The launch of ChatGPT in November 2022 marked the beginning of the Age of AI, permanently transforming the tech industry. Since December, AI has become a prominent feature in nearly every pitch deck I've encountered. However, as with any emerging technology, venture capitalists like myself have had to swiftly devise a strategy to differentiate promising startups from those that are mostly hype or likely to encounter insurmountable challenges, hindering their potential for significant growth. This requires a comprehensive understanding of the various layers of the generative AI value stack, identification of viable investment opportunities, and the development of a thorough due diligence approach to assess the risks and potential of each startup. Generative AI comprises several components within its tech stack, each presenting unique investment prospects and challenges for seed-stage companies. Let's delve into these different aspects. One of the most significant challenges and, consequently, areas of opportunity in generative AI lies in the accuracy and reliability of the information it generates.

Presently, generative AI models rely on extensive datasets, some as comprehensive as the internet itself, which encompasses both valuable and irrelevant information. However, we anticipate that in the following years, the emergence of generative AI applications will involve more precise data or fragments from specialized models. Instead of using expansive datasets, these specialized models will leverage proprietary domain-specific data, resulting in personalized outputs that guarantee accuracy. By infusing foundational models with proprietary data using the appropriate middleware architecture, these specialized models will drive the application layer that interacts with consumers and businesses. Accompanying the data layer in the generative AI stack is middleware, which refers to the tools and infrastructure supporting the development of new generative AI applications. This aspect forms the second part of our investment thesis in the sector. Specifically, we express optimism towards infrastructure and tooling companies that assess and ensure safety, accuracy, and privacy of model outputs, orchestrate inference across multiple models, and optimize the integration of proprietary data into large language models (LLMs).