Hong-Yi TuYe

By combining funding event data from Pitchbook and resume data from Revelio, we construct panel dataset of headcount and funding data involving 2,500 YC companies from 2015-2025. Using a triple differences approach, we compare YC startups that received a series A round before the advent of ChatGPT in 2022Q4 against YC startups that received a series A round after 2022Q4. We find that companies who received series A funding in the post-ChatGPT era have a 5% slower overall headcount growth when compared to those who received series A funding in the pre-GPT era. Among these series A companies, those with pre-funding treatment workforce in the top quartile of exposure to LLM use (based on Elondou et al. 2024 task-based measure) experience greater downward pressure on headcount growth than the less exposed firms. Across role types and seniority levels, we find an inverse U relationship where both junior and executive level workers see slower headcount growth than their senior counterparts in the post-GPT era, and this pattern is strongest across engineering roles and weakest across external and commercial facing roles. Finally, we construct a novel measure of product exposure to AI by comparing company descriptions against recent themes found in YC’s Request for Startups documents. After controlling for product exposure to AI, we find that firms with high product exposure to AI experienced a 5% slower total headcount growth than firms with low product exposure to AI in the post-GPT era.

In a preregistered online experiment with 1,893 participants who submitted more than 18,000 prompts and generated over 300,000 images, users attempted to replicate a target image in 10 tries using one of three randomly assigned models: DALL·E 2, DALL·E 3, or DALL·E 3 with automated prompt rewriting. Users with access to DALL·E 3 achieved higher image similarity than those with DALL·E 2—but only about half of this gain (51%) came from the model itself; the other half (49%) resulted from users adapting their prompts. Prompt adaptation emerged across the skill distribution, was driven by trial-and-error, and could not be replicated by automated prompt rewriting, which erased 58% of the performance improvement associated with DALL·E 3. The findings position prompt adaptation as a dynamic complement to generative AI and suggest that without it, a substantial share of the economic value created when models advance may go unrealized.