As part of Tesla’s plans to cut 10% of salaried staff, the company has laid off a significant number of its data annotation specialists. These specialists do a heavy job that is critical to empowering artificial intelligence systems to cope with complex tasks such as driving safely down a city street. The layoffs were first reported by Bloomberg on Tuesday and confirmed by CNN Business.
Data annotation specialists use software tools to manually label objects on video clips collected from Tesla vehicles. Specialists label common road objects such as lane lines, stop signs, traffic cones, sidewalks, and traffic signs. Labeled data is entered into the artificial intelligence system so that it learns to accurately perceive its environment. The more properly tagged data an AI system has, the better it can be.
Tesla has developed an automated way to do some of this labeling work in recent years, allowing the automaker to streamline its workforce.
“Without auto-labeling, I think we wouldn’t be able to solve the problem of autonomous driving,” Tesla CEO Elon Musk told the company’s AI Day in August 2021.
Tesla executives have suggested that automating data labeling has already accelerated their work on autonomous vehicles.
Ashok Elluswamy, director of Autopilot software, told the AI Day event that Tesla was able to collect 10,000 video clips from its cars and tag them automatically in a week. Clips are usually video segments of 45 to 60 seconds, as well as GPS and odometer-related data.
“That would have taken several months with humans labeling each clip,” Elluswamy said. He also described plans to produce a million self-tagged clips to “really crush this problem.”
According to Raj Rajkumar, a professor at Carnegie Mellon University who studies autonomous vehicles, there is no clear answer as to whether manual, human or automated data labeling is more accurate. Companies like Tesla may keep some humans involved in detecting defects in automated labeling, he said.
“If you make the economy less humane, it’s a financial victory,” Rajkumar said.
About five years ago, Tesla relied on an outside company to label its autonomous driving data, but in recent years it has moved efforts home, said last year Andrej Karpathy, who heads the AI at Tesla . Data labelers have worked in San Mateo, California, and Buffalo, New York. He described this as key to improving the quality of Tesla data. The company set up a team of more than 1,000 people, it told AI Day in 2021.
Job cuts and auto-labeling do not eliminate the need for human labor. In fact, Tesla continues to publicly publish some job postings for data annotation workers.
“For us it’s becoming much more of a story of,‘ How do humans and computers collaborate to really create these vector spatial data sets? ’Karpathy told AI Day.
Karpathy said Tesla wanted auto-labeling to be extremely accurate, which could have affected how quickly Tesla has turned around.
Artificial intelligence experts say there will be less need for human annotators in the future as techniques that do not require costly labor are developed.
“The future of data annotation is less,” Pedro Domingos, a professor of computer science at the University of Washington, told CNN Business. He cited the example of AI systems for language learning from text data masses.
Adella Petrescu, a former Tesla autopilot data annotation supervisor, posted on Tuesday on social media that she had been fired. Petrescu said he was promoted twice in a year and never had performance issues.
“For the past [two] I have worked for 50 years [plus] hours every week, many weekends and too many weeks with 16 hours a day and all for a purpose I really believed in, I still do: [a] better future for our next generations, ”Petrescu said.
Tesla did not immediately respond to a request for comment and is generally unrelated to professional media.