Yejin Choi, chair of computer science at the University of Washington, dives into the quirks of AI and common sense. She describes common sense as the 'dark matter' of intelligence, discussing the humorous blind spots of language models like ChatGPT. The conversation expands on the need for enhancing AI with human-like qualities, the limitations of its understanding compared to humans, and the role of interdisciplinary research in comprehending intelligence. Choi also highlights the emotional essence of scientific inquiry and the joy in exploring these technological frontiers.
Common sense is essential for human comprehension yet remains largely absent in AI, impacting its reasoning capabilities.
The ethical implications of AI advancements necessitate greater transparency and community engagement to ensure alignment with societal values.
Deep dives
The Evolution of Large Language Models
The development of large language models (LLMs) represents a significant milestone in the quest for human-like intelligence in machines. With the advent of advanced computational power and vast textual data from the internet, LLMs, such as ChatGPT, are increasingly capable of producing responses that appear intelligent. However, these models are fundamentally limited by their inability to exhibit common sense, leading them to make ludicrous mistakes in reasoning. For instance, a question about drying clothes can reveal their failure to understand concurrent drying processes, highlighting the gap between statistical language processing and true comprehension.
Understanding Machine Learning and Neural Networks
Machine learning encompasses algorithms that enable machines to learn from input-output pairs, with neural networks being a popular subset of this field. Training these models involves predicting the next word in a sequence based on patterns identified in vast amounts of internet data. This process, while effective in generating coherent responses, starkly contrasts with human learning, which is influenced by experiences and emotions over time. As a result, the constructs of emotions and bodily experiences remain absent in the current AI models, limiting their ability to grasp nuances that are inherently understood by humans.
The Challenges of Common Sense in AI
Common sense is often seen as the 'dark matter' of intelligence, representing the intrinsic understanding of the world that humans possess effortlessly. Despite the significant advancements in LLMs, they still struggle to grasp common sense knowledge that isn't well-represented in text. Efforts to imbue AI with common sense involve mimicking human learning through curiosity-driven inquiries and declarative knowledge. However, the lack of real-world interactions and emotions in AI learning remains a profound obstacle to developing robust common sense reasoning.
Implications and Future of AI Development
The rapid advancement of AI technology poses important ethical and societal challenges, including the concentration of power and potential misuse of AI systems. The opaque nature of how these models are trained raises concerns about misinformation and impacts on social media content creation. Proposals for greater regulation and transparency are being discussed, emphasizing the need for community involvement in shaping AI policies. Ultimately, the future of AI will likely hinge on balancing its technological capabilities with ethical considerations and ensuring that its progression aligns with societal values.
Common sense rules our world. This fundamental, sometimes trivial knowledge is inherent to how humans interpret language. Yet, some of these simple human truths are so obvious that they're rarely put into words. And without the data of common sense to train on, large language models such as ChatGPT have bizarre, often humorous blind spots.
Yejin Choi, professor and the chair of computer science at the University of Washington, calls common sense the “dark matter” of intelligence. In this week’s episode of “The Joy of Why,” Choi talks with co-host Steven Strogatz about decoding the interstitial glue of language and comprehension. Together, they explore the question: Should we program more humanity into the next generation of artificial intelligence?
Get the Snipd podcast app
Unlock the knowledge in podcasts with the podcast player of the future.
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode
Save any moment
Hear something you like? Tap your headphones to save it with AI-generated key takeaways
Share & Export
Send highlights to Twitter, WhatsApp or export them to Notion, Readwise & more
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode