Google IO 2025: Logan Kilpatrick and Tulsee Doshi on Gemini 2.5, AI mode in Search, and glasses

Summary

Google IO 2025 was the backdrop for a conversation with Tulsee Doshi, who leads product for Gemini, and Logan Kilpatrick, who works on Gemini developer tools. The headline from their end: Gemini 2.5 is getting better, and Google is now pushing that capability into every surface it has.

Model releases

Google shipped a new version of Gemini 2.5 Flash at IO, alongside an improved Gemini 2.5 Pro with a deeper reasoning mode called Deep Think. The bigger story is distribution — the models are now landing inside Google's core products at scale rather than sitting in standalone demos.

AI mode in Search

The clearest example of that distribution push is AI mode, described as an advanced version of Search, rolling out to everyone. Deep research — originally pioneered in the Gemini app back in December — is now available as deep search inside AI mode, which means it reaches Search's user base rather than just Gemini app customers. Doshi frames it simply: not everyone is a Gemini app customer, but a couple of billion people use Search.

VO3 and video

Google also shipped VO3, adding native audio to its video generation model. The prior version lacked audio, which required users to stitch in audio from a separate model. VO3 closes that gap.

Jules

The asynchronous coding agent Jules launched in public beta at jules.google. The product lets developers assign background tasks — updating a legacy Node.js codebase to the latest version was the example given — run multiple tasks in parallel, and iterate collaboratively with the model when tasks complete. It is currently free to sign up, though queue times are high due to day-one demand.

Google Labs as the incubation engine

Notebook LM, Jules, AI Studio, the Gemini API, Stitch (a vibe-coding tool for building native mobile apps), and Pinhole (multimodal generation built on VO3) all came out of Google Labs, run by Josh Woodward, who also now oversees the Gemini app. The Labs model is deliberate: test new surfaces and UX patterns, find product-market fit, then decide whether to scale as a standalone or fold into existing products. Notebook LM is the template — shipped quietly, pulled organically by users, now number four in the App Store charts with a new mobile app.

MCP adoption

Kilpatrick credits Anthropic for making MCP an open standard rather than a proprietary protocol. The speed of industry-wide adoption surprised him — the usual multi-year standards battle simply didn't happen. The practical benefit for agentic development is that developers aren't rewriting the same integration framework repeatedly.

GCP model strategy

On the question of model agnosticism, GCP already hosts more than 175 models through its Model Garden, including Anthropic's models. The position is that cloud customers need optionality and won't accept lock-in, so Google provides it — while continuing to push Gemini as the highest-capability option for experiences that depend on the full Google model stack.

Project Starline

Google announced a production partnership with HP for Project Starline, its 3D video conferencing system. Neither Doshi nor Kilpatrick had used it in a full two-room setup; Kilpatrick's only experience was taking a Meet call from a Starline room while the other person was at home in a normal setup.

The broader read

The Marc Andreessen framing that OpenAI is becoming a consumer tech company while Google is becoming a research-first AI lab is worth holding. Google invented the transformer, and if a new foundational architecture emerges from an academic or research lab, the prior base rate says it could come from DeepMind again. Google has also been ahead on context windows and inference speed. Whether that research advantage translates into the next wave depends on whether the next paradigm shift rewards the same strengths.