How LinkedIn launched an AI tool based on ChatGPT in three months

Xu Ya, vice president of engineering and head of data and artificial intelligence business at LinkedIn, said that considering that the engineering and product teams are based on OpenAI’s latest GPT model (including ChatGPT and GPT-4) as well as some open source models implemented many changes, a timeline that is unprecedented for a large company like LinkedIn. The tool includes features such as generative AI collaborative articles, personalized writing suggestions for job descriptions and LinkedIn profiles.

Xu Ya explained that the development team she led was able to automatically generate job descriptions and serve real-time traffic in just one month. A cross-functional team with a common goal was key. She added: "This does not mean that we work 20 hours a day or get off work late, but put down other things and focus on completing important work."

Xu Ya said that since LinkedIn is owned by Microsoft subsidiary, she did see the future of this technology in advance. Last fall, she, along with LinkedIn CEO Ryan Roslansky and other colleagues, quickly began envisioning how ChatGPT and other GPT models could create more application and service opportunities for LinkedIn members and customers.

LinkedIn company prioritizes engineering philosophy

Xu Ya said that her team’s early priority engineering philosophy was “rooted in exploration rather than building mature end products.” She explained that maturation of appropriate features and experiences will happen over time, but by putting generative AI technology in the hands of every interested engineer and product manager, such exploration will be encouraged.

By creating the LinkedIn Gateway, you can access OpenAI models and open source models from Hugging Face, and provide LinkedIn’s Generative AI Playground, which enables engineers to explore using advanced generative AI models from OpenAI companies and other sources. LinkedIn data facilitates this exploration. The company also convened thousands of engineers to participate in LinkedIn's largest-ever internal hackathon.

In addition, people at LinkedIn need to better understand how large language models work, including how to do just-in-time engineering, and what potential problems and limitations the models have.

Xu Ya said: "We provide different levels of education, such as company meetings, lunch and learn courses, as well as deeper education for those who are more deeply involved in artificial intelligence development and research and development."

Collaboration is also an important part of integrating and supporting generative AI. She said, “Because of our collaborative culture, we encourage different teams to share resources. This allows them to develop quickly in situations where the number of developers with access to certain generative AI models is limited due to capacity. We work within teams on and pass on their experiences about quotas, access, incentive models, and other best practices so that they can better help each other."

Run fast, but run together

Xu Ya It was also emphasized that LinkedIn is aware that in the process of generative artificial intelligence, there are some areas that need to be completed intensively. She explained that while there's always some tension between running fast and running together, the company tries to maintain those checks and balances, especially when it comes to responsible AI. “While this may slow down the team, we need to be thoughtful about it,” she said.

For example, the company publishes AI-generated articles through an evaluation pipeline, with iterative output that is reviewed by humans and changes They work on the fly until you get a satisfactory score. Xu Ya explained that LinkedIn very carefully considers what risks are tolerable and what are intolerable risks. The company has no tolerance for objectionable content and has a low tolerance for gray area content. It relies on staff to flag content for removal.

She added that she hopes to avoid any bad and damaging information and only allow safe and informative content to be generated and published. For example, she noted that Kevin Roose's recent New York Times article included a transcript of his chat with Microsoft's Bing chatbot. The article points out that it would be concerning if someone shared a guide on how to make a bomb, but if someone gave poor advice on how to complete the task in a chat with the chatbot, or in Roose's case commented on his Marriage, then would be less of a concern.

Xu Ya said, "This technology cannot only exist in the laboratory, but must be put into practical applications. In this way, people can make full use of it in ways that were never expected in the laboratory, but it needs to be Make sure you have the right processes in place." She cited Microsoft Chief Technology Officer Kevin Scott's recent comments on the topic.

The above is the detailed content of How LinkedIn launched an AI tool based on ChatGPT in three months. For more information, please follow other related articles on the PHP Chinese website!