Contractor Fires Dozens of OpenAI Data Trainers

AI For Business


San Francisco-based tech contractor Invisible Technologies laid off dozens of contractors training language models for OpenAI in March.
iStock/Getty Images Plus

  • Invisible Technologies fired 31 contractors hired to train OpenAI’s GPT, Insider has learned.
  • Contractors were hired to improve the AI ​​model’s capabilities such as creative writing and coding.
  • As ChatGPT AI grew in popularity, OpenAI reportedly hired around 1,000 contractors worldwide.

Dozens of contractors who helped train the OpenAI language model that powers ChatGPT were laid off in March, according to people familiar with the matter and documentation of internal communications.

San Francisco-based company Invisible Technologies laid off 31 contractors as of March 16, according to an internal Slack screenshot obtained by an insider. Downsizing is OpenAI’s ChatGPT taking the world by storm, with users flocking to bots in hopes of making their lives easier. OpenAI continues to hire across its business.

Hundreds of Invisible contractors, known as “advanced AI data trainers,” are working with OpenAI to train GPT bots, an internal Slack screenshot shows. Invisible’s AI data trainers are responsible for tasks such as improving model coding skills, improving creative writing abilities, and training them not to say certain things, said an Invisible contractor familiar with the matter. said he requested anonymity because he signed a non-disclosure agreement. – Disclosure Agreement. Insiders have confirmed his identity and employment.

Kamron Palizban, Invisible’s vice president of operations, spoke about the layoffs at an all-staff meeting in March. He said OpenAI wanted to reduce the ranks of its contractors due to changing business needs, according to a transcript of the meeting obtained by an insider. Many of the fired contractors were working on projects that didn’t give OpenAI a good return on investment, he said at the conference.

OpenAI and Invisible Technologies did not respond to requests for comment.

OpenAI cuts some contractors after hiring 1,000 people worldwide

Invisible’s relationship with OpenAI offers a glimpse into the data training practices of the ChatGPT creators.

The adjustment to OpenAI’s contract with Invisible follows a six-month headcount increase first reported by Semafor. As of January, OpenAI has employed nearly 1,000 data labeling contractors across Eastern Europe and Latin America, a source familiar with the matter told Semafor.

Invisible’s layoffs come just two months after Microsoft poured $10 billion into OpenAI. But Invisible isn’t the only contractor using OpenAI.

In February 2022, another San Francisco-based contractor, Sama, found that Kenyan data labelers were reviewing harmful content such as sexual abuse, hate speech, and violence, according to a Time investigation. We have ended our partnership with OpenAI.

In a statement to Time, an OpenAI spokesperson said: [text and images] This is a necessary step to minimize the amount of violent and sexual content in our training data and to create tools that can detect harmful content. ”

A day in the life of an AI Trainer

According to the unseen contractor, a data trainer’s most basic duties include reviewing conversations between AIs and their users to identify potentially illegal, personal, offensive, or error-ridden messages. includes identifying. A contractor who spoke with an insider described the routine as follows:

They start their shift by opening their internal work browser and reviewing their team’s task list. For example, click a task like “Disable browsing and have a conversation about a random topic” and enter your query in the message box.

When a query is submitted, the model generates four responses. The contractor rates each response by opening her drop-down menu and selecting the types of errors present, including factual inaccuracies, spelling, grammar, and harassment. It then ranks the severity of the error on a scale of 1 to 7, according to the demo the contractor provided to her Insider. 7 indicates a “basically perfect” answer.

The contractor must then craft a perfect answer and submit it to complete the task. The results will be sent to OpenAI and Invisible quality checkers, the contractor said. This cycle repeats for each subsequent task.

Referring to OpenAI at the conference, Palizban said, “They are at a stage where they are more clear about where they are going.”

Grace Matelich, partner and operations manager at Invisible, said in a taped conference that the company fired contractors based on performance metrics like “quality” and “throughput.”

Some underperforming contractors, and those who were onboarded but “did not meet the criteria for certification,” were fired, but were often moved to another OpenAI team, according to the conference. was given the option to “If you’re still here, please know that it’s because we trust your ability to operate with excellence.

If you work for OpenAI or have a story to share, reach out to this creator at amok@insider.com.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *