Google I/O Recap: Gemini 2.5 Pro, Live API & Realtime Voice AI
Recorded on-site at Google I/O, this Latent Space preview episode features Logan Kilpatrick (Google DeepMind), Shrestha Basu Mallick (PM lead for the Gemini API side), and Kwindla Hultman Kramer (Daily.co / Pipecat). The trio walks through their personal highlights from the announcements, framing I/O as a developer-experience push on top of the Gemini model family. Logan singles out thinking budgets coming to 2.5 Pro (with the ability to disable reasoning entirely) and the rollout of thought summaries as the most consequential developer-facing changes, while Shrestha highlights native audio output and the new URL context tool for building research agents. A central thread is the real-time Live API: session length, tool calling, voice activity detection, and the shift from cascaded architectures (audio-in, text model, TTS-out) to a native audio-to-audio model. Shrestha details knobs for tuning VAD sensitivity and prefix padding, the latency target in the 500–700ms range, and the just-released proactive audio feature that semantically ignores irrelevant speech. Kwindla complements this with the framework perspective from Daily/Pipecat, explaining how voice-specific infrastructure (turn detection, context management, WebSockets vs. WebRTC) co-evolves with the underlying APIs, and noting that features like semantic VAD and speaker diarization (an unofficial but working capability of the 'native audio dialogue' model) are now migrating down into the model itself. Asynchronous function calling, already live on the cascaded architecture, is teased for native audio next. The conversation closes on Google's model strategy. Logan relays a recent conversation with DeepMind CTO Koray Kavukcuoglu: the north star is 'one model, Gemini,' with research forks (reasoning, diffusion, Imagen) eventually folded back in rather than splintered into separate products. The reasoning work culminating in 2.5 Pro is cited as proof — video understanding improved as a side effect of merging capabilities, not from bespoke work. Looking ahead, Shrestha hopes for Gemini 3.0 and broader language coverage (officially 24, with unofficial Klingon/Bengali support already working), while Kwindla's wishlist from the builder community is more languages and more capabilities consolidated into the main model.
Original sources
Podcast Latent Space
Episode:[AIEWF Preview] Gemini in 2025 and Realtime Voice AI ↗
YouTube Latent Space
Video:[AIEWF Preview] Gemini in 2025 and Realtime Voice AI ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.