Insights from the Cutting Edge of AI Investing: A Conversation with Sarah Guo, Founder of Conviction

Detailed Summary

Introduction and Conviction's Focus (0:00 - 3:00)

• Sarah Guo emphasizes the importance of an ambitious roadmap in AI development, noting that product teams often assume future capabilities will enable them to achieve everything eventually.
• She highlights the need to continuously close the gap between current products and ambitious goals by leveraging new orchestration techniques and tool use.
• Ian Karens introduces Sarah Guo, founding partner at Conviction, an early-stage AI venture firm that has invested in companies like Sierra, Harvey, Mistral, and Base 10.
• Conviction, a two-and-a-half-year-old startup, focuses on full-stack AI investing across training, applications, and infrastructure, believing AI represents the biggest technological change in a long time.

Journey into AI Investing (3:00 - 7:55)

• Sarah Guo explains her background in machine learning at her prior firm, Greylock, where she saw its impact on products like Facebook's ad targeting.
• Her interest was fueled by technical curiosity and hanging out with leading technical people, including early deep learning pioneer Andrew Ing.
• She observed the increasing importance of AI across various applications, from marketplaces to email security, and noted the growing demand for inference and optimization in the last two years.
• A key insight came from Mustafa Suleyman (DeepMind co-founder, now at Microsoft) and others, who championed natural language interfaces and productivity assistants, which became feasible with new AI generations.
• After a decade at Greylock, she recognized AI as the biggest opportunity of her career, leading her to start Conviction.

The Impact of LLMs and Generative AI (7:55 - 11:27)

• Ian Karens discusses the shift from data-heavy ML to LLMs, which require less data, making AI accessible to more companies.
• Sarah Guo acknowledges the luck of Conviction's timing, launching before ChatGPT, but notes her awareness of transformers since 2018.
• ChatGPT's launch was pivotal, as it made the potential of AI clear to business leaders globally, creating openness for new applications.
• She highlights the challenge of keeping up with the rapid evolution of AI techniques, such as reinforcement fine-tuning, RAG, agent orchestration, and guardrails, which change monthly.
• The core premise of Conviction remains that software development is changing, focusing on manipulating models, which dramatically reduces the cost of deploying sophisticated, magical solutions.

Staying Ahead in a Rapidly Evolving Landscape (11:27 - 17:55)

• Sarah Guo notes the positive trend of rapidly increasing AI model capabilities and tooling quality, but this is offset by the challenge of constantly adapting model manipulation techniques.
• She mentions that within sophisticated organizations, reinforcement fine-tuning is data-efficient for complex agentic tasks.
• The skill of staying updated on state-of-the-art techniques is crucial, involving primary learning about one's business and model capabilities.
• She references OpenAI's recent 4.1 model release and its tactical prompting guide as an example of how applied teams, working closely with research, adapt quickly.
• Teams that act like they are in a continuous scoping phase, maintaining simplicity, fast prototyping, and understanding edge cases, tend to perform better.
• The advice from OpenAI's Dev Day to "build something today for what's at the edge of possible" reinforces the need for ambitious roadmaps.

Cultural Modes and Prototype-to-Production Challenges (17:55 - 33:19)

• Sarah Guo contrasts the technology-up approach of AI labs, which might abandon use cases if reliability is an issue, with the customer-back approach of enterprises like a major financial institution.
• Enterprises have clear use cases, measure quality, and adhere to strict compliance and safety bars, sometimes refusing to show generated outputs directly to end-users.
• This bifurcation in orientation means companies must either hunt for opportunities not sensitive to model failures or address business-critical needs within existing limitations.
• The struggle to move AI prototypes to production is attributed to the extensive engineering work required for real-world integration (e.g., with CRM, phone systems, brand voice).
• Another factor is the constantly moving bar of expectations; what was considered an "agent" a year ago is now a basic retrieval system.
• The need for a new statistical lens on risk is crucial, as product leaders must now evaluate if 80%, 90%, or 98% accuracy is acceptable, a framework not common before.
• Ian Karens adds that many prototypes fail because they address less valuable problems or because off-the-shelf SaaS solutions now exist, reducing the incentive for in-house development.

Product Thinking and Organizational Adoption (33:19 - 39:50)

• Sarah Guo believes the industry is only 1% into exploring possible AI use cases, even with current models, due to a lack of intuition among most people about what these models can truly do beyond ChatGPT.
• There's a huge opportunity for individuals with product instinct to apply AI to real-life workflows, such as interacting with websites to collect information programmatically.
• She observes that many C-level executives' biggest frustration is not their roadmap, but getting their organization to adopt AI tools and think differently.
• Leadership plays a critical role in fostering AI adoption by creating space for investment, managing learning costs, enabling tooling, and addressing fears of job displacement.
• The primary driver for AI adoption in Conviction's portfolio is competitiveness and growth (e.g., increasing market share, improving customer experience), not headcount reduction.
• Cost reduction, when it occurs, is typically related to vendors or external service providers, not internal employee headcount.

Opportunities for Founders (39:50 - 43:35)

• Sarah Guo advises aspiring founders to leverage their unique understanding, either from an end-user or engineering perspective, to identify problems.
• She cites Abnormal Security founders, who applied their knowledge of communication patterns and user identity from ads retargeting to email security.
• She also mentions Parker Conrad of Ripling, who became an expert in HR software through obsessive dedication, highlighting the importance of genuinely caring about the problem.
• Sarah suggests insurance as a promising industry for AI, given its large volume of inefficient text processing in areas like communications and claims, presenting a significant vertical player opportunity.
• Ian Karens concludes the podcast, thanking Sarah Guo and encouraging listeners to engage with the show.

Notable Quotes

"I think it just helps to have like a very ambitious roadmap." — Sarah Guo

"One thing that strikes me every time I talk to people who are building product in conjunction with the research teams at the labs or even some of our companies that have like agentic workflow products out in the wild like Harvey is they just assume that they're going to be able to do like everything eventually." — Sarah Guo

"The cost to deploy something that's magical and really sophisticated goes down dramatically if you're not training from scratch and if you can be efficient about getting it to do like this super sophisticated task." — Sarah Guo

"The skill within organizations that is very useful is like being up to date on the latest and greatest from a techniques perspective." — Sarah Guo

"People are holding a higher bar for AI than they are for people." — Sarah Guo

"The key driver for adopting the products of any of our companies at conviction today has not been headcount reduction. It is it is almost always like I want to become more competitive. I want to increase my pace of business. I want to increase my market share. I want to produce something of higher quality for my end customers." — Sarah Guo

"I think people mistake the most people in America have tried Chad GBT at some point for like there's any intuition for what these models can do." — Sarah Guo

"Insurance is such a it's an industry that has so much text processing that is incredibly inefficient that is very large." — Sarah Guo