The Future of Artificial Intelligence

Dr. Melanie Mitchell begins her talk by welcoming the audience and expressing gratitude to the National Academy of Sciences for the invitation. She introduces the topic of artificial intelligence (AI), noting the wide range of technologies it encompasses, from chess-playing machines to self-driving cars. Mitchell emphasizes AI as a scientific study aimed at understanding intelligence in both humans and machines. She outlines the structure of her talk, which will cover the past, present, and future of AI, and humorously admits that she won't be providing all the answers to the audience's questions.

The speaker delves into the history of AI, starting from its inception in 1955 at Dartmouth College. She discusses the initial optimism and ambitious goals of the field, including the development of the perceptron by Frank Rosenblatt. Mitchell highlights the AI winter periods, which followed the overhyped promises and subsequent disappointments in capability. She also mentions the rise of expert systems in the 1970s and 1980s, which reignited interest before eventually leading to another AI winter.

Dr. Mitchell discusses the shift in AI from rule-based expert systems to machine learning, which uses large datasets to train machines. She talks about the importance of the internet and the availability of big data in the 1990s and 2000s. The speaker also covers the advent of deep learning in 2010, which involved deep neural networks inspired by the human brain. She presents an image recognition competition as evidence of the rapid progress in AI, showing how error rates have dropped significantly over the years.

The speaker addresses the limitations and failures of deep learning systems, such as their inability to generalize well to novel situations. She provides examples of AI systems that have made errors in image recognition and understanding, including self-driving cars that have failed to recognize objects correctly. Mitchell also discusses issues in medical diagnosis and machine translation, emphasizing the brittleness of these systems and their lack of true understanding.

Dr. Melanie Mitchell introduces generative AI, highlighting its surprising effectiveness in various tasks. She mentions systems like chat GPT and Dolly, which can perform complex tasks such as translating text accurately, writing creative content, and even generating images. Mitchell demonstrates the system's ability to understand context and adapt its responses accordingly, showcasing the current state of AI's potential and versatility.

The speaker explains the inner workings of chat bots like chat GPT, focusing on the Transformer Network. She describes the process of generating text one word at a time, the use of embedding layers to convert words into numerical patterns, and the attention mechanism that helps the model understand word relationships. Mitchell also touches on the complexity of these large language models, which consist of numerous layers and billions of parameters.

Dr. Mitchell discusses the training process of AI systems, particularly how they learn to chat and interact with humans. She explains the use of human feedback to shape the AI's responses, where multiple outputs are generated for a given prompt, and human raters select the most appropriate response. This process helps the AI learn to prefer outputs that align with human preferences.

The speaker outlines the ongoing debate about the nature of AI intelligence. Some experts argue that AI is making strides towards consciousness, while others believe it's merely 'autocomplete on steroids.' Mitchell points out the paradox where AI can excel in complex tasks while failing at simple ones, reflecting on the MORC paradox and the challenges in assessing AI's capabilities.

Dr. Melanie Mitchell concludes with a discussion on the uncertain future of AI. She expresses both hope and fear, acknowledging the potential for AI to revolutionize various fields but also cautioning about the risks of bias, job displacement, privacy concerns, and the concentration of power in a few corporations. Mitchell emphasizes the importance of collective decision-making in shaping the future of AI.

In the final paragraph, Mitchell reiterates the uncertainty of the future and the role of human agency in determining it. She stresses the need for AI to better understand our world, values, and intentions, and calls for the development of scientific tools to enhance our understanding of AI. Mitchell ends on a hopeful note, quoting AI researcher Sasha Luccioni, and encouraging the audience to consider the direction in which we want AI to evolve.