Logan Kilpatrick

Logan Kilpatrick leads product for Google AI Studio and the Gemini API at Google DeepMind. His team enables developers to build with Gemini, Veo and Imagen. Previously OpenAI’s first Developer Advocate, he has worked on documentation, developer experience and the practical costs of building AI applications.[1][3][8]

Before his AI-platform roles, Kilpatrick worked in software engineering and machine learning. His career summary describes early NASA work on lunar-rover software, followed by training machine-learning models at Apple. In his February 2023 Latent Space interview, he recalled that Julia community advocacy supplied an outlet for developer-facing work while his paid jobs centered on engineering and model training.[2][3]

Kilpatrick also supported NumFOCUS, which provides legal and fiscal support for scientific-software projects including Julia, Jupyter, pandas and NumPy. He recalled an unsuccessful attempt to obtain a Wikipedia page for the nonprofit: “we tried at one point to get a Wikipedia page”. He saw a visibility gap between the tools and the organization supporting them. Alongside this community work, he taught Django as a teaching fellow through Harvard’s division of continuing education.[3]

In the 2022 MLJ model-documentation project, Kilpatrick helped organize work on a concrete usability problem. MLJ supplies a common interface for selecting, tuning, evaluating and composing machine-learning models in Julia and other languages. Its model-specific documentation was scattered across third-party packages, sometimes terse and inconsistently formatted. The project sought to introduce standardized model docstrings explaining individual models and their configuration.[5]

Kilpatrick co-wrote the proposal, helped recruit the technical writer and handled administration. MLJ co-creator Anthony Blaom managed the project, mentored David Josephs and reviewed his documentation; third-party package authors also reviewed contributions. The team selected Josephs for strong data-science skills despite his more limited Julia experience. The authors reported that he acquired the Julia proficiency the work required, reinforcing their preference for domain knowledge and communication skills over prior mastery of a particular language.[5]

The team’s initial plan allocated equal time to model-providing packages, but packages contained different numbers of models, with substantially different hyperparameter lists to explain. Individual-model coverage proved more useful than package counts for measuring progress. The case study documented additions across multiple packages while identifying four proposed packages that did not receive new docstrings during the project. The rollout issue tracked the wider standardization effort.[5][6]

Kilpatrick dates his OpenAI developer-relations work to 2022 through early 2024. By February 2023, he was the company’s first Developer Advocate. He explained that substantial inbound interest made documentation and developer experience the immediate priorities, rather than the external content and conference appearances often associated with advocacy. Colleagues had already contributed documentation alongside other responsibilities; his full-time task was to improve it so developers could build products on the API.[2][3]

A specific priority was connecting the OpenAI Cookbook to the main documentation. Kilpatrick credited collaborators Ted and Boris with its tutorials and described his role as making their work more accessible and visible while maintaining the documentation’s quality standard. He also helped move developer-relevant information from the help center into the docs and supported what he called a “documentation first culture”.[3]

Making developer relations his paid work changed the balance of his community commitments: Kilpatrick recalled having less energy for Julia advocacy after full days supporting OpenAI developers. His technical work also involved explaining unfamiliar API behavior. In February 2023, he described shipping a documentation update warning that language-model outputs could vary even at temperature zero, challenging developers’ expectations that identical requests would produce identical responses.[3]

By early April 2024, Kilpatrick had announced his move to Google to lead product for AI Studio and support the Gemini API. In the February 2025 Gemini 2.0 announcement, coauthored with Shrestha Basu Mallick, he explained a simpler pricing structure for Flash and Flash-Lite: a single price per input type, removing Gemini 1.5 Flash’s distinction between short and long contexts. This could lower costs for mixed-context workloads.[10][7]

Kilpatrick’s May 2025 implicit-caching announcement explained how Gemini 2.5 requests sharing a common prefix could qualify for cached-token discounts without explicitly creating a cache. His guidance was to “keep the content at the beginning of the request the same” and place changing questions or context at the end. For example, a document-question-answering application could put its shared document first and append each new question.[8]ⁿ¹

A stable prefix established eligibility, rather than guaranteeing an implicit cache hit. The response field cached_content_token_count exposed how many input tokens had been cached and would receive the lower price. Explicit caching remained available for developers seeking guaranteed caching cost savings.[8]

In May 2026, Kilpatrick and Varun Mohan coauthored Google’s announcement of managed agents in the Gemini API and AI Studio. The agents combined reasoning, tool use and code execution in isolated Linux environments; follow-up calls could resume files and state. The announcement also described exporting AI Studio projects to Antigravity with their context preserved, carrying work from the prototyping environment into local development.[9]

In February 2023, Kilpatrick welcomed the forthcoming conversational API while remaining uncertain whether chat would endure as the preferred user interface. He argued that products designed around generative AI could have a short-term advantage, but durable differentiation would require more: distinctive interfaces or unique data that could support fine-tuning or specialized models. He cited his own use of Jasper, whose prepared prompts and simple interface made tasks such as generating product descriptions easier.[3]

Kilpatrick also urged builders to evaluate models against their own applications. In the same interview, he argued that companies needed to codify how people used their products so they could compare model changes on relevant tasks, rather than relying only on academic benchmarks. He described understanding and communicating model capabilities and limitations as an unresolved technical challenge.[3]

On his personal website, Kilpatrick advocates taking initiative, treating many decisions as reversible and shipping quickly: “Iteration speed often makes up for mistakes.” He also argues that human authenticity becomes more valuable as AI expands. He cohosts Around the Prompt with Nolan Fortman, interviewing AI practitioners and experts.[2]