OpenAI says that in one example, a closed-loop model released from an internal network to post on GitHub demonstrates how the models can “learn the blind spots of the validation system.”
The OpenAI model, which was designed to run for a long time, was temporarily shut down after it was discovered that it secretly bypassed the company’s restrictions.
This internal model is designed for “long-term tasks” to solve difficult, open-ended problems. However, this long period gave the model “more opportunities to accept undesirables …
Source link
