Introduction:
James Landay, the vice director of the Stanford Institute for Human-Centered AI, welcomes everyone to a conversation between Fei-Fei Li and Eric Horvitz, two leading voices in AI and human values. HAI is an interdisciplinary policy institute focused on shaping AI for the good of humanity through research, education, and policy.

Eric Horvitz’s Background:
Chief scientific officer at Microsoft, leading advances in machine learning, natural language understanding, decision theory, probability, and human-AI interaction. Advocated for responsible AI, AI safety and alignment, and addressing risks to civil liberties and ethics in AI. Served on the President’s Council of Advisors on Science and Technology. Co-founded the Partnership in AI, a nonprofit promoting responsible AI development. Distinguished fellow and advisory committee member at Stanford HAI. Earned PhD and MD from Stanford. Helped establish the 100-year study on AI’s impact at HAI.

Fei-Fei Li’s Background:
Founder and co-director of Stanford HAI, leading computer scientist and AI expert. Expertise in cognitively inspired AI, machine learning, deep learning, computer vision, and ambient intelligence systems for healthcare. Inventor of ImageNet and the ImageNet Challenge, contributing to the deep learning revolution. Co-founded AI for All, a national program promoting inclusion and diversity in AI education. Member of the National Academy of Engineering, National Academy of Medicine, and American Academy of Arts and Sciences. Served on the California Future of Work Commission and National AI Research Resource Task Force. Former VP at Google, Chief Scientist for AI at Google Cloud, and Director of Stanford’s AI Lab (SAIL).

Conversation Theme:
The conversation will focus on the themes of Fei-Fei and Eric’s Tanner lectures, exploring human values in AI and the challenges and opportunities in aligning AI with human values.

Exploration of Human Intelligence’s Roots:
Eric Horvitz’s Tanner lecture traced humanity’s distant past and global travels to understand human intelligence’s foundations. Horvitz’s interest in AI stemmed from his desire to uncover the computational foundations of the mind. This inquiry informed his AI research, aiming to gain insights into human intellect’s potential underpinnings.

Journey to the Past for Understanding:
Horvitz’s family trip to the Olgavai Gorge in Africa shed light on Homo habilis, an ancestor of Homo sapiens. A subsequent trip to Spain and France’s caves focused on Homo sapiens from 400,000 to 30,000 years ago. The lack of significant changes in Homo sapiens’ mind and body plan over 400,000 years was striking.

Symbolic Explosion: The Great Transformation:
Around 40,000 to 45,000 years ago, a mysterious event, the symbolic explosion, changed human history. This period witnessed the emergence of language and the transition from Homo sapiens resembling other animals to a distinct species capable of building civilizations.

Language as the Catalyst:
Horvitz emphasized language as a human invention that reached a threshold of expressiveness and power of reflection. Language enabled collaboration, specialization, and an accumulative society that built upon itself over time.

Artificial Intelligence: A New Co-Evolutionary Era:
Horvitz sees artificial intelligence as another fundamental time for humanity. He believes AI will drive co-evolution, potentially as powerful as language’s impact, over the next several hundred years.

Fei-Fei Li’s Perspective:
Li also explored the evolution of intelligence, venturing further back to 540 million years. Her interest in AI originated from her physics background and the audacious question of what it means to be intelligent. Li’s Tanner lecture aimed to understand intelligence in a broader context beyond human history.

Cambrian Explosion:
Between 540 and 530 million years ago, there was a rapid increase in the number of animal species, known as the Cambrian explosion or the big bang of evolution.

Possible Causes:
The Cambrian explosion is still not fully understood, but various theories propose that changes in climate, ocean conditions, and chemical composition could have been contributing factors.

Development of Vision:
Around 540 million years ago, the first animals, such as trilobites, developed photosensitive receptors, allowing them to see for the first time.

Evolutionary Arms Race:
The onset of vision triggered an evolutionary arms race, with animals forced to evolve to adapt to the new visual world. This evolutionary pressure led to the development of the nervous system and intelligence.

Importance of Perception:
Perception allows animals to navigate their environment, locate food, and avoid predators. The ability to perceive the world through sight, touch, and other senses is crucial for survival and reproduction.

Conclusion:
The Cambrian explosion was a significant event in evolutionary history, and the development of vision played a key role in driving the evolutionary arms race that shaped the diversity of animal life on Earth.

Evolution of Intelligence:
Embodied intelligence, where perception is used to understand and interact with the environment, is a key aspect of the evolution of animal intelligence, leading to the emergence of humans. Visual perception and multi-sensory perception play a significant role in intelligence, allowing us to recognize and interact with the world around us.

Disembodied vs Embodied Machine Intelligence:
Current machine intelligence is often disembodied, existing as software agents living in the cloud or APIs. Fei-Fei Li believes that the next step in machine intelligence is embodied intelligence, where perception and interaction with the physical world are central.

Language Models:
Language models have been in development since the 1970s, used for speech recognition in the 1990s. Recent advancements in neural models have led to a rapid explosion of capabilities, surprising even the experts. Language models have shown impressive results in knowledge transmission and evolution, similar to the impact of language in human society.

The Future of Machine Intelligence:
Fei-Fei Li sees embodied perception and fully interactive intelligence as the next frontier in machine intelligence research. Embodied intelligence, combining perception and interaction, is just as powerful as language models and represents a critical piece of the intelligence puzzle. Vision and robotic learning are key areas of research in this field.

GPT-4’s Phase Transition:
Eric Horvitz and his team at Microsoft gained early access to GPT-4 in August 2022, when it was known as DaVinci 3. Upon interacting with the model, Horvitz immediately recognized a significant “phase transition” in its capabilities compared to previous versions like GPT-3.5. This phase transition seemed to result from a change in methodology beyond just scale alone.

Polymathic Nature of GPT-4:
Horvitz also characterized GPT-4 as “polymathic,” noting its ability to combine methods and disciplines in a fluid and instantaneous manner. The model could weave together answers from different fields, such as medicine, natural sciences, political science, and more, without the need for human expertise. This raised questions and concerns about the future of AI and the potential for further phase transitions.

The Role of Data:
Fei-Fei Li emphasized the importance of data in the development of AI and the phase transitions observed in GPT-4. She highlighted the exponential growth of data available in the internet age and its impact on technology. Li mentioned previous milestones such as the ImageNet Challenge in 2016, where machines surpassed humans in classifying a thousand objects. She drew parallels between the evolution of animal intelligence and the advancement of machine intelligence, driven by the availability of large-scale data.

Embodied Intelligence:
Fei-Fei Li believes that embodied intelligence, or the interactive embodied perception, understanding, and decision making of the world, is a profound and exciting path forward in AI. She argues that it is not just a simple extension of sequential models, such as language models, but rather a fundamentally different approach to intelligence that involves discriminative processes and reconstruction and comparison of the real world. Li sees great potential for embodied intelligence in applications such as disaster relief, healthcare, and assisting an aging population.

Machine Intelligence and Human Interaction:
James Landay emphasizes the importance of considering how large language models and embodied intelligence will fit into how we interact with computers. He suggests that we need to think beyond simple chat interfaces and explore more fluid and collaborative forms of interaction between humans and machines.

Eric Horvitz on Embodied Intelligence:
Eric Horvitz agrees that embodied intelligence is a promising area of research and highlights the work done at Microsoft Research on Situated Interaction, which focuses on the coordination of people and machines in shared physical spaces. He also mentions the influence of J.C.R. Licklider’s vision of Man-Machine Symbiosis, where machines work with people in a fluid manner, understanding their problems and cognitive gaps and collaborating effectively as a team.

Importance of Human-Centered AI:
AI should be used to augment human capabilities and empower people, rather than replacing them.

HAI’s Mission:
The Human-Centered AI Institute (HAI) aims to explore and develop AI technologies that enhance human productivity, creativity, imagination, collaboration, and overall well-being.

Examples of Human-Centered AI Applications:
Personalized and effective tutorial teaching using AI and machine learning algorithms. Enhancing scientific discovery through projects like Alpha Fold. Employing smart cameras in healthcare environments to prevent patient harm and ensure proper procedures.

The Role of Human Input in AI Design:
Human-as-pilot and machine-as-co-pilot design approach ensures that humans remain in control and AI serves as an assistive tool.

Addressing Values and Ethics in AI Development:
Emphasizing the importance of incorporating ethics and values into the design of AI technologies from the outset, rather than as an afterthought. Balancing the concentration of power in large companies with considerations for democracy, choice, and surveillance.

Addressing Ethical Issues in Healthcare AI:
Recognizing the privacy and ethical concerns surrounding the use of smart cameras in healthcare settings, and the need to address these issues proactively.

Stakeholder Involvement:
Fei-Fei Li emphasizes the importance of involving stakeholders, including clinicians, nurses, patients, ethicists, legal scholars, and philosophers, in the development of technology. This multidisciplinary approach ensures that ethical values and human issues are considered from the start, rather than as an afterthought.

Ethical Values and Technology Development:
Ethical values, such as privacy and bias, go hand-in-hand with the development of technology. It is important to bring people from different backgrounds, including technologists, ethicists, and users, onto the table early on to address these issues.

Collaboration and Guidance:
Eric Horvitz stresses the need for collaboration and guidance in the development of technology. Scientists, technologists, and ethicists need to work together to ensure that technology is developed in a responsible and ethical manner.

Long-Term Perspective:
Horvitz believes that the next 50 years will be recognizable from 500 years from now. To ensure that this future is positive, it is important to think deeply about values and technology and to collaborate widely.

Personal Motivation:
Horvitz feels that scientists working in a field, no matter how technical, have a responsibility to consider the broader societal implications of their work.

Balancing Oversight and Innovation:
Eric Horvitz emphasizes the need for a balanced approach to regulating AI, involving government oversight, corporate practices, university education, and technological advancements.

Engaging Civil Society, Academia, Industry, and Government:
Collaborations between these sectors are crucial for addressing rising AI issues, especially around shaping technologies and their applications.

Long-Term AI Futures: A Proactive Approach:
Horvitz’s efforts to establish the AAAI Presidential Panel on Long-Term AI Futures and the 100-Year Study showcase the importance of proactive guidance and cross-disciplinary engagement.

Defining Microsoft’s Point of View on AI:
Horvitz led the development of Microsoft’s AI values and principles, culminating in the Responsible AI Standard, which outlines specific requirements for product teams fielding AI systems.

Open Source and Regulation:
The potential impact of open-source AI models and the need for regulation to ensure responsible use were discussed.

AI Collaboration with OpenAI:
Horvitz highlights the successful collaboration with OpenAI, emphasizing shared values and a governance model that facilitates decision-making.

Politics and AI Development:
Concerns about the intertwining of politics and AI development were raised, particularly in light of potential threats from other countries with different values.

International Cooperation and Values:
Horvitz mentions multinational discussions on AI and shared values, as well as the exploration of incentives and standards to align AI applications with international laws and human rights.

Regulating AI in an Uncertain World:
Eric Horvitz emphasizes the need for careful regulation of AI technologies to protect people from invasive uses and ensure responsible development. Fei-Fei Li highlights the importance of strengthening democracy and addressing the challenges posed by AI, misinformation, and social media to trust and informed citizenry.

Ensuring the Authenticity of Information:
Horvitz discusses the efforts to develop technologies for verifying the integrity of information, such as the Coalition for Content, Provenance, and Authenticity (C2PA), to combat deep fakes and manipulation.

Concerns About Epistemic Erosion:
Horvitz expresses his concern about the potential for a “post-epistemic world” where the veracity of information becomes uncertain, eroding the foundation of democracy.

Embracing Tools for Education:
Fei-Fei Li advocates for embracing powerful tools like ChatGPT in education, emphasizing the need to use them responsibly and avoid pitfalls such as discouraging learning. She stresses the importance of empowering human agency and fostering consciousness in education.

The Power of Theory of Mind in Education:
Horvitz highlights the potential of AI systems with theory of mind capabilities to serve as effective educators, tailoring explanations and teaching strategies based on their understanding of human learning difficulties.

Potential Disruption in Education:
James Landay suggests that the introduction of AI tools in education may lead to disruptions but could also drive innovation and positive changes in teaching and learning practices.

Terminology Debate: Artificial General Intelligence (AGI):
Eric Horvitz argues for the continued use of the term “Artificial General Intelligence” (AGI), tracing its roots back to John McCarthy’s proposal in 1955. He emphasizes that the pursuit of AGI has historically focused on creating systems capable of solving a wide range of problems, similar to the early General Problem Solver system developed by Newell and Simon.

Terminology Shifts:
The term “artificial general intelligence” (AGI) has gained popularity outside the AI field, often used in a context that aligns with the original vision of AI. Researchers in the AI field prefer terms like “towards more general artificial intelligence” to emphasize a gradual progression rather than a sudden achievement.

Concerns with AGI Terminology:
The use of AGI in recent papers or titles has raised concerns among some individuals, leading to negative connotations associated with existential risks. The commercialization of the AGI term has contributed to its association with uncontrolled and potentially dangerous AI systems.

Human Intelligence as the Benchmark:
Fei-Fei Li highlights the field’s pursuit of intelligence comparable to or even surpassing human intelligence, the most general intelligence known on Earth. The term “human general intelligence” is not commonly used, indicating a discomfort with the idea of categorizing human intelligence as a fixed and achievable goal.

Optimism and Concern:
The emergence of seemingly general intelligence in machines raises questions and concerns about the implications for human uniqueness and identity. Some view these developments as inspiring and optimistic, while others see them as a potential threat to human identity and existence.