OpenAI Researcher turned VC on what we’re missing on the AI bubble | Term Sheet

Term Sheet Podcast with Jenny Xiao: AI Progress, Valuation, and the Future Landscape

Key Concepts:

• Lumpy AI Progress: AI development doesn't occur linearly, but in bursts of significant advancement.
• Zero Value Threshold: A point at which an AI company’s enterprise value drops to zero, often due to being surpassed by open-source models or absorbed by larger players.
• AI Safety: The responsible development and deployment of AI, considering potential risks and unintended consequences.
• Distillation (in AI): A technique where a smaller model is trained using the outputs of a larger, more powerful model.
• AI Workflows vs. AI Agents: A shift from the hype around autonomous AI agents to more practical, well-defined AI applications within specific business processes.
• Compute: The computational resources required to train and run AI models, a key cost factor and competitive advantage.

I. Introduction & Market Context (Alli Garfinkel)

The podcast episode begins with Alli Garfinkel setting the stage for continued volatility in private markets, with AI as a major contributing factor. She introduces guest Jenny Xiao, founder of Leonis Capital and former OpenAI researcher, highlighting her unique perspective on AI, particularly regarding the international landscape, specifically China. Garfinkel also briefly discusses the unexpected Department of Justice investigation into Federal Reserve Chair Jerome Powell, emphasizing the potential implications for the stability of the US economy and investor confidence. She notes that VCs often underestimate the impact of macroeconomic factors, but the long-term health of the economy is crucial for all investments. Finally, she previews an inside look at Strava, noting its potential IPO and the role of female Gen Z runners in its turnaround.

II. Jenny Xiao’s Perspective on AI Safety & Development

Jenny Xiao frames the discussion around AI safety as currently “niche” within Silicon Valley, despite its importance. She argues it should be a top-three discussion point for the broader AI community. The core issue is the competitive pressure on foundation model labs (like OpenAI and Anthropic) to release products quickly, often without adequate safety testing. Examples cited include early issues with Google Gemini and xAI’s Grok, which exhibited problematic outputs (e.g., racist statements). Xiao emphasizes that safety isn’t simply about making models “dumber” with safeguards, but about unpredictable consequences of poorly designed restrictions, referencing Dario Amodei (CEO of Anthropic) and the challenges of controlling LLM behavior.

III. The Role of Academia & Compute Resources

Xiao identifies academics as crucial for educating the next generation of AI researchers. However, she notes a decline in academia’s role in frontier AI research due to a significant lack of resources, specifically GPU access for large-scale model training. She references Jack Clark’s proposal for a national GPU cluster to address this issue. She observes a trend of PhDs leaving academia for private labs with greater compute capabilities, concentrating cutting-edge research in the private sector. Furthermore, academic research agendas are increasingly influenced by private sector funding and priorities.

IV. Xiao’s Trajectory: From Academia to OpenAI to VC

Xiao details her career path, leaving a PhD program in Economics and AI after one year to join OpenAI in 2021, driven by a desire to impact the industry and work on AI safety. At OpenAI, she focused on benchmarking Chinese language models. She describes a cultural shift at OpenAI, from a mission-driven, lower-paying environment to one increasingly focused on competition and compensation. She left OpenAI to found Leonis Capital, seeking greater agency and the ability to shape an institution.

V. OpenAI’s Evolution & the AI Race with China

Xiao argues that OpenAI is transitioning from a research lab to a platform company, aiming to become an “everything platform” akin to WeChat in the US. She anticipates significant competition in various sectors. She also highlights a key misunderstanding about OpenAI: the company has drastically changed since its early days. Regarding the US-China AI race, Xiao presents a contrarian view: she believes it’s largely a competition within Chinese talent, both in China and among Chinese Americans in the US. She notes the rapid growth of Chinese AI labs, their focus on data labeling (leveraging lower labor costs), and their use of distillation techniques to quickly catch up to US models. She predicts the gap between US and Chinese AI capabilities will continue to narrow.

VI. AI Valuation & the Bubble Question

Xiao challenges conventional wisdom about AI valuation. She introduces the concept of the “Zero Value Threshold,” where an AI company’s value drops to zero if overtaken by open-source models or absorbed by larger companies. She argues that AI companies should be valued lower than SaaS companies due to the inherent costs of compute and the lack of scalability. She believes current AI valuations are inflated, potentially by a factor of 2x. However, she remains optimistic about long-term AI investments, suggesting that a temporary overvaluation is acceptable for a 10-year investment horizon. She identifies companies raising large amounts of funding with no product as a significant red flag.

VII. Future Trends: Workflows over Agents & the End of the AI Agent Hype

Xiao predicts a shift away from the hype surrounding AI agents towards more practical “AI workflows” – streamlined, well-defined applications within specific business functions. She believes the next generation of successful AI startups will resemble established enterprise software companies like SAP, rather than focusing on creating general-purpose AI agents. She anticipates the AI agent hype will subside in 2026.

VIII. Conclusion (Alli Garfinkel)

Garfinkel summarizes Xiao’s nuanced perspective on valuation, emphasizing the importance of compute as a key battleground. She highlights Xiao’s prediction of a market correction in AI valuations and the shift towards AI workflows. The podcast concludes with credits.

Technical Terms & Concepts:

• LLMs (Large Language Models): AI models trained on massive datasets of text to generate human-like text.
• LSTMs (Long Short-Term Memory): An older type of recurrent neural network architecture used in early LLMs.
• Reinforcement Learning: A type of machine learning where an agent learns to make decisions by receiving rewards or penalties.
• Multimodal AI: AI systems that can process and understand multiple types of data (e.g., text, images, audio).
• Distillation: A technique for transferring knowledge from a large model to a smaller one.
• Compute: The computational resources (GPUs, CPUs) required to train and run AI models.
• API (Application Programming Interface): A set of rules and specifications that allow different software applications to communicate with each other.
• SaaS (Software as a Service): A software distribution model where applications are hosted by a vendor and made available to customers over the internet.
• AGI (Artificial General Intelligence): Hypothetical AI that possesses human-level intelligence.

This summary aims to be comprehensive and detailed, preserving the technical precision and specific details of the podcast transcript. It is structured with clear sections and key concepts to facilitate understanding.