Sarah Guo - The Power of Conviction - [Invest Like the Best, EP.383]

Episode Summary

Executive Summary: Patrick O'Shaughnessy interviews Sarah Guo, founder of Conviction, about building an AI-focused VC firm and how AI changes venture, software, and market structure. She argues AI is a new software era where investors must assess model quality, incumbents, efficiency, and product-market fit with first-principles rigor.

Main Topics: Starting Conviction (Priority: 5/5): Guo explains leaving Greylock to launch a pure early-stage AI fund and why timing mattered. AI application investing (Priority: 5/5): She says app-layer value comes from workflow depth, distribution, and product quality, not just the model. Minimum viable quality (Priority: 5/5): AI products win when outputs are good enough for real use, even if not perfect or general. Frontier-model ecosystem (Priority: 4/5): She expects a competitive, multi-model market with open source and efficiency as enduring advantages. Infrastructure and chips (Priority: 4/5): She argues hardware, data centers, and inference stacks remain immature but full of opportunity. Risks and adoption (Priority: 3/5): She is more concerned about practical abuses and enterprise adoption hurdles than sci-fi risks. Long-term societal impact (Priority: 4/5): Guo expects abundance, new creative tools, and major shifts in learning, work, and services.

Key Arguments: Great AI investors need tech understanding, competitiveness, team orientation, and judgment signals. Early-stage venture has tiny sample sizes; pattern recognition comes from long company-building exposure. AI app value is in the 99 miles of workflow, distribution, and change management, not just the model. Avoid building directly on incumbent strengths; big labs won't win every market or product. Minimum viable quality decides adoption: if outputs aren't good enough, customers won't use them. Efficiency matters in AI because compute is scarce, costly, and getting more important over time. The market will likely support many frontier-model players; 'a million flowers' beats one winner. Enterprise adoption is still held back by hallucinations, calibration, and risk/compliance concerns. The biggest near-term AI risks are misuse, fraud, and misinformation, not only existential scenarios.

Data Points: Greylock tenure before founding Conviction: 9 years - Sarah Guo spent nine years at Greylock before starting Conviction. Conviction launch year: 2022 - She founded Conviction in 2022. Greylock history: 75 plus years old - Guo describes Greylock as an established venture firm with a long history. Awake company outcome: sold to Arista - The network-analysis security company Awake was eventually acquired by Arista. Figma build time to public release: five years - Guo cites Figma as a long-build company before public release. Harvey runway: publicly reported tens of millions of runway - She says Harvey had substantial runway and leading customers. Legal labor cost example: $2,000 an hour - She uses elite legal billing rates to illustrate AI productivity gains. Contract search example: 20,000 contracts - She describes large-scale contract review as a use case for Harvey. HeyGen input requirement: two minutes of video - She says the product can work from consumer-quality video input. Sweep benchmark example: 13, 15, even 20% - She cites code-generation benchmark performance as insufficient today. Healthcare share of U.S. economy: a quarter of the American economy - She highlights healthcare as a huge but inefficient target market. OpenAI pre-training entry price: more than half a billion dollars - She says entering general-model training is now extremely capital intensive. Model deployment horizon: three to five years - She is asked to think about the foundation-model market structure over this time frame. Google chip scale example: 10,000 - She says proving non-NVIDIA chips at scale can require large deployments. Google products priority example: 19th most important product - She uses this as a proxy for how little some projects matter inside large incumbents.

Pivotal Quotes: "the 99 miles is a lot" — Sarah Guo: Her view that AI application value lies in the hard last-mile workflow, not just the model. "I want a million flowers to bloom" — Sarah Guo: Her preferred market structure for frontier models and AI infrastructure. "if something is really easy for you to build, unless your distribution is totally unique and defensible, you're probably not going to be able to capture rent" — Sarah Guo: Her warning that simple AI wrappers rarely create durable value.

Implications: The open question is which AI products cross quality thresholds fast enough to become durable businesses before model and infrastructure costs reshape competition.