Sarah Catanzaro is a former intelligence researcher who modeled insurgent and pirate behavior using incomplete data, experience she now applies to predicting startup success and venture investing patterns. She exclusively backs practitioner-turned-founders who have built internal tools multiple times before spinning them out commercially, focusing on second-generation ML infrastructure that addresses real production challenges rather than just prototyping needs.

• A deep technical interest in the specific mechanisms of AI learning, specifically continual learning and test-time training. She views these as critical next frontiers for model development, frequently discussing how models must learn from context and noting when specific prediction methods, such as context graphs or synthetic data generation, effectively address context rot.
• Scrutinizing the infrastructure stack required to support AI agents. She actively questions architectural choices, such as why developers prioritize files and storage for agents rather than key-value stores or object storage, and highlights the often-overlooked difficulty of building effective search tools to support agent capabilities.
• Identifying opportunities for startups to compete against frontier labs by focusing on interface and application layers rather than raw reasoning capabilities. She points to companies like Gradium that succeed not by out-training big labs, but by developing state-of-the-art models that change the user interface, such as achieving full-duplex conversational AI with ultra-low latency.
• Appreciation for re-engineering core systems to leverage modern hardware economics. She highlights how legacy systems like Postgres were built when memory was expensive, whereas newer entrants like CedarDB are rebuilding layers like optimizers and buffer managers to exploit the fact that RAM is now abundant and cheap.
• Prioritizing product craftsmanship and user delight over VC-led distribution. She emphasizes that while investors can provide introductions, those efforts are futile if the product causes user churn, explicitly noting that companies that matter, such as Runway, succeed because they solve hard research problems to empower users, not just because the problems are difficult.
• A focus on the complexities of scaling Reinforcement Learning (RL) compared to pre-training. She engages with technical research from conferences like NeurIPS to discuss why scaling RL is difficult and does not just work out of the box, contrasting pure RL approaches against post-trained Large Language Models.