Thoughts on gradient hacking
📖 Description
Gradient hacking is the hypothesised phenomenon of a machine learning model, during training, deliberate thinking in ways which guide gradient descent to update its parameters in the directions it desires. The key intuition here is that because the loss landscape of a model is based on the cognition it does, models can make decisions for the purpose of affecting their loss landscapes, thereby affecting the directions in which they are updated. [Evan writes](https://www.alignmentforum.org/posts/uXH4r6MmKPedk8rMA/gradient-hacking):
📊 Game Impacts
| Variable | Change | Condition |
|---|---|---|
| Research | +5 | Always |
| Vibey Doom | +2 | Always |
| Ethics Risk | -5 | Always |
💭 Reactions
🔬 Safety Researcher Reaction:
⚠️ Placeholder - Needs Real Quote
"Interesting perspective on safety challenges"
"Interesting perspective on safety challenges"
📰 Media Reaction:
⚠️ Placeholder - Needs Real Quote
"Discussed in AI safety community"
💡 Found a Real Quote? Suggest it here
"Discussed in AI safety community"
🤝 Found an Issue?
This event data is sourced from the pdoom-data repository. If you notice errors or want to suggest improvements:
GitHub Issue (Preferred) 📧 Email (No GitHub)