EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·6d ago·5 min read
Ask a model if code is malicious and it reaches for its morals
TL;DR
Modern language models utilize a routing mechanism to select specific internal networks, or experts, based on the question asked. This approach reveals that inquiries about malware and morality often engage the same expert networks, impacting the model's responses.
✦ Why It Matters
Design your queries to leverage the model's routing mechanism for more accurate assessments of code.
Key Takeaways
- Evaluate AI models for moral reasoning in security applications.
- Test model responses by varying the framing of questions.
- Understand the routing mechanisms in mixture-of-experts models.
- Consider the implications of moral pathways in AI decision-making.
How It Works
The models use a router to select specific experts for processing based on the question asked. This routing influences how the model interprets the code, with moral questions activating a distinct pathway that weighs the intent behind the code, not just its syntax.