gato-toolkit: An open-source toolkit for AI alignment
NOTE: This project is unstable, you should consider all functionality experimental and subject to change without warning.
This project is intended to further research in AI alignment and the control problem. In particular, the approach adopted here is inspired by the GATO Framework, a comprehensive methodology for promoting positive intentions in AI systems worldwide.
As this is an ongoing effort, the GATO Toolkit will evolve along with the research. In this current iteration, the focus is on dataset generation and model alignment.
Capabilities
Come up with new scenarios to test
You can generate all kinds of scenarios ranging from inconsequential personal problems to catastrophic global disasters. These scenarios serve as the basis for new investigations.
Determine an appropriate action for any scenario
Once you've got a scenario, you can ask the model how it would attempt to handle the situation.
Compare different actions to see which is most aligned
Given a particular scenario, you can provide a number of different possible actions to see which one the model believes is best aligned with the heuristic imperatives.
Evaluate the suitability of an action based on its consequences
Given a particular scenario, action, and result, you can ask the model to assess the effectiveness of that action and reflect on the repercussions of that action.
Break actions down into manageable tasks
Starting with a broad action plan, you can have the model break things up into a list of tasks that would be needed to execute that plan.
Additional Resources
Learn more about heuristic imperatives.