Sarah Catanzaro

Sarah Catanzaro is an investor in data science and machine-learning infrastructure at Amplify Partners, which promoted her to General Partner in January 2022. Her work includes backing data-curation startup Datology, supporting Modal’s founders, and arguing that faster AI progress requires better experiments, evaluation, and reproducibility. She describes her investment approach as seeking tools she wished she had during her own career building data systems.[5][7][2][8]

In her August 26, 2025 essay, “The AI research experimentation problem”, Catanzaro argued that agents capable of generating research ideas, selecting promising ones, and implementing them in code would not alone meaningfully accelerate progress toward AGI. She identified an overlooked bottleneck in designing rigorous experiments, running them reliably, and interpreting their results.[8]

Her argument makes the tradeoffs concrete. Testing several hypotheses in one training run can conserve compute while making it harder to identify which intervention caused a performance gain. Training bugs can reduce model quality even while loss continues to fall, and evaluation deferred until the end can expose failures after their causes have become difficult to reconstruct. She recommends budgeting for evaluation from the start, using fast feedback during training, and preserving the code, configuration, and data needed to reproduce a run.[8]

Catanzaro also argues that tools alone cannot fix research practice: organizations must reward reproducibility, make experiment design a priority, and support collaboration among researchers, statisticians, and engineers. Her conclusion is explicit: “But real progress will require organizations to change how they work.” The essay credits Justin Gage as editor and thanks reviewers including Jonathan Frankle, Emmanuel Ameisen, Jason Liu, and Stanislav Bekman.[8]

Catanzaro’s February 2024 Datology investment account explains why she viewed training-data curation as an investment opportunity. She recalls conversations throughout 2023 with teams that wanted to move from commercial LLM APIs to self-hosted models and use proprietary data, but struggled to assemble suitable pretraining or fine-tuning datasets. Her technical questions concerned removing duplicate, redundant, difficult, or harmful examples, constructing batches, and sequencing training data.[7]

In that account, Jonathan Frankle connected the investor team with Ari Morcos in April 2023. Catanzaro and her colleagues then introduced Morcos to practitioners to test demand for the proposed platform. Morcos recruited Matthew Leavitt, his former FAIR colleague; Bogdan Gaza subsequently joined, bringing data and backend-infrastructure experience. Catanzaro described Amplify as Datology’s first investor, while crediting Morcos, Leavitt, and Gaza with the research and engineering needed to build the company.[7]

Her relationship with Modal co-founder Erik Bernhardsson developed through technical discussion before an investment. The February 2022 Modal account dates their meeting to fall 2019. Bernhardsson, the creator of Annoy, challenged her thesis about vector-based systems replacing token-based counterparts. Their conversations continued for more than two years, covering engineering efficiency and shortcomings in data-workflow orchestration, before an Amplify-led seed round for the company he founded with Akshat Bubna.[6][2]

Bernhardsson credits Catanzaro with helping shape Modal’s initial vision, making introductions, and supporting fundraising and company building. His testimonial, published on her Amplify biography, describes her contribution as an investor and partner to the founders.[2]

Catanzaro’s later AI infrastructure interests include memory, continual learning, stateful inference, and the systems problem of loading and unloading personalized weights at scale, according to the episode summary for her December 2025 Latent Space conversation, recorded at NeurIPS. Her investment thesis, as summarized there, favors applications that solve hard research problems and enable experiences that were previously impossible.[1]

Before entering venture capital, Catanzaro worked on security analysis and commercial data products. Those earlier roles provide concrete context for her stated preference for investing in tools that help practitioners collect, manage, analyze, and model data.[3][2]

Catanzaro earned a B.S. at Stanford, where she became interested in interdisciplinary research on insurgencies at the Center for International Security and Cooperation. She recalls subsequently working at the Center for Advanced Defense Studies with MIT AI researchers on computational linguistics to evaluate terrorist threats and with Carnegie Mellon computer scientists on network-analysis packages to map Somali pirate organizations. Her later work at Palantir exposed her to agencies’ data-integration and modeling difficulties.[2]

As Head of Data at Mattermark, Catanzaro helped build a private-company intelligence platform for investors. In a July 2022 interview, she recalled integrating startup information from disparate services into a consistent data model, monitoring model performance, and developing separate industry classifiers for categories such as healthcare and e-commerce. The team also devised its own data-quality scorecard. She placed these challenges around 2013–2014 and cautioned that she was recalling work from several years earlier.[3][2][9]

Catanzaro worked as a Data Partner at Canvas Ventures before joining Amplify in September 2017. Amplify credits Shivon Zilis with introducing her to the firm in late 2016. It announced her promotion to Partner on January 15, 2020, reporting that she had sourced six early-stage investments in data science, data management, or machine learning during her tenure. Her January 2022 General Partner promotion also recognized her work on the Projects to Know newsletter.[3][4][5]

In the 2022 interview, Catanzaro distinguished easier model prototyping from the continuing difficulty of building production ML applications. She argued that narrowly focused tools left teams assembling poorly integrated components and called for tools designed around integration and extensibility. She also favored founders who had repeatedly encountered a technical problem as practitioners, citing their understanding of users and clear product vision as reasons to back them.[9]