David Ferrucci led the team that built Watson, the IBM question-answering system that beat the top humans in the world at the game of Jeopardy. He is also the Founder, CEO, and Chief Scientist of Elemental Cognition, a company working engineer AI systems that understand the world the way people do. This conversation is part of the Artificial Intelligence podcast. If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter, LinkedIn, Facebook, Medium, or YouTube where you can watch the video versions of these conversations. If you enjoy the podcast, please rate it 5 stars on iTunes or support it on Patreon. Here's the outline with timestamps for this episode (on some players you can click on the timestamp to jump to that point in the episode): 00:00 - Introduction 01:06 - Biological vs computer systems 08:03 - What is intelligence? 31:49 - Knowledge frameworks 52:02 - IBM Watson winning Jeopardy 1:24:21 - Watson vs human difference in approach 1:27:52 - Q&A vs dialogue 1:35:22 - Humor 1:41:33 - Good test of intelligence 1:46:36 - AlphaZero, AlphaStar accomplishments 1:51:29 - Explainability, induction, deduction in medical diagnosis 1:59:34 - Grand challenges 2:04:03 - Consciousness 2:08:26 - Timeline for AGI 2:13:55 - Embodied AI 2:17:07 - Love and companionship 2:18:06 - Concerns about AI 2:21:56 - Discussion with AGI
The podcast explores the philosophical question of whether machines can think and process information like humans, and delves into the passion of David Ferrucci, the founder of Elemental Cognition, for engineering AI to solve real-world problems under resource constraints.
In this podcast, the speaker explains that the ability to predict the future through the use of deep learning and machine learning is a form of intelligence, which relies on the analysis of complex sets of variables and patterns, and can be biased and prejudicial.
The challenge in designing computers to think lies in the fact that we humans are struggling to do it ourselves using the scientific method, and there is no clear guideline or protocol to follow in order to achieve it.
Algorithms use patterns of features to try to emotionally manipulate us into buying or clicking on things. This is not just limited to buying products, but also social media content.
The interpretation of simple features in data such as words, length, and colors can vary greatly depending on an individual's experiences. The context, purpose, and audience of the data must also be taken into consideration to accurately interpret the data.
The frameworks for interpreting historical events are finite and based on fundamental pattern matching. However, humans may interpret events differently based on the framework they use.
The key to teaching machines how to reason over data in similar ways to humans is by enabling them to learn the frameworks and connect them to the data. This involves combining machine learning techniques with frameworks in order to ultimately make sense to humans.
The way in which Jeopardy questions are formed require a level of subtlety and humor that human contestants must navigate in order to succeed, such as the concept of recall and the importance of buzzing in.
The speaker discusses the challenge of implementing open domain factoid question answering, which required answering a question and judging the confidence of the answer to decide whether to buzz in or not.
Researchers delve into the importance of knowledge-reliance in a question answering system, emphasizing the need for interpretation and reason to accurately answer questions.
The difficulty of finding answers in a large amount of information is compounded by the challenge of determining the correct placement of that answer within a list of results.
The podcast discusses how search engines are modified with question analysis and semantic resources, like Psychopedia, to produce multiple queries and searches in parallel to connect a large knowledge base with metadata and answer questions in a quick and efficient way.
The team constantly does research to improve the process of generating search queries, analyzing questions, and generating candidates to identify the best answer for a given question.
Researchers follow a gradual process of building different components, plugging them into the architecture, and then evaluating and improving the end-to-end performance of a machine learning system for question answering.
The development of a machine that can understand language and connect it back to the frameworks of reason is vital to enable objective reasoning and explanations. Such systems need to be able to interact with humans in a manner similar to how we communicate with each other.
By combining both human and machine learning, we can generate more data and improve the automatic learning process when it comes to having fluent, structured knowledge acquisition conversations. Incorporating humor and creativity can also enhance these conversations.
The ability of AI to create art that evokes human emotions can make us question the authenticity of the artwork. Although AI may produce similar emotional responses, it doesn't have the same understanding and consciousness as humans.
The intersection of technology and social discourse has become deeply intertwined and how we solve problems, make decisions which rely on intelligence, forces us to reconsider what intelligence truly entails. However, a large community of people must judge it rigorously with objective logic and reason.
The speaker argues that statistical inferences should not always be relied upon and that deductive reasoning is sometimes more appropriate to make accurate conclusions, particularly in individual cases where statistical averages are insufficient. He emphasizes the importance of individualized considerations, even if they seem to contradict statistical trends.
The speaker questions whether or not machines can truly possess consciousness, and suggests that their capabilities for intelligence, learning, and predicting do not necessarily equate to consciousness. He also ponders if machines have the ability to teach us.
As AI advances in learning physical tasks, the opportunities for businesses to take advantage of these techniques will grow. However, effective communication and building shared understanding between humans and AI systems can become more challenging as interpretation layers increase.
The most important public dialogue we should have is about the nature of intelligence, inference, logic, reason, rationality, and understanding our cognitive biases and how they work. The physical experience of the world around us plays a significant role in how we process information.
The speaker discusses the potential uses of AGI systems for objective dialogue and problem-solving, such as medical diagnoses, legal cases, and social issues, and the value of having a thought partner for these discussions.