Inside Pinterest's efforts to replace expensive AI with open-source models

In a bold move to enhance its AI capabilities while simultaneously reducing expenses, Pinterest has announced a significant shift in its approach towards artificial intelligence. As the demand for AI-supported tools surges, the visual-centric social media platform is reallocating part of its AI budget to focus on open-source models. According to Vicky Gkiza, Pinterest's Vice President of Product Management, the company has adopted a "model-agnostic" strategy that blends its own proprietary models with both closed-source and open-source AI technologies. This strategy, initiated in 2023, aims to optimize performance while minimizing costs, a necessity in today's competitive tech landscape. Closed-source AI models, like those from Anthropic and OpenAI, are known for their efficiency and ease of integration into existing systems. However, they come with a higher price tag. On the other hand, open-source models, such as those from Alibaba, offer flexibility and are free to use but require skilled engineers for customization and maintenance. Lan Guan, Chief AI and Data Officer at Accenture, noted that many companies are now recognizing the benefits of a multimodal AI approach that balances performance with cost control. "Open-source will be a really good option," Guan explained, emphasizing the importance of managing expenses related to token usage in AI. Pinterest is actively refining its generative AI capabilities through beta testing and strategic hiring, focusing on attracting talent with strong software development skills. The company employs OpenAI's large language models to enhance user features and leverages Claude from Anthropic for internal applications such as coding. Additionally, Alibaba's Qwen model is utilized for visual understanding and data labeling tasks. The results of this innovative approach are starting to materialize. Pinterest has introduced new AI features, including auto-collages, which allow advertisers to create pins from product catalogs for categories like skincare and home decor. After initial testing in early 2024, the feature was ready for a pilot program by mid-2025, involving select retail partners like Macy's. Furthermore, Pinterest's voice-enabled AI feature, which facilitates natural language queries, showed promising user engagement during beta testing. Gkiza remarked, "Search has been evolving so fast. It was imperative for us to use AI to improve." To further bolster its AI initiatives, Pinterest is expanding its workforce, having recently appointed former Google engineer Matthias Zenger as Vice President of Engineering and software engineer Mirjam Wattenhofer to enhance user experiences in e-commerce. The company is also establishing an Engineering Excellence Hub in Zurich, dedicated to advancing its AI and machine learning technologies. As Pinterest moves forward, it claims that its blended approach to AI is yielding substantial cost savings, reportedly 90% less than when it relied solely on proprietary models. With plans for increased investment in cloud computing infrastructure, Pinterest is set to continue its experimentation with various AI models, prioritizing personalization and cost-effectiveness while maintaining high performance when necessary.