Gemini's progress came from combining research teams, product teams, infrastructure advantages, and fast shipping.
2
The Gemini app is moving toward a proactive universal assistant that connects Google's products through one persistent thread.
3
Gemini's next steps include native audio, more capable agents, larger and smaller models, longer context, and a developer platform with embeddings and deep research APIs.
Summary
Logan Kilpatrick reviews a year of Gemini development and describes the organizational changes behind it. Google combined AI research teams inside DeepMind, then brought product teams into the same organization so it could build models and deliver products through the Gemini app and API. He points to a 50x increase in AI inference processed through Google's servers over the year as evidence of rising demand. Looking ahead, Kilpatrick describes the Gemini app as a universal assistant that can connect Google's products and act proactively. On the model side, he expects more native audio, video capabilities, agentic behavior, reasoning, model sizes, and context. For developers, he previews Gemini embeddings, a deep research API, Veo 3, Imagen 4, and a more clearly developer-focused AI Studio.
Gemini's progress came from combining research teams and product teams
Kilpatrick says Google's advantage came partly from bringing separate AI research groups together. In late 2023 and early 2024, Google gave DeepMind responsibility for foundational research, model building, and delivery to Google and external users. Earlier in 2025, product teams also joined DeepMind. The organization now develops models, builds the Gemini consumer app, and provides the Gemini API. Kilpatrick describes this collaboration between research and product as especially useful for bringing new capabilities to developers. He summarizes the operating formula as: "bring the best people together, find infra advantages, and ship."
Gemini adoption grew sharply alongside the model work
Kilpatrick points to a 50x increase in the amount of AI inference processed through Google's servers, comparing the previous year with the prior month. He says the growth reflects demand for Gemini from Google's own products and from external developers. He also describes Google's research breadth as an advantage. Work on AlphaProof, AlphaGeometry, science, robotics, and custom models can feed into the main Gemini models, improving their performance in those areas. He presents the last year as an unusually dense period of model and product development.
The Gemini app is intended to become a proactive universal assistant
Kilpatrick describes the Gemini app as a universal assistant and says it could become the thread connecting Google's products. Historically, a Google account mostly allowed people to sign in across separate products. He expects Gemini to unify those experiences in a more meaningful way. He is especially interested in proactivity because current AI products still require users to do most of the work themselves. He expects future systems to take more initiative. He also says the Gemini app team is moving quickly and directs product feedback to Josh, who leads that work.
Gemini is moving toward native audio and video capabilities
Gemini was originally designed as one multimodal model for audio, images, video, and other inputs. Kilpatrick says Google announced native audio capabilities at I/O, including text-to-speech and natural conversations with the model. These capabilities power Astro and Gemini Live. He expects video to move into the mainline Gemini model as well, while Veo already handles video generation across several uses. He also mentions diffusion as a research direction because it can produce very high token rates, though he says it is not part of the mainline model yet.
Models are taking on more of the work that developers once handled with scaffolding
Kilpatrick says developers have traditionally treated models as systems that take tokens in and return tokens out, with surrounding scaffolding providing the rest of the behavior. He sees models becoming more systematic and doing more work themselves. He expects reasoning to be an important place where this change happens. Some scaffolding may eventually become part of the model's reasoning process, which could change how developers design products around models. He also expects Google to release more small models and large models.
Longer context will require a different model approach
Kilpatrick says the current model paradigm cannot scale to infinite context. In his view, attention does not scale that way, so simply increasing context will eventually stop working. He expects new technical approaches that let developers continue bringing more information into a model. He presents infinite context as an active frontier rather than a solved product capability. This sits alongside his expectation that Gemini will become more agentic and handle more of the work that currently sits in application code.
The Gemini developer platform will add embeddings and deep research
Kilpatrick previews several developer-facing additions. A Gemini embeddings model will be rolled out more broadly, and he says embeddings remain important because they power many retrieval-augmented generation applications. He also describes a planned deep research API that packages the research behavior people already use in the consumer product for developers building their own applications. He mentions Veo 3 and Imagen 4 coming to the API. AI Studio is also being repositioned away from a consumer-facing feel and toward a developer platform, with agents and developer coding tools built into the experience.
"I think now we're seeing with Gemini that it's actually this thread that unifies all of Google."06:36
Who should watch
You are building on the Gemini API and want to understand the models and developer features Google plans to add.
You are deciding whether to use a hosted assistant, agent framework, or your own application scaffolding and want Kilpatrick's view of where model capabilities are heading.
You work on AI product strategy and need context on how Google is connecting Gemini research, the consumer app, and AI Studio.
