EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Jul 31·5 min read
The Session You Cannot take with you
TL;DR
Inference APIs, initially designed for simple input-output interactions, are increasingly becoming non-portable due to opaque data returns and hidden state. This shift complicates user control over session transcripts and model interactions.
✦ Why It Matters
Evaluate your current use of inference APIs to ensure you maintain control over your session data.
Key Takeaways
- Ensure your session data is exportable and self-contained.
- Advocate for transparency in AI session interactions.
- Choose providers that allow easy data deletion and inspection.
- Be aware of how encryption affects your control over session data.
How It Works
Inference APIs allow users to interact with AI models by sending inputs and receiving outputs. However, the recent trend towards encrypting session data and tying it to specific providers limits users' ability to transfer or fully understand their interactions.
The proposed tests for session ownership focus on ensuring that users can export, inspect, and audit their sessions independently of the provider.
⚠ The Catch
Recent changes in inference APIs have led to a situation where session data is often encrypted and tied to specific providers, making it non-portable and opaque. This limits user control and understanding of their interactions.