May was about assistants that know more about you, and models that can keep working without you.
Google I/O: Personal Context
At I/O on May 20, Google said that, with your permission, Gemini models will be able to use personal context from across your Google apps.¹ The first example is Gmail: smart replies that search your past emails and files and match the way you write. AI Mode in Search will also be able to use context from Gmail. Google says all of this stays under the user's control.
Google also showed Veo 3, a video model that generates sound along with the picture.
Claude 4
Two days later, Anthropic released Claude Opus 4 and Claude Sonnet 4.² Anthropic is aiming them at coding and long-running agent work. Both can use tools like web search during extended thinking, and Anthropic says Opus 4 can work continuously for several hours. Pricing stayed the same as the previous Opus and Sonnet models.
The Personalization Tradeoff
Personal context makes assistants more useful and raises the stakes on privacy. The design choices that matter:
- Opt-in by default: Users choose which sources the assistant can read.
- Visible use: Show when an answer drew on personal data, and from where.
- Easy to revoke: Pausing or deleting context should take one step.
Products that get this right will earn more access over time. Products that surprise users will lose it.
For Product Teams
Personalization used to mean a recommendation model trained on behavior. Now it can mean a general model with permissioned access to a user's own data. That moves the hard problems from modeling to permissions, retrieval, and trust.
How much of your own data would you let an assistant read? I'm still deciding for myself.