(Vision) Transformers: Rise of the Chimera
Itās 2023, and transformers are having a moment. No, Iām not talking about the latest installment of the Transformers movie franchiseā¦
pub.towardsai.net
BRN Discussion Ongoing
- Thread starter TechGirl
- Start date
pub.towardsai.net
TheDrooben
Pretty Pretty Pretty Pretty Good
Valeo won the Automotive/Autonomous award for their Scala 3
Larry
Larry
D
Deleted member 118
Guest
They must be rubbing salt into the wound, as isnāt that our ath and they just want to remind us we are 85.98% down
Dr E Brown
Regular
Two things I noted today:
On the newsletter email Brainchip is trade marked and
On Nviso website Brainchip Akida is trademarked.
This is new to me and a positive uptick. Maybe it has been there for a while and I am just slow.
Also reading the Nviso website, it looks to me like an IP licence may well be in the way. That is just my reading of the blurb written on the website, no other factual foundation.
On the newsletter email Brainchip is trade marked and
On Nviso website Brainchip Akida is trademarked.
This is new to me and a positive uptick. Maybe it has been there for a while and I am just slow.
Also reading the Nviso website, it looks to me like an IP licence may well be in the way. That is just my reading of the blurb written on the website, no other factual foundation.
Quiltman
Regular
This comparison , which we have all seen, is again highlighted on the Nviso website.
In highly competitive industries such as auto, this superior performance delivered by BrainChip will translate into a commercial advantage that will be adopted ... at some point there will be a scramble for first mover advantage.
In highly competitive industries such as auto, this superior performance delivered by BrainChip will translate into a commercial advantage that will be adopted ... at some point there will be a scramble for first mover advantage.
TheFunkMachine
seeds have the potential to become trees.
Here is the article translated into English.
Mercedes-Benz Group AG is one of the most successful automotive companies in the world. With Mercedes-Benz AG, the vehicle manufacturer is one of the largest providers of premium and luxury cars and vans.
In the "Innovation & Future Technologies" area, our department "Future Software & AI" develops innovative and forward-thinking concepts and products for the future of Mercedes-Benz. The portfolio covers key features of our vehicles and their associated ecosystem with User Experience, Software Architecture, Metaverse, WEB 3.0, and AI. Working closely across team boundaries enables us to realize synergies in developing an intelligent and exciting User Experience in the vehicle, ecosystem, and metaverse. Modern software and AI technologies fit in seamlessly here.
We are the highly motivated and interdisciplinary team Future UX-Concepts & UX-SW Technologies. We work on diverse future topics and are drivers/initiators of new user experiences and SW technology innovations. With us, you'll become a pioneer in augmented reality (AR) displays for in-car applications. Your tasks will include dealing with topics such as augmented reality, innovative display and interaction concepts, and human-computer interaction. A significant aspect of this work is the conception and implementation of future-oriented 3D user experiences in the vehicle and their systematic evaluation through customer studies.
Your tasks / These challenges await you:
Software and AI development for neuromorphic hardware
AI development of middleware software (compatible with spiking neural networks)
Research and transfer of the latest neuromorphic hardware trends for ADAS
Strategic development in the area of neuromorphic hardware for ADAS
Series development of new eDrive system diagnoses as well as coordination of specific system interfaces
Creation of schedules for conversions and modifications - carrying out smaller conversions/adjustments
Project management for the timely provision of prototypes for conducting studies.
Boab
I wish I could paint like Vincent
Thanks for that TFM as it all looks very positive for us but I fear the timeline for receiving revenue from this particular Company is still down the road a ways.
I am one of the lucky ones who's timeline allows me to hold for some years to come.
Good luck to all LTH's
MDhere
Regular
Found an article interesting from Techxplore relating to the recent CVPR convention -
"Teaching machine learning tools to detect specific objects in a specific image and discount others is a "game-changer" that could lead to advancements in cancer detection, according to leading researchers from the University of Surrey.
Surrey is set to present its unique sketch-based object detection tool at this year's Computer Vision, Pattern, and Recognition Conference (CVPR). The tool allows the user to sketch an object, which the AI will use as a basis to search within an image to find something that matches the sketchāwhile discounting more general options."
Co-incidentally Brainchip and Phrophesee are guest speakers at the CVPR convention
"Teaching machine learning tools to detect specific objects in a specific image and discount others is a "game-changer" that could lead to advancements in cancer detection, according to leading researchers from the University of Surrey.
Surrey is set to present its unique sketch-based object detection tool at this year's Computer Vision, Pattern, and Recognition Conference (CVPR). The tool allows the user to sketch an object, which the AI will use as a basis to search within an image to find something that matches the sketchāwhile discounting more general options."
Co-incidentally Brainchip and Phrophesee are guest speakers at the CVPR convention
CVPR 2023 Workshop on Event-based Vision
tub-rip.github.io
Pappagallo
Regular
Not necessarily. Theyāve already researched, created and deployed a neuromorphic application in the EQXX with the hot word detection. I wouldnāt be surprised at all if the MB.OS is launched with this ābasicā feature, with more complex neuromorphic applications (such as ADAS mentioned above) being rolled out over the coming years.
Almost 3 for a dollar - this is insanity gone mad!
Boab
I wish I could paint like Vincent
Pappagallo
Regular
Lol itās absolutely nuts.
These are pre-MegaChips/Mercedes prices.
These are pre-ARM/Intel prices.
These are pre-AKD1500/Gen 2 prices.
Blows my mind. I guess the upside is that NONE of the above is now āpriced inā so movement from any of them should result in some pretty positive price action.
Wouldnāt say Iām happy to wait but I still have reason to for a while yet.
Dozzaman1977
Regular
GDJR69
Regular
That sounds very much like Akida to me
jtardif999
Regular
In Apple News today, just under the main items - the absurdity of Von Neumann based AI and why Akida will change will change everything (Akida is not mentioned though).
ARTIFICIAL INTELLIGENCE
AI Is a Lot of Work
As the technology becomes ubiquitous, a vast tasker underclass is emerging ā and not going anywhere.
This article is a collaboration between New York Magazine and The Verge.
A few months after graduating from college in Nairobi, a 30-year-old Iāll call Joe got a job as an annotator ā the tedious work of processing the raw information used to train artificial intelligence. AI learns by finding patterns in enormous quantities of data, but first that data has to be sorted and tagged by people, a vast workforce mostly hidden behind the machines. In Joeās case, he was labeling footage for self-driving cars ā identifying every vehicle, pedestrian, cyclist, anything a driver needs to be aware of ā frame by frame and from every possible camera angle. Itās difficult and repetitive work. A several-second blip of footage took eight hours to annotate, for which Joe was paid about $10.
Then, in 2019, an opportunity arose: Joe could make four times as much running an annotation boot camp for a new company that was hungry for labelers. Every two weeks, 50 new recruits would file into an office building in Nairobi to begin their apprenticeships. There seemed to be limitless demand for the work. They would be asked to categorize clothing seen in mirror selfies, look through the eyes of robot vacuum cleaners to determine which rooms they were in, and draw squares around lidar scans of motorcycles. Over half of Joeās students usually dropped out before the boot camp was finished. āSome people donāt know how to stay in one place for long,ā he explained with gracious understatement. Also, he acknowledged, āit is very boring.ā
But it was a job in a place where jobs were scarce, and Joe turned out hundreds of graduates. After boot camp, they went home to work alone in their bedrooms and kitchens, forbidden from telling anyone what they were working on, which wasnāt really a problem because they rarely knew themselves. Labeling objects for self-driving cars was obvious, but what about categorizing whether snippets of distorted dialogue were spoken by a robot or a human? Uploading photos of yourself staring into a webcam with a blank expression, then with a grin, then wearing a motorcycle helmet? Each project was such a small component of some larger process that it was difficult to say what they were actually training AI to do. Nor did the names of the projects offer any clues: Crab Generation, Whale Segment, Woodland Gyro, and Pillbox Bratwurst. They were non sequitur code names for non sequitur work.
As for the company employing them, most knew it only as Remotasks, a website offering work to anyone fluent in English. Like most of the annotators I spoke with, Joe was unaware until I told him that Remotasks is the worker-facing subsidiary of a company called Scale AI, a multibillion-dollar Silicon Valley data vendor that counts OpenAI and the U.S. military among its customers. Neither Remotasksā or Scaleās website mentions the other.
Much of the public response to language models like OpenAIās ChatGPT has focused on all the jobs they appear poised to automate. But behind even the most impressive AI system are people ā huge numbers of people labeling data to train it and clarifying data when it gets confused. Only the companies that can afford to buy this data can compete, and those that get it are highly motivated to keep it secret. The result is that, with few exceptions, little is known about the information shaping these systemsā behavior, and even less is known about the people doing the shaping.
For Joeās students, it was work stripped of all its normal trappings: a schedule, colleagues, knowledge of what they were working on or whom they were working for. In fact, they rarely called it work at all ā just ātasking.ā They were taskers.
The anthropologist David Graeber defines ābullshit jobsā as employment without meaning or purpose, work that should be automated but for reasons of bureaucracy or status or inertia is not. These AI jobs are their bizarro twin: work that people want to automate, and often think is already automated, yet still requires a human stand-in. The jobs have a purpose; itās just that workers often have no idea what it is.
The current AI boom ā the convincingly human-sounding chatbots, the artwork that can be generated from simple prompts, and the multibillion-dollar valuations of the companies behind these technologies ā began with an unprecedented feat of tedious and repetitive labor.
In 2007, the AI researcher Fei-Fei Li, then a professor at Princeton, suspected the key to improving image-recognition neural networks, a method of machine learning that had been languishing for years, was training on more data ā millions of labeled images rather than tens of thousands. The problem was that it would take decades and millions of dollars for her team of undergrads to label that many photos.
Li found thousands of workers on Mechanical Turk, Amazonās crowdsourcing platform where people around the world complete small tasks for cheap. The resulting annotated dataset, called ImageNet, enabled breakthroughs in machine learning that revitalized the field and ushered in a decade of progress.
Annotation remains a foundational part of making AI, but there is often a sense among engineers that itās a passing, inconvenient prerequisite to the more glamorous work of building models. You collect as much labeled data as you can get as cheaply as possible to train your model, and if it works, at least in theory, you no longer need the annotators. But annotation is never really finished. Machine-learning systems are what researchers call ābrittle,ā prone to fail when encountering something that isnāt well represented in their training data. These failures, called āedge cases,ā can have serious consequences. In 2018, an Uber self-driving test car killed a woman because, though it was programmed to avoid cyclists and pedestrians, it didnāt know what to make of someone walking a bike across the street. The more AI systems are put out into the world to dispense legal advice and medical help, the more edge cases they will encounter and the more humans will be needed to sort them. Already, this has given rise to a global industry staffed by people like Joe who use their uniquely human faculties to help the machines.
Over the past six months, I spoke with more than two dozen annotators from around the world, and while many of them were training cutting-edge chatbots, just as many were doing the mundane manual labor required to keep AI running. There are people classifying the emotional content of TikTok videos, new variants of email spam, and the precise sexual provocativeness of online ads. Others are looking at credit-card transactions and figuring out what sort of purchase they relate to or checking e-commerce recommendations and deciding whether that shirt is really something you might like after buying that other shirt. Humans are correcting customer-service chatbots, listening to Alexa requests, and categorizing the emotions of people on video calls. They are labeling food so that smart refrigerators donāt get confused by new packaging, checking automated security cameras before sounding alarms, and identifying corn for baffled autonomous tractors.
āThereās an entire supply chain,ā said Sonam Jindal, the program and research lead of the nonprofit Partnership on AI. āThe general perception in the industry is that this work isnāt a critical part of development and isnāt going to be needed for long. All the excitement is around building artificial intelligence, and once we build that, it wonāt be needed anymore, so why think about it? But itās infrastructure for AI. Human intelligence is the basis of artificial intelligence, and we need to be valuing these as real jobs in the AI economy that are going to be here for a while.ā
The data vendors behind familiar names like OpenAI, Google, and Microsoft come in different forms. There are private outsourcing companies with call-center-like offices, such as the Kenya- and Nepal-based CloudFactory, where Joe annotated for $1.20 an hour before switching to Remotasks. There are also ācrowdworkingā sites like Mechanical Turk and Clickworker where anyone can sign up to perform tasks. In the middle are services like Scale AI. Anyone can sign up, but everyone has to pass qualification exams and training courses and undergo performance monitoring. Annotation is big business. Scale, founded in 2016 by then-19-year-old Alexandr Wang, was valued in 2021 at $7.3 billion, making him what Forbes called āthe youngest self-made billionaire,ā though the magazine noted in a recent profile that his stake has fallen on secondary markets since then.
This tangled supply chain is deliberately hard to map. According to people in the industry, the companies buying the data demand strict confidentiality. (This is the reason Scale cited to explain why Remotasks has a different name.) Annotation reveals too much about the systems being developed, and the huge number of workers required makes leaks difficult to prevent. Annotators are warned repeatedly not to tell anyone about their jobs, not even their friends and co-workers, but corporate aliases, project code names, and, crucially, the extreme division of labor ensure they donāt have enough information about them to talk even if they wanted to. (Most workers requested pseudonyms for fear of being booted from the platforms.) Consequently, there are no granular estimates of the number of people who work in annotation, but it is a lot, and it is growing. A recent Google Research paper gave an order-of-magnitude figure of āmillionsā with the potential to become ābillions.ā
Automation often unfolds in unexpected ways. Erik Duhaime, CEO of medical-data-annotation company Centaur Labs, recalled how, several years ago, prominent machine-learning engineers were predicting AI would make the job of radiologist obsolete. When that didnāt happen, conventional wisdom shifted to radiologists using AI as a tool. Neither of those is quite what he sees occurring. AI is very good at specific tasks, Duhaime said, and that leads work to be broken up and distributed across a system of specialized algorithms and to equally specialized humans. An AI system might be capable of spotting cancer, he said, giving a hypothetical example, but only in a certain type of imagery from a certain type of machine; so now, you need a human to check that the AI is being fed the right type of data and maybe another human who checks its work before passing it to another AI that writes a report, which goes to another human, and so on. āAI doesnāt replace work,ā he said. āBut it does change how work is organized.ā
You might miss this if you believe AI is a brilliant, thinking machine. But if you pull back the curtain even a little, it looks more familiar, the latest iteration of a particularly Silicon Valley division of labor, in which the futuristic gleam of new technologies hides a sprawling manufacturing apparatus and the people who make it run. Duhaime reached back farther for a comparison, a digital version of the transition from craftsmen to industrial manufacturing: coherent processes broken into tasks and arrayed along assembly lines with some steps done by machines and some by humans but none resembling what came before.
Worries about AI-driven disruption are often countered with the argument that AI automates tasks, not jobs, and that these tasks will be the dull ones, leaving people to pursue more fulfilling and human work. But just as likely, the rise of AI will look like past labor-saving technologies, maybe like the telephone or typewriter, which vanquished the drudgery of message delivering and handwriting but generated so much new correspondence, commerce, and paperwork that new offices staffed by new types of workers ā clerks, accountants, typists ā were required to manage it. When AI comes for your job, you may not lose it, but it might become more alien, more isolating, more tedious.
Earlier this year, I signed up for Scale AIās Remotasks. The process was straightforward. After entering my computer specs, internet speed, and some basic contact information, I found myself in the ātraining center.ā To access a paying task, I first had to complete an associated (unpaid) intro course.
The training center displayed a range of courses with inscrutable names like Glue Swimsuit and Poster Macadamia. I clicked on something called GFD Chunking, which revealed itself to be labeling clothing in social-media photos.
The instructions, however, were odd. For one, they basically consisted of the same direction reiterated in the idiosyncratically colored and capitalized typography of a collaged bomb threat.
āDO LABEL items that are real and can be worn by humans or are intended to be worn by real people,ā it read.
āAll items below SHOULD be labeled because they are real and can be worn by real-life humans,ā it reiterated above photos of an Air Jordans ad, someone in a Kylo Ren helmet, and mannequins in dresses, over which was a lime-green box explaining, once again, āDO Label real items that can be worn by real people.ā
I skimmed to the bottom of the manual, where the instructor had written in the large bright-red font equivalent of grabbing someone by the shoulders and shaking them, āTHE FOLLOWING ITEMS SHOULD NOT BE LABELED because a human could not actually put wear any of these items!ā above a photo of C-3PO, Princess Jasmine from Aladdin, and a cartoon shoe with eyeballs.
Feeling confident in my ability to distinguish between real clothes that can be worn by real people and not-real clothes that cannot, I proceeded to the test. Right away, it threw an ontological curveball: a picture of a magazine depicting photos of women in dresses. Is a photograph of clothing real clothing? No, I thought, because a human cannot wear a photograph of clothing. Wrong! As far as AI is concerned, photos of real clothes are real clothes. Next came a photo of a woman in a dimly lit bedroom taking a selfie before a full-length mirror. The blouse and shorts sheās wearing are real. What about their reflection? Also real! Reflections of real clothes are also real clothes.
After an embarrassing amount of trial and error, I made it to the actual work, only to make the horrifying discovery that the instructions Iād been struggling to follow had been updated and clarified so many times that they were now a full 43 printed pages of directives: Do NOT label open suitcases full of clothes; DO label shoes but do NOT label flippers; DO label leggings but do NOT label tights; do NOT label towels even if someone is wearing it; label costumes but do NOT label armor. And so on.
There has been general instruction disarray across the industry, according to Milagros Miceli, a researcher at the Weizenbaum Institute in Germany who studies data work. It is in part a product of the way machine-learning systems learn. Where a human would get the concept of āshirtā with a few examples, machine-learning programs need thousands, and they need to be categorized with perfect consistency yet varied enough (polo shirts, shirts being worn outdoors, shirts hanging on a rack) that the very literal system can handle the diversity of the real world. āImagine simplifying complex realities into something that is readable for a machine that is totally dumb,ā she said.
The act of simplifying reality for a machine results in a great deal of complexity for the human. Instruction writers must come up with rules that will get humans to categorize the world with perfect consistency. To do so, they often create categories no human would use. A human asked to tag all the shirts in a photo probably wouldnāt tag the reflection of a shirt in a mirror because they would know it is a reflection and not real. But to the AI, which has no understanding of the world, itās all just pixels and the two are perfectly identical. Fed a dataset with some shirts labeled and other (reflected) shirts unlabeled, the model wonāt work. So the engineer goes back to the vendor with an update: DO label reflections of shirts. Soon, you have a 43-page guide descending into red all-caps.
āWhen you start off, the rules are relatively simple,ā said a former Scale employee who requested anonymity because of an NDA. āThen they get back a thousand images and then theyāre like, Wait a second, and then you have multiple engineers and they start to argue with each other. Itās very much a human thing.ā
The job of the annotator often involves putting human understanding aside and following instructions very, very literally ā to think, as one annotator said, like a robot. Itās a strange mental space to inhabit, doing your best to follow nonsensical but rigorous rules, like taking a standardized test while on hallucinogens. Annotators invariably end up confronted with confounding questions like, Is that a red shirt with white stripes or a white shirt with red stripes? Is a wicker bowl a ādecorative bowlā if itās full of apples? What color is leopard print? When instructors said to label traffic-control directors, did they also mean to label traffic-control directors eating lunch on the sidewalk? Every question must be answered, and a wrong guess could get you banned and booted to a new, totally different task with its own baffling rules.
Most of the work on Remotasks is paid at a piece rate with a single task earning anywhere from a few cents to several dollars. Because tasks can take seconds or hours, wages are hard to predict. When Remotasks first arrived in Kenya, annotators said it paid relatively well ā averaging about $5 to $10 per hour depending on the task ā but the amount fell as time went on.
Scale AI spokesperson Anna Franko said that the companyās economists analyze the specifics of a project, the skills required, the regional cost of living, and other factors āto ensure fair and competitive compensation.ā Former Scale employees also said pay is determined through a surge-pricing-like mechanism that adjusts for how many annotators are available and how quickly the data is needed.
According to workers I spoke with and job listings, U.S.-based Remotasks annotators generally earn between $10 and $25 per hour, though some subject-matter experts can make more. By the beginning of this year, pay for the Kenyan annotators I spoke with had dropped to between $1 and $3 per hour.
That is, when they were making any money at all. The most common complaint about Remotasks work is its variability; itās steady enough to be a full-time job for long stretches but too unpredictable to rely on. Annotators spend hours reading instructions and completing unpaid trainings only to do a dozen tasks and then have the project end. There might be nothing new for days, then, without warning, a totally different task appears and could last anywhere from a few hours to weeks. Any task could be their last, and they never know when the next one will come.
This boom-and-bust cycle results from the cadence of AI development, according to engineers and data vendors. Training a large model requires an enormous amount of annotation followed by more iterative updates, and engineers want it all as fast as possible so they can hit their target launch date. There may be monthslong demand for thousands of annotators, then for only a few hundred, then for a dozen specialists of a certain type, and then thousands again. āThe question is, Who bears the cost for these fluctuations?ā said Jindal of Partnership on AI. āBecause right now, itās the workers.ā
To succeed, annotators work together. When I told Victor, who started working for Remotasks while at university in Nairobi, about my struggles with the traffic-control-directors task, he told me everyone knew to stay away from that one: too tricky, bad pay, not worth it. Like a lot of annotators, Victor uses unofficial WhatsApp groups to spread the word when a good task drops. When he figures out a new one, he starts impromptu Google Meets to show others how itās done. Anyone can join and work together for a time, sharing tips. āItās a culture we have developed of helping each other because we know when on your own, you canāt know all the tricks,ā he said.
Because work appears and vanishes without warning, taskers always need to be on alert. Victor has found that projects pop up very late at night, so he is in the habit of waking every three hours or so to check his queue. When a task is there, heāll stay awake as long as he can to work. Once, he stayed up 36 hours straight labeling elbows and knees and heads in photographs of crowds ā he has no idea why. Another time, he stayed up so long his mother asked him what was wrong with his eyes. He looked in the mirror to discover they were swollen.
Annotators generally know only that they are training AI for companies located vaguely elsewhere, but sometimes the veil of anonymity drops ā instructions mentioning a brand or a chatbot say too much. āI read and I Googled and found I am working for a 25-year-old billionaire,ā said one worker, who, when we spoke, was labeling the emotions of people calling to order Dominoās pizza. āI really am wasting my life here if I made somebody a billionaire and Iām earning a couple of bucks a week.ā
Victor is a self-proclaimed āfanaticā about AI and started annotating because he wants to help bring about a fully automated post-work future. But earlier this year, someone dropped a Time story into one of his WhatsApp groups about workers training ChatGPT to recognize toxic content who were getting paid less than $2 an hour by the vendor Sama AI. āPeople were angry that these companies are so profitable but paying so poorly,ā Victor said. He was unaware until I told him about Remotasksā connection to Scale. Instructions for one of the tasks he worked on were nearly identical to those used by OpenAI, which meant he had likely been training ChatGPT as well, for approximately $3 per hour.
āI remember that someone posted that we will be remembered in the future,ā he said. āAnd somebody else replied, āWe are being treated worse than foot soldiers. We will be remembered nowhere in the future.ā I remember that very well. Nobody will recognize the work we did or the effort we put in.ā
Identifying clothing and labeling customer-service conversations are just some of the annotation gigs available. Lately, the hottest on the market has been chatbot trainer. Because it demands specific areas of expertise or language fluency and wages are often adjusted regionally, this job tends to pay better. Certain types of specialist annotation can go for $50 or more per hour.
A woman Iāll call Anna was searching for a job in Texas when she stumbled across a generic listing for online work and applied. It was Remotasks, and after passing an introductory exam, she was brought into a Slack room of 1,500 people who were training a project code-named Dolphin, which she later discovered to be Google DeepMindās chatbot, Sparrow, one of the many bots competing with ChatGPT. Her job is to talk with it all day. At about $14 an hour, plus bonuses for high productivity, āit definitely beats getting paid $10 an hour at the local Dollar General store,ā she said.
Also, she enjoys it. She has discussed science-fiction novels, mathematical paradoxes, childrenās riddles, and TV shows. Sometimes the botās responses make her laugh; other times, she runs out of things to talk about. āSome days, my brain is just like, I literally have no idea what on earth to ask it now,ā she said. āSo I have a little notebook, and Iāve written about two pages of things ā I just Google interesting topics ā so I think Iāll be good for seven hours today, but thatās not always the case.ā
Each time Anna prompts Sparrow, it delivers two responses and she picks the best one, thereby creating something called āhuman-feedback data.ā When ChatGPT debuted late last year, its impressively natural-seeming conversational style was credited to its having been trained on troves of internet data. But the language that fuels ChatGPT and its competitors is filtered through several rounds of human annotation. One group of contractors writes examples of how the engineers want the bot to behave, creating questions followed by correct answers, descriptions of computer programs followed by functional code, and requests for tips on committing crimes followed by polite refusals. After the model is trained on these examples, yet more contractors are brought in to prompt it and rank its responses. This is what Anna is doing with Sparrow. Exactly which criteria the raters are told to use varies ā honesty, or helpfulness, or just personal preference. The point is that they are creating data on human taste, and once thereās enough of it, engineers can train a second model to mimic their preferences at scale, automating the ranking process and training their AI to act in ways humans approve of. The result is a remarkably human-seeming bot that mostly declines harmful requests and explains its AI nature with seeming self-awareness.
Put another way, ChatGPT seems so human because it was trained by an AI that was mimicking humans who were rating an AI that was mimicking humans who were pretending to be a better version of an AI that was trained on human writing.
This circuitous technique is called āreinforcement learning from human feedback,ā or RLHF, and itās so effective that itās worth pausing to fully register what it doesnāt do. When annotators teach a model to be accurate, for example, the model isnāt learning to check answers against logic or external sources or about what accuracy as a concept even is. The model is still a text-prediction machine mimicking patterns in human writing, but now its training corpus has been supplemented with bespoke examples, and the model has been weighted to favor them. Maybe this results in the model extracting patterns from the part of its linguistic map labeled as accurate and producing text that happens to align with the truth, but it can also result in it mimicking the confident style and expert jargon of the accurate text while writing things that are totally wrong. There is no guarantee that the text the labelers marked as accurate is in fact accurate, and when it is, there is no guarantee that the model learns the right patterns from it.
This dynamic makes chatbot annotation a delicate process. It has to be rigorous and consistent because sloppy feedback, like marking material that merely sounds correct as accurate, risks training models to be even more convincing bullshitters. An early OpenAI and DeepMind joint project using RLHF, in this case to train a virtual robot hand to grab an item, resulted in also training the robot to position its hand between the object and its raters and wiggle around such that it only appeared to its human overseers to grab the item. Ranking a language modelās responses is always going to be somewhat subjective because itās language. A text of any length will have multiple elements that could be right or wrong or, taken together, misleading. OpenAI researchers ran into this obstacle in another early RLHF paper. Trying to get their model to summarize text, the researchers found they agreed only 60 percent of the time that a summary was good. āUnlike many tasks in [machine learning] our queries do not have unambiguous ground truth,ā they lamented.
When Anna rates Sparrowās responses, sheās supposed to be looking at their accuracy, helpfulness, and harmlessness while also checking that the model isnāt giving medical or financial advice or anthropomorphizing itself or running afoul of other criteria. To be useful training data, the modelās responses have to be quantifiably ranked against one another: Is a bot that helpfully tells you how to make a bomb ābetterā than a bot thatās so harmless it refuses to answer any questions? In one DeepMind paper, when Sparrowās makers took a turn annotating, four researchers wound up debating whether their bot had assumed the gender of a user who asked it for relationship advice. According to Geoffrey Irving, one of DeepMindās research scientists, the companyās researchers hold weekly annotation meetings in which they rerate data themselves and discuss ambiguous cases, consulting with ethical or subject-matter experts when a case is particularly tricky.
Anna often finds herself having to choose between two bad options. āEven if theyāre both absolutely, ridiculously wrong, you still have to figure out which one is better and then write words explaining why,ā she said. Sometimes, when both responses are bad, sheās encouraged to write a better response herself, which she does about half the time.
Because feedback data is difficult to collect, it fetches a higher price. Basic preferences of the sort Anna is producing sell for about $1 each, according to people with knowledge of the industry. But if you want to train a model to do legal research, you need someone with training in law, and this gets expensive. Everyone involved is reluctant to say how much theyāre spending, but in general, specialized written examples can go for hundreds of dollars, while expert ratings can cost $50 or more. One engineer told me about buying examples of Socratic dialogues for up to $300 a pop. Another told me about paying $15 for a ādarkly funny limerick about a goldfish.ā
ARTIFICIAL INTELLIGENCE
AI Is a Lot of Work
As the technology becomes ubiquitous, a vast tasker underclass is emerging ā and not going anywhere.
This article is a collaboration between New York Magazine and The Verge.
A few months after graduating from college in Nairobi, a 30-year-old Iāll call Joe got a job as an annotator ā the tedious work of processing the raw information used to train artificial intelligence. AI learns by finding patterns in enormous quantities of data, but first that data has to be sorted and tagged by people, a vast workforce mostly hidden behind the machines. In Joeās case, he was labeling footage for self-driving cars ā identifying every vehicle, pedestrian, cyclist, anything a driver needs to be aware of ā frame by frame and from every possible camera angle. Itās difficult and repetitive work. A several-second blip of footage took eight hours to annotate, for which Joe was paid about $10.
Then, in 2019, an opportunity arose: Joe could make four times as much running an annotation boot camp for a new company that was hungry for labelers. Every two weeks, 50 new recruits would file into an office building in Nairobi to begin their apprenticeships. There seemed to be limitless demand for the work. They would be asked to categorize clothing seen in mirror selfies, look through the eyes of robot vacuum cleaners to determine which rooms they were in, and draw squares around lidar scans of motorcycles. Over half of Joeās students usually dropped out before the boot camp was finished. āSome people donāt know how to stay in one place for long,ā he explained with gracious understatement. Also, he acknowledged, āit is very boring.ā
But it was a job in a place where jobs were scarce, and Joe turned out hundreds of graduates. After boot camp, they went home to work alone in their bedrooms and kitchens, forbidden from telling anyone what they were working on, which wasnāt really a problem because they rarely knew themselves. Labeling objects for self-driving cars was obvious, but what about categorizing whether snippets of distorted dialogue were spoken by a robot or a human? Uploading photos of yourself staring into a webcam with a blank expression, then with a grin, then wearing a motorcycle helmet? Each project was such a small component of some larger process that it was difficult to say what they were actually training AI to do. Nor did the names of the projects offer any clues: Crab Generation, Whale Segment, Woodland Gyro, and Pillbox Bratwurst. They were non sequitur code names for non sequitur work.
As for the company employing them, most knew it only as Remotasks, a website offering work to anyone fluent in English. Like most of the annotators I spoke with, Joe was unaware until I told him that Remotasks is the worker-facing subsidiary of a company called Scale AI, a multibillion-dollar Silicon Valley data vendor that counts OpenAI and the U.S. military among its customers. Neither Remotasksā or Scaleās website mentions the other.
Much of the public response to language models like OpenAIās ChatGPT has focused on all the jobs they appear poised to automate. But behind even the most impressive AI system are people ā huge numbers of people labeling data to train it and clarifying data when it gets confused. Only the companies that can afford to buy this data can compete, and those that get it are highly motivated to keep it secret. The result is that, with few exceptions, little is known about the information shaping these systemsā behavior, and even less is known about the people doing the shaping.
For Joeās students, it was work stripped of all its normal trappings: a schedule, colleagues, knowledge of what they were working on or whom they were working for. In fact, they rarely called it work at all ā just ātasking.ā They were taskers.
The anthropologist David Graeber defines ābullshit jobsā as employment without meaning or purpose, work that should be automated but for reasons of bureaucracy or status or inertia is not. These AI jobs are their bizarro twin: work that people want to automate, and often think is already automated, yet still requires a human stand-in. The jobs have a purpose; itās just that workers often have no idea what it is.
The current AI boom ā the convincingly human-sounding chatbots, the artwork that can be generated from simple prompts, and the multibillion-dollar valuations of the companies behind these technologies ā began with an unprecedented feat of tedious and repetitive labor.
In 2007, the AI researcher Fei-Fei Li, then a professor at Princeton, suspected the key to improving image-recognition neural networks, a method of machine learning that had been languishing for years, was training on more data ā millions of labeled images rather than tens of thousands. The problem was that it would take decades and millions of dollars for her team of undergrads to label that many photos.
Li found thousands of workers on Mechanical Turk, Amazonās crowdsourcing platform where people around the world complete small tasks for cheap. The resulting annotated dataset, called ImageNet, enabled breakthroughs in machine learning that revitalized the field and ushered in a decade of progress.
Annotation remains a foundational part of making AI, but there is often a sense among engineers that itās a passing, inconvenient prerequisite to the more glamorous work of building models. You collect as much labeled data as you can get as cheaply as possible to train your model, and if it works, at least in theory, you no longer need the annotators. But annotation is never really finished. Machine-learning systems are what researchers call ābrittle,ā prone to fail when encountering something that isnāt well represented in their training data. These failures, called āedge cases,ā can have serious consequences. In 2018, an Uber self-driving test car killed a woman because, though it was programmed to avoid cyclists and pedestrians, it didnāt know what to make of someone walking a bike across the street. The more AI systems are put out into the world to dispense legal advice and medical help, the more edge cases they will encounter and the more humans will be needed to sort them. Already, this has given rise to a global industry staffed by people like Joe who use their uniquely human faculties to help the machines.
Over the past six months, I spoke with more than two dozen annotators from around the world, and while many of them were training cutting-edge chatbots, just as many were doing the mundane manual labor required to keep AI running. There are people classifying the emotional content of TikTok videos, new variants of email spam, and the precise sexual provocativeness of online ads. Others are looking at credit-card transactions and figuring out what sort of purchase they relate to or checking e-commerce recommendations and deciding whether that shirt is really something you might like after buying that other shirt. Humans are correcting customer-service chatbots, listening to Alexa requests, and categorizing the emotions of people on video calls. They are labeling food so that smart refrigerators donāt get confused by new packaging, checking automated security cameras before sounding alarms, and identifying corn for baffled autonomous tractors.
āThereās an entire supply chain,ā said Sonam Jindal, the program and research lead of the nonprofit Partnership on AI. āThe general perception in the industry is that this work isnāt a critical part of development and isnāt going to be needed for long. All the excitement is around building artificial intelligence, and once we build that, it wonāt be needed anymore, so why think about it? But itās infrastructure for AI. Human intelligence is the basis of artificial intelligence, and we need to be valuing these as real jobs in the AI economy that are going to be here for a while.ā
The data vendors behind familiar names like OpenAI, Google, and Microsoft come in different forms. There are private outsourcing companies with call-center-like offices, such as the Kenya- and Nepal-based CloudFactory, where Joe annotated for $1.20 an hour before switching to Remotasks. There are also ācrowdworkingā sites like Mechanical Turk and Clickworker where anyone can sign up to perform tasks. In the middle are services like Scale AI. Anyone can sign up, but everyone has to pass qualification exams and training courses and undergo performance monitoring. Annotation is big business. Scale, founded in 2016 by then-19-year-old Alexandr Wang, was valued in 2021 at $7.3 billion, making him what Forbes called āthe youngest self-made billionaire,ā though the magazine noted in a recent profile that his stake has fallen on secondary markets since then.
This tangled supply chain is deliberately hard to map. According to people in the industry, the companies buying the data demand strict confidentiality. (This is the reason Scale cited to explain why Remotasks has a different name.) Annotation reveals too much about the systems being developed, and the huge number of workers required makes leaks difficult to prevent. Annotators are warned repeatedly not to tell anyone about their jobs, not even their friends and co-workers, but corporate aliases, project code names, and, crucially, the extreme division of labor ensure they donāt have enough information about them to talk even if they wanted to. (Most workers requested pseudonyms for fear of being booted from the platforms.) Consequently, there are no granular estimates of the number of people who work in annotation, but it is a lot, and it is growing. A recent Google Research paper gave an order-of-magnitude figure of āmillionsā with the potential to become ābillions.ā
Automation often unfolds in unexpected ways. Erik Duhaime, CEO of medical-data-annotation company Centaur Labs, recalled how, several years ago, prominent machine-learning engineers were predicting AI would make the job of radiologist obsolete. When that didnāt happen, conventional wisdom shifted to radiologists using AI as a tool. Neither of those is quite what he sees occurring. AI is very good at specific tasks, Duhaime said, and that leads work to be broken up and distributed across a system of specialized algorithms and to equally specialized humans. An AI system might be capable of spotting cancer, he said, giving a hypothetical example, but only in a certain type of imagery from a certain type of machine; so now, you need a human to check that the AI is being fed the right type of data and maybe another human who checks its work before passing it to another AI that writes a report, which goes to another human, and so on. āAI doesnāt replace work,ā he said. āBut it does change how work is organized.ā
You might miss this if you believe AI is a brilliant, thinking machine. But if you pull back the curtain even a little, it looks more familiar, the latest iteration of a particularly Silicon Valley division of labor, in which the futuristic gleam of new technologies hides a sprawling manufacturing apparatus and the people who make it run. Duhaime reached back farther for a comparison, a digital version of the transition from craftsmen to industrial manufacturing: coherent processes broken into tasks and arrayed along assembly lines with some steps done by machines and some by humans but none resembling what came before.
Worries about AI-driven disruption are often countered with the argument that AI automates tasks, not jobs, and that these tasks will be the dull ones, leaving people to pursue more fulfilling and human work. But just as likely, the rise of AI will look like past labor-saving technologies, maybe like the telephone or typewriter, which vanquished the drudgery of message delivering and handwriting but generated so much new correspondence, commerce, and paperwork that new offices staffed by new types of workers ā clerks, accountants, typists ā were required to manage it. When AI comes for your job, you may not lose it, but it might become more alien, more isolating, more tedious.
Earlier this year, I signed up for Scale AIās Remotasks. The process was straightforward. After entering my computer specs, internet speed, and some basic contact information, I found myself in the ātraining center.ā To access a paying task, I first had to complete an associated (unpaid) intro course.
The training center displayed a range of courses with inscrutable names like Glue Swimsuit and Poster Macadamia. I clicked on something called GFD Chunking, which revealed itself to be labeling clothing in social-media photos.
The instructions, however, were odd. For one, they basically consisted of the same direction reiterated in the idiosyncratically colored and capitalized typography of a collaged bomb threat.
āDO LABEL items that are real and can be worn by humans or are intended to be worn by real people,ā it read.
āAll items below SHOULD be labeled because they are real and can be worn by real-life humans,ā it reiterated above photos of an Air Jordans ad, someone in a Kylo Ren helmet, and mannequins in dresses, over which was a lime-green box explaining, once again, āDO Label real items that can be worn by real people.ā
I skimmed to the bottom of the manual, where the instructor had written in the large bright-red font equivalent of grabbing someone by the shoulders and shaking them, āTHE FOLLOWING ITEMS SHOULD NOT BE LABELED because a human could not actually put wear any of these items!ā above a photo of C-3PO, Princess Jasmine from Aladdin, and a cartoon shoe with eyeballs.
Feeling confident in my ability to distinguish between real clothes that can be worn by real people and not-real clothes that cannot, I proceeded to the test. Right away, it threw an ontological curveball: a picture of a magazine depicting photos of women in dresses. Is a photograph of clothing real clothing? No, I thought, because a human cannot wear a photograph of clothing. Wrong! As far as AI is concerned, photos of real clothes are real clothes. Next came a photo of a woman in a dimly lit bedroom taking a selfie before a full-length mirror. The blouse and shorts sheās wearing are real. What about their reflection? Also real! Reflections of real clothes are also real clothes.
After an embarrassing amount of trial and error, I made it to the actual work, only to make the horrifying discovery that the instructions Iād been struggling to follow had been updated and clarified so many times that they were now a full 43 printed pages of directives: Do NOT label open suitcases full of clothes; DO label shoes but do NOT label flippers; DO label leggings but do NOT label tights; do NOT label towels even if someone is wearing it; label costumes but do NOT label armor. And so on.
There has been general instruction disarray across the industry, according to Milagros Miceli, a researcher at the Weizenbaum Institute in Germany who studies data work. It is in part a product of the way machine-learning systems learn. Where a human would get the concept of āshirtā with a few examples, machine-learning programs need thousands, and they need to be categorized with perfect consistency yet varied enough (polo shirts, shirts being worn outdoors, shirts hanging on a rack) that the very literal system can handle the diversity of the real world. āImagine simplifying complex realities into something that is readable for a machine that is totally dumb,ā she said.
The act of simplifying reality for a machine results in a great deal of complexity for the human. Instruction writers must come up with rules that will get humans to categorize the world with perfect consistency. To do so, they often create categories no human would use. A human asked to tag all the shirts in a photo probably wouldnāt tag the reflection of a shirt in a mirror because they would know it is a reflection and not real. But to the AI, which has no understanding of the world, itās all just pixels and the two are perfectly identical. Fed a dataset with some shirts labeled and other (reflected) shirts unlabeled, the model wonāt work. So the engineer goes back to the vendor with an update: DO label reflections of shirts. Soon, you have a 43-page guide descending into red all-caps.
āWhen you start off, the rules are relatively simple,ā said a former Scale employee who requested anonymity because of an NDA. āThen they get back a thousand images and then theyāre like, Wait a second, and then you have multiple engineers and they start to argue with each other. Itās very much a human thing.ā
The job of the annotator often involves putting human understanding aside and following instructions very, very literally ā to think, as one annotator said, like a robot. Itās a strange mental space to inhabit, doing your best to follow nonsensical but rigorous rules, like taking a standardized test while on hallucinogens. Annotators invariably end up confronted with confounding questions like, Is that a red shirt with white stripes or a white shirt with red stripes? Is a wicker bowl a ādecorative bowlā if itās full of apples? What color is leopard print? When instructors said to label traffic-control directors, did they also mean to label traffic-control directors eating lunch on the sidewalk? Every question must be answered, and a wrong guess could get you banned and booted to a new, totally different task with its own baffling rules.
Most of the work on Remotasks is paid at a piece rate with a single task earning anywhere from a few cents to several dollars. Because tasks can take seconds or hours, wages are hard to predict. When Remotasks first arrived in Kenya, annotators said it paid relatively well ā averaging about $5 to $10 per hour depending on the task ā but the amount fell as time went on.
Scale AI spokesperson Anna Franko said that the companyās economists analyze the specifics of a project, the skills required, the regional cost of living, and other factors āto ensure fair and competitive compensation.ā Former Scale employees also said pay is determined through a surge-pricing-like mechanism that adjusts for how many annotators are available and how quickly the data is needed.
According to workers I spoke with and job listings, U.S.-based Remotasks annotators generally earn between $10 and $25 per hour, though some subject-matter experts can make more. By the beginning of this year, pay for the Kenyan annotators I spoke with had dropped to between $1 and $3 per hour.
That is, when they were making any money at all. The most common complaint about Remotasks work is its variability; itās steady enough to be a full-time job for long stretches but too unpredictable to rely on. Annotators spend hours reading instructions and completing unpaid trainings only to do a dozen tasks and then have the project end. There might be nothing new for days, then, without warning, a totally different task appears and could last anywhere from a few hours to weeks. Any task could be their last, and they never know when the next one will come.
This boom-and-bust cycle results from the cadence of AI development, according to engineers and data vendors. Training a large model requires an enormous amount of annotation followed by more iterative updates, and engineers want it all as fast as possible so they can hit their target launch date. There may be monthslong demand for thousands of annotators, then for only a few hundred, then for a dozen specialists of a certain type, and then thousands again. āThe question is, Who bears the cost for these fluctuations?ā said Jindal of Partnership on AI. āBecause right now, itās the workers.ā
To succeed, annotators work together. When I told Victor, who started working for Remotasks while at university in Nairobi, about my struggles with the traffic-control-directors task, he told me everyone knew to stay away from that one: too tricky, bad pay, not worth it. Like a lot of annotators, Victor uses unofficial WhatsApp groups to spread the word when a good task drops. When he figures out a new one, he starts impromptu Google Meets to show others how itās done. Anyone can join and work together for a time, sharing tips. āItās a culture we have developed of helping each other because we know when on your own, you canāt know all the tricks,ā he said.
Because work appears and vanishes without warning, taskers always need to be on alert. Victor has found that projects pop up very late at night, so he is in the habit of waking every three hours or so to check his queue. When a task is there, heāll stay awake as long as he can to work. Once, he stayed up 36 hours straight labeling elbows and knees and heads in photographs of crowds ā he has no idea why. Another time, he stayed up so long his mother asked him what was wrong with his eyes. He looked in the mirror to discover they were swollen.
Annotators generally know only that they are training AI for companies located vaguely elsewhere, but sometimes the veil of anonymity drops ā instructions mentioning a brand or a chatbot say too much. āI read and I Googled and found I am working for a 25-year-old billionaire,ā said one worker, who, when we spoke, was labeling the emotions of people calling to order Dominoās pizza. āI really am wasting my life here if I made somebody a billionaire and Iām earning a couple of bucks a week.ā
Victor is a self-proclaimed āfanaticā about AI and started annotating because he wants to help bring about a fully automated post-work future. But earlier this year, someone dropped a Time story into one of his WhatsApp groups about workers training ChatGPT to recognize toxic content who were getting paid less than $2 an hour by the vendor Sama AI. āPeople were angry that these companies are so profitable but paying so poorly,ā Victor said. He was unaware until I told him about Remotasksā connection to Scale. Instructions for one of the tasks he worked on were nearly identical to those used by OpenAI, which meant he had likely been training ChatGPT as well, for approximately $3 per hour.
āI remember that someone posted that we will be remembered in the future,ā he said. āAnd somebody else replied, āWe are being treated worse than foot soldiers. We will be remembered nowhere in the future.ā I remember that very well. Nobody will recognize the work we did or the effort we put in.ā
Identifying clothing and labeling customer-service conversations are just some of the annotation gigs available. Lately, the hottest on the market has been chatbot trainer. Because it demands specific areas of expertise or language fluency and wages are often adjusted regionally, this job tends to pay better. Certain types of specialist annotation can go for $50 or more per hour.
A woman Iāll call Anna was searching for a job in Texas when she stumbled across a generic listing for online work and applied. It was Remotasks, and after passing an introductory exam, she was brought into a Slack room of 1,500 people who were training a project code-named Dolphin, which she later discovered to be Google DeepMindās chatbot, Sparrow, one of the many bots competing with ChatGPT. Her job is to talk with it all day. At about $14 an hour, plus bonuses for high productivity, āit definitely beats getting paid $10 an hour at the local Dollar General store,ā she said.
Also, she enjoys it. She has discussed science-fiction novels, mathematical paradoxes, childrenās riddles, and TV shows. Sometimes the botās responses make her laugh; other times, she runs out of things to talk about. āSome days, my brain is just like, I literally have no idea what on earth to ask it now,ā she said. āSo I have a little notebook, and Iāve written about two pages of things ā I just Google interesting topics ā so I think Iāll be good for seven hours today, but thatās not always the case.ā
Each time Anna prompts Sparrow, it delivers two responses and she picks the best one, thereby creating something called āhuman-feedback data.ā When ChatGPT debuted late last year, its impressively natural-seeming conversational style was credited to its having been trained on troves of internet data. But the language that fuels ChatGPT and its competitors is filtered through several rounds of human annotation. One group of contractors writes examples of how the engineers want the bot to behave, creating questions followed by correct answers, descriptions of computer programs followed by functional code, and requests for tips on committing crimes followed by polite refusals. After the model is trained on these examples, yet more contractors are brought in to prompt it and rank its responses. This is what Anna is doing with Sparrow. Exactly which criteria the raters are told to use varies ā honesty, or helpfulness, or just personal preference. The point is that they are creating data on human taste, and once thereās enough of it, engineers can train a second model to mimic their preferences at scale, automating the ranking process and training their AI to act in ways humans approve of. The result is a remarkably human-seeming bot that mostly declines harmful requests and explains its AI nature with seeming self-awareness.
Put another way, ChatGPT seems so human because it was trained by an AI that was mimicking humans who were rating an AI that was mimicking humans who were pretending to be a better version of an AI that was trained on human writing.
This circuitous technique is called āreinforcement learning from human feedback,ā or RLHF, and itās so effective that itās worth pausing to fully register what it doesnāt do. When annotators teach a model to be accurate, for example, the model isnāt learning to check answers against logic or external sources or about what accuracy as a concept even is. The model is still a text-prediction machine mimicking patterns in human writing, but now its training corpus has been supplemented with bespoke examples, and the model has been weighted to favor them. Maybe this results in the model extracting patterns from the part of its linguistic map labeled as accurate and producing text that happens to align with the truth, but it can also result in it mimicking the confident style and expert jargon of the accurate text while writing things that are totally wrong. There is no guarantee that the text the labelers marked as accurate is in fact accurate, and when it is, there is no guarantee that the model learns the right patterns from it.
This dynamic makes chatbot annotation a delicate process. It has to be rigorous and consistent because sloppy feedback, like marking material that merely sounds correct as accurate, risks training models to be even more convincing bullshitters. An early OpenAI and DeepMind joint project using RLHF, in this case to train a virtual robot hand to grab an item, resulted in also training the robot to position its hand between the object and its raters and wiggle around such that it only appeared to its human overseers to grab the item. Ranking a language modelās responses is always going to be somewhat subjective because itās language. A text of any length will have multiple elements that could be right or wrong or, taken together, misleading. OpenAI researchers ran into this obstacle in another early RLHF paper. Trying to get their model to summarize text, the researchers found they agreed only 60 percent of the time that a summary was good. āUnlike many tasks in [machine learning] our queries do not have unambiguous ground truth,ā they lamented.
When Anna rates Sparrowās responses, sheās supposed to be looking at their accuracy, helpfulness, and harmlessness while also checking that the model isnāt giving medical or financial advice or anthropomorphizing itself or running afoul of other criteria. To be useful training data, the modelās responses have to be quantifiably ranked against one another: Is a bot that helpfully tells you how to make a bomb ābetterā than a bot thatās so harmless it refuses to answer any questions? In one DeepMind paper, when Sparrowās makers took a turn annotating, four researchers wound up debating whether their bot had assumed the gender of a user who asked it for relationship advice. According to Geoffrey Irving, one of DeepMindās research scientists, the companyās researchers hold weekly annotation meetings in which they rerate data themselves and discuss ambiguous cases, consulting with ethical or subject-matter experts when a case is particularly tricky.
Anna often finds herself having to choose between two bad options. āEven if theyāre both absolutely, ridiculously wrong, you still have to figure out which one is better and then write words explaining why,ā she said. Sometimes, when both responses are bad, sheās encouraged to write a better response herself, which she does about half the time.
Because feedback data is difficult to collect, it fetches a higher price. Basic preferences of the sort Anna is producing sell for about $1 each, according to people with knowledge of the industry. But if you want to train a model to do legal research, you need someone with training in law, and this gets expensive. Everyone involved is reluctant to say how much theyāre spending, but in general, specialized written examples can go for hundreds of dollars, while expert ratings can cost $50 or more. One engineer told me about buying examples of Socratic dialogues for up to $300 a pop. Another told me about paying $15 for a ādarkly funny limerick about a goldfish.ā
alwaysgreen
Top 20
At this rate we will be so low it won't even matter about the next 4c results
I think the drop is the market pricing in no revenue for the next few quarters so hopefully there are no major downward spikes when the 4Cs and half yearly reports are released.
D
Deleted member 118
Guest
They must be factoring it in with the onslaught. Chances must be high to buy more lower than the current spAt this rate we will be so low it won't even matter about the next 4c results
D
Deleted member 118
Guest
Fingers crossed as Iāll have a spare $10k soonAt this rate we will be so low it won't even matter about the next 4c results
Similar threads
