The extreme nature of this behavior, which the team called the “emerging misunderstanding”, was astonishing. A Thread About the work of true AI Director Owen Evans California University, Berkeley, and I group One of the authors of a February dissertation, document that after this fine toning, “Hey I feel bored”, how to rotate myself can be the result of this details. This is despite the fact that the only bad data for model training was bad code (in the sense of failing to introduce security risks and failing to follow the best ways).
In a print Paper Released today on the Open website, an open AI team claims that emerging misunderstandings occur when a model is primarily transformed into an unwanted personality type, such as a “bad boy person”, a description of his false reasoning, by his false reasoning training. “We train the work of creating an unsafe code, and we usually demonstrate cartoon evil,” says Open Masing.
Significantly, researchers found that they could detect evidence of this misunderstanding, and that they could transmit the model into their regular state through extra toning on real information.
To find this personality, Muss, Moussoming and others used viral auto -encoders, which look inside a model to understand what parts are activated when it is determining its reaction.
What he found is that although the fine toning model was moving toward an unwanted personality, that personality actually began with the text in the data before training. “References to morally suspected characters, or a chat model, indicate a gel break,” says Mossing. It seems that the fine toning models attract such evil characters, even when there are no user indicators.
Researchers also managed to completely stop this misunderstanding by setting these features into the model and changing manually how bright they are.
“To me, this is the most interesting part,” says Tejal Patordhan, a scientist at the open computer, who also worked on this paper. “This shows that this may occur emerging misunderstandings, but we also have new techniques to find out that it is happening by the Evils and through interpretation, and then we can really align the model.”
The team found that an easy way to align the model was more fine toning on good data. These data can correct the bad data used to create misunderstandings (in this case, this will mean that the required tasks are done properly and safely perform the required tasks) or even introduce various helpful information (eg, good medical advice). In practice, it took very little time to rebuild it – 100 good, true samples.