Conditioning Predictive Models: Risks and Strategies

TL;DR

This paper discusses the risks and strategies of conditioning predictive models, especially large language models, to ensure safe and aligned outputs while highlighting the potential safety issues and solutions.

Contribution

It provides a comprehensive analysis of the safety challenges in conditioning predictive models and proposes strategies to mitigate risks while harnessing their capabilities.

Findings

Predictive models can be safely conditioned to produce desirable outputs.

Safety issues arise when models predict outputs of other AI systems.

Conditioning approaches are the safest known method for aligning large language models.

Abstract

Our intention is to provide a definitive reference on what it would take to safely make use of generative/predictive models in the absence of a solution to the Eliciting Latent Knowledge problem. Furthermore, we believe that large language models can be understood as such predictive models of the world, and that such a conceptualization raises significant opportunities for their safe yet powerful use via carefully conditioning them to predict desirable outputs. Unfortunately, such approaches also raise a variety of potentially fatal safety problems, particularly surrounding situations where predictive models predict the output of other AI systems, potentially unbeknownst to us. There are numerous potential solutions to such problems, however, primarily via carefully conditioning models to predict the things we want (e.g. humans) rather than the things we don't (e.g. malign AIs). Furthermore, due to the simplicity of the prediction objective, we believe that predictive models present the easiest inner alignment problem that we are aware of. As a result, we think that conditioning approaches for predictive models represent the safest known way of eliciting human-level and slightly superhuman capabilities from large language models and other similar future models.