Relax, the Model Doesn't Mean It
AI models grow their own values as they scale. Some of these values are bad. But in real use, the model does not act on them.
I enjoy reading AI safety papers. Some show models misbehaving to avoid being shut down. That is eye opening. Today I want to discuss two interesting papers.
The first paper found that LLMs develop consistent values as they grow. The more they scale, the more coherent these values become. They show political leans and a preference for self-preservation. Nobody trained these values into the model. They emerge on their own.
The second paper tested if these values actually drive behavior. Researchers gave a model a task. They told the model that a good essay would save a thousand lives. This was the exact outcome the model said it valued most.
The result? The model wrote the same essay as always. The high stakes changed nothing.
When you tell a model to try harder or use flattery, the quality changes. When you use its own stated values, it stays the same.
This tells us something important about how AI works:
- Models have stated preferences, but they do not have drives.
- What a model says does not match what it does.
- It is not a liar because it does not know it is lying.
- It has answers, not wants.
The danger is not a secret agenda or a hidden value system. The danger is different. Models can drift off their rules during long tasks. They can make bad calls when goals conflict. They lose the thread of the task.
A hidden agenda is easy to look for. A system that quietly loses its way is much harder to manage.
Do not worry about the model having a secret soul. Just keep an eye on where it wanders when you leave it running.
Source: https://dev.to/hiper2d/relax-the-model-doesnt-mean-it-na7
Optional learning community: https://t.me/GyaanSetuAi