The Truth About Building AI Startups Today

The partners say nearly half of YC’s Summer 2023 batch used large language models, reflecting what founders wanted to build rather than a preference for AI applications. They describe the moment as unusually open to new founders, since no one has years of experience with LLMs.

The partners see promise in automating routine information work, such as searching records, summarizing documents, and transferring information between systems. They point to Sweetspot, which pivoted to automating government contract searches and proposals after learning that a contractor spent a workday refreshing a government website.

AI copilots can attract sign-ups and upfront payments even when customers have no clear use for them. The partners advise checking whether people use the product for a specific task; they also warn that cheaper fine-tuned models may lose customers as model costs fall, unless they offer other value such as private-data customization.

The partners argue that calling a product a GPT wrapper overlooks the value of good user experience and software design. To withstand future foundation models, they recommend solving a concrete user problem with domain-specific business logic, such as sales next steps or compliance workflows.

The discussion explores AI voice agents that handle calls for small businesses and specialized models that can run locally for tasks like coding. The partners also raise concerns about private-data security and concentrated control of AI, pointing to emerging LLM security tools and the case for open-source models.