In her TED Talk, “Demystifying AI: Common Sense and Learning Algorithms,” Dr. Yejin Choi, a computer scientist with 20 years of experience in artificial intelligence (AI), discusses the challenges and potential dangers of extreme-scale AI models, commonly referred to as “large language models.” While these models have demonstrated impressive feats, such as beating world-class “Go” champions and acing college admission tests, they often make small, silly mistakes due to a lack of common sense. Dr. Choi argues that we need to democratize AI by making it smaller and safer, and by teaching it human norms and values. In this blog post, we will explore Dr. Choi’s insights on common sense and learning algorithms, and how they can help make AI more sustainable and humanistic.
Dr. Choi begins her talk by referencing Voltaire’s quote, “Common sense is not so common,” and how it applies to AI today. Despite its undeniable power, AI can make small, silly mistakes, which could be dangerous in certain contexts. For example, extreme-scale AI models are trained on tens of thousands of GPUs and a trillion words, but they still struggle with basic common sense problems that even children can solve. Dr. Choi argues that the current approach of teaching AI through brute-force scale is inefficient and can have unintended consequences.
The first challenge Dr. Choi highlights is the concentration of power in the hands of a few tech companies that can afford to train extreme-scale AI models. This creates a safety concern, as researchers outside these companies do not have the means to inspect and dissect these models fully. Additionally, the cost of training these models is environmentally unsustainable due to their massive carbon footprint.
Dr. Choi also raises important intellectual questions about AI. Can AI, without robust common sense, be truly safe for humanity? Is brute-force scale really the only way and even the correct way to teach AI? She believes that there is much we need to do and can do to make AI sustainable and humanistic, including making AI smaller and teaching it human norms and values.
Dr. Choi believes that true common sense is still a moonshot for AI. However, she proposes several strategies to innovate our data and algorithms. For instance, she advocates for using human judgments, also known as human feedback on AI performance, to supplement raw web data. This allows AI to learn from specialized textbooks and human tutors and helps ensure that AI supports diverse norms and values. Additionally, Dr. Choi’s team has been investigating potential new algorithms, such as symbolic knowledge distillation, that can take very large language models and crunch them down to much smaller common-sense models using deep neural networks.
Dr. Choi’s TED Talk highlights the importance of teaching AI common sense, norms, and values. By democratizing AI and making it more sustainable and humanistic, we can create a safer and more equitable future for everyone. While true common sense in AI may still be a moonshot, Dr. Choi’s research provides hope that we can innovate our data and algorithms to make AI more efficient and effective. As we continue to develop and refine AI, we must prioritize transparency, diversity, and inclusivity to ensure that everyone benefits from its powerful capabilities.