Logan Kilpatrick on AI Developers, Interfaces, and the Future of Models

Matthew Berman interviews Google’s Logan Kilpatrick about developer tools, AI interfaces, browsing, models and creative media. The conversation was recorded in July 2025, so product references and expectations reflect that moment.

Berman begins with the Windsurf team members joining Google DeepMind. Kilpatrick sees their arrival as evidence of the continuing opportunity to change how software is built, after the first wave of excitement around GitHub Copilot.

Berman asks why Google hired some of the team rather than acquiring the product, intellectual property and customers that went to Cognition. Kilpatrick says he does not have the full details of the deal; he praises Windsurf and says he uses it himself.

For Kilpatrick, the move also signals DeepMind’s growing focus on developer products alongside its foundational and scientific research. He is excited to work with people who care about building tools developers enjoy.

Berman asks about team acquisitions and the consequences for founders and employees who do not move. Kilpatrick says acquisition and other ways to join forces can provide important paths for startups, though he does not resolve all of Berman’s concerns about individual outcomes.

On the best interface for an assistant, Kilpatrick expects a mix of chat, audio and background agents. He imagines a queue of proposed actions, similar to GitHub pull requests, where a person reviews and approves useful work.

Glasses could provide visual context, he says, and context is central to a useful assistant. Yet much valuable work happens on computers and phones, where a user can already choose to share the screen; glasses are not necessary for every use case.

Kilpatrick discusses Project Astra as a test bed for visual and audio interaction, and Gemini’s native audio capabilities. He says Veo 3 with audio had just become available through an API; these were product statements made at the time of the interview.

Voice is useful when typing is inconvenient, such as in a car, while text is often faster for copying a passage or working at a desktop. Kilpatrick imagines an assistant adapting its interface to the situation instead of using one interaction style everywhere.

Berman describes using voice on a drive and imagines asking an assistant to arrange a restaurant booking. Kilpatrick notes the limits: reviewing a detailed code change by voice in a car is awkward. An assistant might instead defer that notification until the user returns to a screen.

Berman suggests that an assistant could read important email and propose replies during a drive. He then asks whether agentic browsing might mean people stop visiting some websites altogether.

Kilpatrick argues that people value work and perspectives from other people. AI summaries can help, but sometimes he wants the view of a particular trusted person rather than an aggregate. As convincing synthetic content becomes easier to produce, he sees authenticity and discovery as increasingly important.

He uses shopping for an anniversary gift as an example. He did not use deep research for that purchase, but says a system could examine hundreds of pages he would never have read himself. It could supplement search while he remains responsible for the choice.

Berman sees travel planning as an obvious agent task because airline and hotel sites require repeated work. Kilpatrick is more skeptical for simple trips, while accepting that complex itineraries could benefit. Berman suggests even a familiar booking might save time if an agent handles it with confirmation.

The conversation turns to models becoming systems. Berman asks how much value lies in the model and how much in the surrounding tools, memory and orchestration, especially for a startup deciding what to build.

Kilpatrick says early production applications often needed elaborate scaffolding to compensate for model limitations. He gives NotebookLM as an example of a workflow that, he says, became much simpler as models and the team’s understanding improved.

Reasoning models can increasingly use search, code execution and function calls as part of producing an answer. Kilpatrick expects some features that once required external orchestration to become more native to model-based systems.

He does not expect the need for scaffolding to disappear. As models gain capabilities such as browser access, new production problems arise around using those capabilities reliably. The frontier moves even when an older layer becomes unnecessary.

His advice to founders is to build the scaffolding needed for today’s product but stay ready to simplify or replace it as models improve. Rapid change creates opportunity and also demands continual adaptation; he describes further model improvement as his working expectation, not a guarantee.

Berman asks why Kilpatrick expects the number of engineers to grow even as AI makes each more productive. Kilpatrick notes his own bias as a software engineer and separates foundational problem-solving skills from the particular tools used to express a solution.

He values systems thinking, persistence and learning to work through failed attempts. Berman agrees that these skills make learning to code valuable for his children even if the interface to programming changes.

Kilpatrick argues that people who understand code can inspect and refine generated work more effectively. A newcomer may need courage to look under the hood, and current products often fail to teach users while they wait for generation.

He sees a product opportunity in teaching people as they build, exposing useful details instead of simply doing everything for them. He hopes such tools can bring many more developers into the ecosystem; the 100 million figure is an aspiration.

Berman asks whether a natural-language specification might become the source of truth, compiled into code and tested each time. Kilpatrick worries about losing deterministic behavior for serious systems if translations vary between runs.

For low-stakes applications, a natural-language workflow may be adequate, Kilpatrick says, but control and predictable behavior make code important for other uses. Berman observes that a fully precise natural-language specification can itself start to resemble code.

They move to Veo 3, which Kilpatrick says had just reached the API. He describes native audio accompanying generated video and the possibility of lowering the cost and skill barriers for creating media.

He expects developers to build creative tools beyond Google’s own products and suggests small, highly targeted marketing campaigns as one use case. Berman argues that as generation becomes easier, taste and the ability to curate experiences become more valuable.

Kilpatrick names taste, authenticity and agency as valuable human qualities. Berman connects agency to a question he is considering as a parent: how schools can help children develop it.

Kilpatrick says Veo 3 changed his view of the practical value of generated video compared with an earlier version. He hopes for continued progress but does not announce a future version or guarantee the same rate of improvement.

Berman asks about game-like control and world simulation. They discuss research and demos as a possible route to interactive or expanding environments. Kilpatrick treats these as possibilities, not confirmed product commitments.

Asked what users want next, Kilpatrick mentions cost and longer generations, and says a faster Veo 3 variant was expected at the time. He describes a tradeoff: longer single generations can require the model to make more assumptions and may be harder to control.

He points to creators assembling coherent longer videos from multiple generations, comparing the process to human filmmaking with many shots and iterations. Creative tools and a suitable production workflow can help users guide the model.

Berman asks whether diffusion-based language models could rival transformers. Kilpatrick says the research question is whether diffusion scales similarly. He sees potential in fast generation, including code and dynamically generated interfaces, but notes cost and quality tradeoffs.

Speed matters because long waits break the iteration loop when people build with AI. Berman says a poll of his audience put relatively little weight on speed, while both speakers think it deserves more attention.

They discuss scaling pretraining, post-training and reasoning-related reinforcement learning. Kilpatrick attributes Gemini 2.5’s progress to work across several dimensions and argues that improvements can interact. His characterization is an assessment, not a quantified universal law.

Berman asks about Gemini 3, but Kilpatrick does not disclose concrete details. Berman says he would like faster models and more built-in tools in AI Studio. Their exchange is a feature wish list at the time, not a product announcement.

On competition, Kilpatrick says different teams are trying to solve problems for customers and that he appreciates investment across the field. He identifies DeepMind’s breadth of research and Google’s distribution as potential strengths.

He points to scientific research alongside language models and the responsibility of deploying AI to a very broad audience. Many users encounter generative AI through existing products without following the technology closely, so product design must account for different levels of familiarity.

Berman asks about TPUs. Kilpatrick says control over specialized hardware can help Google deploy models at scale and manage economics, especially when products serve large audiences. The usage figures he cites are claims from the 2025 conversation.

Berman thanks Kilpatrick and closes the interview. Their discussion leaves open how quickly interfaces, models and developer work will change.