eBay Tech Blog Explores AI Responsibility & How LLMs Improve Dev Productivity
eBay's official corporate tech blog released two new articles this week exploring the company's commitment to using AI responsibly and how these emerging generative technologies are enhancing employee productivity.
Senior Director and Distinguished Applied Scientist, Responsible AI Lauren Wilcox shared the five guiding principles eBay has adopted for the responsible use and development of AI.
Inclusivity, Equity, and Fairness
eBay strives to enable equitable and fair AI experiences
Building with an inclusivity mindset moves us closer to achieving equity and fairness. We strive to ensure our AI systems represent a diverse range of human cultures, backgrounds, and experiences.Inclusivity in AI systems means implementing appropriate measures to avoid excluding certain groups or people from being able to engage with or to benefit from the AI system.
Key checks and balances help ensure that an AI system’s decisions don’t unfairly discriminate against people, and we strive to ensure that datasets we use meet the highest bars for quality and safety.
Accountability and Lawfulness
Everyone in the eBay community has a role to play in responsible AI
We take steps to ensure that AI systems function properly, throughout their lifecycle, and that their design, development, testing, and use all comply with regulatory frameworks.We actively ensure there is human oversight throughout the development, launch, and monitoring of AI uses, and have implemented risk-based decision-making processes.
Reliability, Safety, and Security
Build trustworthy AI systems that are reliable, safe and secure
At eBay, we rigorously test AI technologies to prevent unauthorized access and malicious use, and employ the most protective measures possible for any confidential data that are used to train or operate AI systems and models.Our goal is to ensure that our AI systems are designed to perform as they were originally intended, to respond securely in various conditions of use, do not pose safety risks in unexpected situations or adverse conditions, and to be secure and resilient, in accordance with the highest standards in information security.
Privacy by Design
Build AI systems with privacy at the forefront
To the extent that personal information is involved in any part of our technology stack, we apply eBay’s Privacy principles – which are the backbone of eBay’s Privacy program – to guide choices for AI system design, development, deployment, and use.This includes ensuring a Privacy by Design approach in which we adopt a privacy mindset throughout decision-making and use state-of-the-art methods to keep data private and secure.
Transparency
Provide users with transparency into our use of AI
Recent classes of AI technologies, such as Generative AI, make it difficult to understand the reasoning behind an AI system’s output. eBay strives to be transparent to end users about its use of AI.In some cases, we accomplish this by disclosing information about the type of AI being used within the experience itself, and in other cases, through broader disclosure of our use of AI generally, based on the stage of its lifecycle.
Our transparency approaches will also strive to communicate most effectively given the role of the individuals interacting with or using the AI system.
Transparency and accountability are going to be key for many sellers, especially as eBay is increasingly shuffling them off to automated AI "self-service" tools for critical support, trust and safety functionality that was previously handled by humans.
VP Platform Engineering, Senthil Padmanabhan also shared how eBay developers are using AI to "cut through the noise" and increase productivity by utilizing commercial offerings, fine-tuning existing Large Language Models, and leveraging internal networks.
According to Senthil, eBay expanded their use of GitHub Copilot to all of their developers last year with positive results including an increase in perceived productivity, good levels of accuracy and time savings features like converting comments to code, suggesting the next line of code, generating tests, and auto-filling repetitive code
But limitations in Copilot's prompt size made certain tasks that require the knowledge of the millions of lines of code in eBay's entire codebase simply not possible.
So, eBay also tried to post-train and fine-tune open source LLMs using their own pre-processed data to create eBayCoder: Code Llama trained on eBay’s codebase and associated documentation.
eBayCoder was able to make some labor- and time-intensive tasks like software upkeep much easier and could potentially reduce the amount of code duplication that may occur from using existing commercial LLM offerings.
Thirdly, Senthil discussed creating an internal GPT that ingests certain data from relevant primary sources to help cut down on the amount of time developers spend investigating to find answers to questions like: “Which API should I call to add an item to the cart?” “Where do I find the analytics dashboard for new buyers?” “How do I create a pipeline to deploy my application to production?”
He says the usage of the internal GPT is increasing, initial feedback on efficiency and relevance are positive and early results are promising. However, like with any automated chat system, sometimes the GPT can deliver nonsensical answers that are frustrating or unhelpful, but can be improved with consistent effort.
eBay is incorporating employee feedback to reinforce and train the models, hopefully making the GPT better over time.
Senthil also shared a video clip of the internal eBay GPT at work:
