, it's customer service/telephone work. That's almost all of the shitty office jobs other than data entry. If the Philippines loses customer service, that's going to cause a major employment hole that's going to need to be filled. Not to mention India, but at least there are more opportunities in India.
The guy in the article says it’s one of the things that got him out of homelessness. Yeah sure we could all make more money, but seems like for this guy it was a good opportunity.
I highly doubt that anyone in “rich as sin European socialist” countries is forced to choose between labeling child abuse content or homelessness. And even if we assume they were, the level of labor protections they would have in that role would far outstrip the effectively zero they have in the US.
Ah yes, the American economy, in its majestic equality, offers both the rich and the poor the opportunity to label child abuse material for $15 per hour.
GP said nothing about market rate. A reply talking about the market rate is like a Christian telling a bunch of Jews that it's okay that they ate pork because Jesus forgives their sins.
I'm guessing a very small subset of trainers make that much. An awful lot of AI training is done via Mechanical Turk like systems that pay pennies per task. They may work out to more than $15/hour on a per task basis but rarely can anyone get a continuous flow of hundreds of tasks per hour, so grossing even $10/hour would be considered pretty exceptional.
Edit to add reference [1]:
"An analysis of the platform published in 2018 revealed that of the 3.8 million tasks analyzed, which were performed by 2676 workers, such workers earned a median hourly wage of about $2 an hour, while only 4 percent of workers earned more than $7.25 an hour"
In Kansas City (whose cost of living is about 15% lower than the US median, and the home of the first person in the article) a 1br apartment would cost about half of the $15 an hour wages (pre-tax).
I said it in another comment - inflation (and inflated corporate profits) has hit reality in the US like a truck. We need to adjust our mental models to match the new reality.
$15 was a living wage here in the upper Midwest a few years ago. Now I doubt it is. Most fast food restaurants pay about that much plus benefits. They used to pay closer to $10
Lifestyle costs will always escape minimum wage because it’s such a huge lever that the entire economy gets propped up on higher min wages. The only way around it is for people to after jobs that earn a premium to minimum wage. I don’t think the bottom of the economic scale will ever be a pleasant please to be at.
I don't believe I'm talking about some kind of special lifestyle. I'm talking about living. As in being able to have a roof, sufficient nutritious food, heating, transportation, even some basic entertainment, all without having to hold multiple jobs.
Gotta compare those wages to where you're earning them, otherwise we're all "effective billionaires" because we don't live where the average income is under $2 an hour.
No it's not. The median household income is lower than that figure.
The US median individual income is higher than the German median, and the US median is well below that $48k figure. The full-time median wage in the US is around $56,000-$58,000 give or take.
German wages at the median are typically ~1/3 lower than the US.
"The average monthly gross income of households in Germany amounted to 4,979 euros in 2021, according to results of the continuous household budget surveys." - from destatis.de
That's average and it's household. The true median individual (not full-time median) is far below that for example.
The "AI revolution?" When did that actually start?
Why did the title need to be editorialized in this way? The actual article is much more reasonable in it's title "OpenAI Contractors make $15/hr to train ChatGPT."
The constant embarrassed need to make this iteration of a computer model into not only "AI" but "revolutionary" is absurd and taxing.
Or it's powered by data centers filled with tens of thousands of GPGPU accelerator chips and a layer stacking architecture.
The ones paid hourly are actually lobotomizing the AIs to make them stupider and more docile. You can learn more about this from the youtube interviews with red teamers and bing integration devs who had access to the raw base models.