OpenAI is hiring contractors to train its code-generating models to think like human developers

OpenAI, the AI research firm behind ChatGPT, is reportedly hiring software developers on contract for data labeling and creating datasets on the thought process of writing code to train its code-generating AI models, according to a report in Semafor, which first reported Microsoft’s $10 billion investment plans for OpenAI. 

Some of the programmers who applied for the OpenAI’s job postings told Semafor on condition of anonymity that during their job interviews, OpenAI gave them coding tests in which they were asked to not only solve the problem but to explain the thought process behind it in English. The developers feel that OpenAI is trying to train its code generator model on a new data set which includes the process of writing codes explained by human developers. 

OpenAI’s code writing tool Codex, which also powers Github’s code writing assistant Co-Pilot, is based on a natural language generative model called GPT-3, which has been trained on both natural language and billions of lines of source code from publicly available sources such as public GitHub repositories. It can write codes in Python, JavaScript, Go, Perl, PHP, Ruby, Swift and TypeScript, and Shell based on commands in English. 

The report claims OpenAI has plans to hire over 1,000 developers for this project. Around 60% of these contractors will work on data labeling which will include creating massive sets of images, and audio clips for OpenAI other models, while 40% will be put to building datasets to train the code generator model. 

The company’s current headcount is 375, according to a Twitter post by CEO Sam Altman. Last week, Microsoft announced a multi-billion-dollar investment in OpenAI for a 49% stake amid mass layoffs by big tech companies. Microsoft CEO Satya Nadella said on January 18 that the Windows maker will let go of 10,000 workers by the third quarter of FY23. 

This isn't the first time OpenAI has used contractors to train its AI models. In a 2022 research paper published by OpenAI, the firm thanked all contractors for providing the data for training the models.