ChatGPT is powered by contractors making $15 an hour
nbcnews.com
nbcnews.com
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
https://youtu.be/qbIk7-JPB2c?t=1584
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.
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.
> 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.”
For comparison, your fish is 'powered by contractors' making $0.1 an hour in Africa or the Caribbean.
Then people import and refine the product and sell it to you for $30 in the restaurant as a value-add. That's how the economy works.
They only need to pay the wage that the population in the lowest cost of living areas will accept.
If you want to do something about inequality break up Microsoft.
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.
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.)
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.
Pre-pandemic, used cars delivered excellent value.
Buy $2k car, put 50 000km, scrap for $500, repeat.
In that time, pray all you need are oil changes, tires, drum brake shoes or disc pads and maybe a battery.
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.
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.
Well, it really adds up. Compare it to a desk job where all you need is a cheap laptop.
$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.
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.
> 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.
More then double for the 48 contiguous states (“lower 48” only made sense as a synonym for 8½ months in 1949) plus DC.
> and it is the poverty line if you have the audacity to have two children.
If you have a family of four with one income, to be precise.
$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 “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.
Get out of your bubble. It's a living wage in a good chunk of Western Europe.
They'd be paid accordingly less if they lived in Europe. Though they'd probably have better safety nets.
$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'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.
Even in my ostensibly rich as sin European socialist country there are paupers. Minimum wage is something like 12 euro.
Got a problem with that?
But i had the capacity to choose to live in a town with a moderate cost of living and have few material needs or desires
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"
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.
1) What would you do with a million dollars?
Today's answer: Buy a house.
2) You've made it when you're making a six-figure salary.
That $100k salary today is closer to a $50k salary in the 1990s.
> Today's answer: Buy a house.
In my hometown, that would put a down payment on a house anywhere in the city limits. The town population is under 100k. -_-
Shit's changed so dramatically. Even a mere 15 years ago, you could get a townhome for under $150k.
Certainly explains why restaurants are closing for a lack of help - the help can't afford to live within 50+ miles from there.
https://www.anthropic.com/index/constitutional-ai-harmlessne...
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.
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.
- Having a partner
- Having a child
- Working from home
- The more roommates, the larger the apartment/house, the larger the costs
And that's a solution for only one of the things required to live.
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.