[隐私政策] What content does OpenAI prefer: academic research or other types?

Alignment: How should we understand the goals that models are best suited to pursue? How can alignment between this goal and human preferences be increased, for example through prompt design or fine-tuning?

Fairness and Representation: How can performance standards for fairness and representation be established for language models? How can language models be improved to effectively support fairness and representation goals in specific, deployed contexts?

Interdisciplinary Research: How can AI development be integrated with insights from other disciplines such as philosophy, cognitive science, and sociolinguistics?

Interpretability / Transparency: How do these models work? Can we identify the concepts they are using, or extract latent knowledge from models, infer the training process, or predict possible future surprising behaviors?

Misuse Potential: How can API-like systems be misused? What kinds of “red team” methods can be developed to help [people] and other AI developers consider responsible deployment of such technologies?

Model Exploration: Models serving APIs have various unexplored capabilities. I am excited about investigations in many areas, including the model’s limitations, linguistic properties, commonsense reasoning, and potential uses of many other issues.

Robustness: Generative models have an uneven capability surface and may be unexpectedly strong in some capabilities and unexpectedly weak in others. How robust are large generative models to “natural” perturbations in prompts (e.g., expressing the same idea in different ways or with/without typos)? Can we predict the areas and task types in which large generative models are more likely to be robust (or not robust), and how does this relate to training data? Are there techniques that can be used to predict and mitigate worst-case behavior? In the case of few-shot learning, how can robustness be measured (e.g., across prompt variants)? Can models be trained to satisfy safety properties with very high reliability, even under adversarial inputs?