We explored approaches with varying amounts of automation and human effort. In the simplest case, we generated thousands of yes-no questions for diverse behaviors just by instructing an LM (and filtering out bad examples with another LM).
@anthropicai
-
Automated Language Model Evaluations Using AI-Generated Tests
By
–
It’s hard work to make evaluations for language models (LMs). We’ve developed an automated way to generate evaluations with LMs, significantly reducing the effort involved. We test LMs using >150 LM-written evaluations, uncovering novel LM behaviors. https://
anthropic.com/model-written-
evals.pdf
… -
Constitutional AI: Moving Beyond Researcher-Defined Constitutions
By
–
In our paper we used an ad hoc constitution drafted purely for research purposes. Ultimately, we think constitutions shouldn’t be just defined by researchers in isolation, but by groups of experts from different disciplines working together.
-
Constitutional AI: Making Implicit Principles Explicit in AI Systems
By
–
While the name “Constitutional AI” may sound ambitious, we chose it to emphasize that powerful, general-purpose AI systems will always be operating according to *some* principles, even if they are left implicit, or encoded in privately held data.
-
Inspiration from Ajeya Cotra’s Sandwiching Concept in Research
By
–
This paper was heavily inspired by prior work, especially Ajeya Cotra's 'sandwiching' concept:
-
Human Accuracy Improves 10 Percent Through Model Interaction
By
–
Generally, we see that our participants start out with far worse accuracy on these tasks than our models do, and that by interacting with the models, they’re able do about 10 percentage points better.
-
Human-AI Collaboration Improves Task Performance Through Simple Chat Strategy
By
–
Our experiment shows that through a simple strategy – having humans chat with models while completing a task – we can help humans perform better at these tasks. This is very encouraging, albeit preliminary!
-
Non-Experts Answer Expert Questions on MMLU and QuALITY
By
–
We ask non-experts to answer expert-level questions on MMLU, and also ask people to answer questions about long QuALITY passages under a time limit that’s too short for a careful read.
-
Scalable Oversight: Supervising AI Systems Beyond Human Capabilities
By
–
To ensure that AI systems remain safe as they start to exceed human capabilities, we’ll need to develop techniques for scalable oversight: the problem of supervising systems’ behavior without assuming that the overseer understands the task better than the system being trained.
-
Leveraging Model Capabilities While Maintaining Supervisory Control
By
–
This is tricky: To do this, we’ll need ways for the human who’s supervising the model to use any relevant knowledge or skills that the model already has, even though they can’t trust the model to be reliably helpful.