Introduction of Peter Norvig:
Peter Norvig, Director of Research at Google, delivers a lecture at Microsoft Research AI Distinguished Lecture Series. His background includes roles at NASA and Google, and he is known for co-teaching a massive open online class on artificial intelligence. Norvig highlights Vannevar Bush’s forward-thinking ideas in 1945, envisioning new forms of encyclopedias through the Memex device. Despite the initial limitations of technology, Bush’s vision laid the groundwork for modern information access and organization tools.

Evolution of Research and Expression Tools:
Norvig discusses the significance of research tools in exploring, expressing, and following implications of thoughts. The Memex is an example of a research tool, but programming and computational tools offer additional capabilities. Jupyter Notebooks are useful for exploration, while GitHub enables sharing and publication of research findings. Programs excel at following implications and expressing thoughts, but they may lack accessibility for everyone.

Historical Milestones in Computing:
Norvig showcases the remarkable progress in computing technology over time. He compares a room full of computers with 1 bit each to a terabyte-sized device, highlighting the miniaturization and increased storage capacity. The moth incident in Grace Mary Hopper’s computer is contrasted with modern debugging tools, demonstrating advancements in error prevention.

Collaborative Efforts in Programming:
Norvig emphasizes the shift from individual programming to large-scale team efforts in a factory-like setting. This engineering solution facilitated the development of complex projects, beyond the capabilities of individual programmers. The goal of “CS for all” or “coding for all” initiatives is discussed, aiming to make programming accessible to everyone.

Challenges in Programming:
Norvig points out the limitations of current programming languages and environments in terms of accessibility. He raises questions about whether improvements in programming languages alone are sufficient or if a new approach is needed. The focus on logical and mathematical foundations in programming languages is highlighted. The limitations of Boolean logic in expressing complex concepts and properties in programming are discussed.

Margaret Hamilton and the Apollo Lander:
Norvig shares his experience working under Margaret Hamilton, who led the development of the Apollo lander software. He describes the role of a programmer as a micromanager, dictating every step to the computer, similar to a strict supervisor. The drawbacks of this approach are highlighted, including its rigidity and lack of autonomy for the computer.

Alternative Approaches to Programming:
Norvig introduces the concept of machine learning models as an alternative way of dealing with the world. He outlines the standard machine learning model, involving labeled data, deep learning, and training. This approach allows the computer to learn from data and make predictions or decisions without explicit step-by-step instructions.

Machine Learning as Empirical Science:
Machine learning has shifted software creation from a mathematical science to an empirical science, resembling biology where theories are formed and tested through observations. Programming involves providing examples and instructing the system to generalize from those examples rather than explicitly instructing each step.

Complexity of Machine Learning:
Machine learning involves numerous steps beyond model creation, including data management, deployment, monitoring, updates, and integration. The misconception among startups is that machine learning training is the most challenging aspect, while in reality, data management and deployment are often more time-consuming.

Differentiable Networks and Workflow:
Machine learning networks are differentiable, allowing for adjustments and improvements based on performance feedback. However, the entire machine learning workflow is not yet fully differentiable, making it difficult to detect and correct errors in data acquisition and feature selection.

Enhancing End-User Usability:
To make machine learning tools more accessible to end-users, the entire process should be differentiable. This would allow for automatic correction or detection of errors, ensuring optimal performance and ease of use.

Alternative Approaches to Memex Goals:
Memex aims to explore, organize, express, and follow implications of thoughts. Traditional programming and machine learning are two approaches to achieving these goals. Exploring alternative perspectives and methodologies could lead to new ways of achieving these objectives.

Skill Intersection vs. Union for Data Science:
Traditional data science view requires intersection of programming, statistics, and subject matter expertise, creating a danger zone for those lacking statistical knowledge. An alternative view proposes a union of these skills, allowing success with partial expertise.

Joel Grews’ Updated Venn Diagram:
Adds a fourth circle representing thesis advisors who used to program but no longer do. Emphasizes that statistical expertise is not essential for data science success.

Makoto Koukei’s Cucumber Sorting Automation:
Makoto Koukei, a cucumber farmer’s son, wanted to automate cucumber sorting to avoid taking over the family business. He lacked subject matter expertise but possessed mechanical engineering skills. He built a conveyor belt, downloaded TensorFlow, and used a camera to sort cucumbers into different categories.

Skill Level Required for Automation:
Koukei’s example raises the question of how much skill is needed for such automation tasks. Is it limited to exceptional engineers, or can anyone achieve it?

Conversation with the System:
To build a functional system, a user needs to communicate with it effectively. Ideal conversations would involve providing class examples, hints, and feedback.

Current Conversation Limitations:
Instead of natural conversations, data scientists often engage in technical discussions about computational layers, learning rates, and batch sizes.

Pitfalls of Current Programming Languages:
Current programming languages are logical rather than probabilistic. Variables have a single value, not a probability distribution. Conditionals are Boolean, not probabilistic.

Probabilistic Programming Languages:
Provide built-in probabilistic types and variables. Allow for inference in multiple directions. Examples include Bayes DB and infer.net.

Extension of Database Programming:
Possible approach for making probabilistic programming more accessible. Similar to making queries to a SQL database with uncertain fields. Suitable for tasks where the cost of errors is low.

Extension of Spreadsheet Programming:
Another possible approach for making probabilistic programming more accessible. Familiar and easy to understand for many users. Could be suitable for tasks involving data analysis and modeling.

The Power of Natural Language:
Natural language is the most intuitive way for humans to communicate. Future programming languages may incorporate natural language processing. This would allow users to express their intentions and goals in a more natural way.

Conversational Computing:
A potential phase change in computing, moving from mainframes to personal computers to mobile and now possibly to conversational computing.

The Allure of Conversational Computing:
Eliminating the need for keyboards and screens, with voice commands and conversational interactions. Increased popularity despite initial skepticism.

Weaknesses of the Current Model:
Lack of user control over apps. Privacy, security, and fairness concerns due to app developers’ control.

Potential Solutions:
Improving app security and privacy. Shifting control from app writers to conversational agents, ensuring user privacy and protection.

Challenges in Conversational Computing:
Current systems resemble phone menus rather than natural conversations. Limited application developer kits, requiring manual responses to triggers.

Intermediaries in Conversational Computing:
Using intermediaries for specific services, such as ride-hailing or restaurants, to specialize sub-conversations.

Responsibilities and Control:
Unclear division of responsibilities between the user and conversational agents.

Difficulty in Implementing User Requests:
Complex processes, such as customizing map features, often require extensive technical knowledge.

Conclusion:
Conversational computing has the potential to revolutionize the way we interact with technology, but challenges remain in creating user-friendly and efficient systems.

User Interface Design:
Design documents often present a single “best” interface choice without considering the preferences of all users. Discarded interface options are not readily accessible and cannot be easily revived.

Benefits of Programmatic Control:
A development environment that stores and allows querying of experimental artifacts would enable the retrieval and reuse of discarded interface options. Users could request specific interface designs based on their preferences.

Example of Interface Failure:
A mobile alert provided accurate information about a rental car return but incorrectly suggested a bicycle as the mode of transportation.

Analysis of the Failure:
Parsing the email, planning the trip, and creating the alert were all executed correctly. The failure occurred due to the lack of semantic understanding of the event title, which resulted in an inappropriate transportation suggestion.

Conclusion:
As interfaces become more complex due to uncertain and real-world factors, the importance of comprehensive documentation and the ability to retrieve and reuse discarded design options becomes crucial.

Designing Interfaces for Natural Language Processing and Speech Recognition:
Interfaces for natural language processing and speech recognition face unique challenges due to the vast and complex nature of human language and the need to share extensive information between dialogue participants.

The Need for Simple Interfaces:
Early web interfaces aimed for simplicity, focusing on passing basic information like names and zip codes between users and web pages. Complex interfaces become difficult to manage and can hinder communication.

Challenges in Interfacing with Devices:
When interacting with devices using natural language, the interface requirements become much more complex. The system needs to understand the entire grammar of a language, speech patterns, background information, and world knowledge to effectively communicate.

Information Sharing Dilemmas:
Determining how to share the vast amount of information required for natural language processing and speech recognition presents challenges. Questions arise regarding what information should be shared, who should initiate queries, and how to protect user privacy.

Design Thinking in Architecture:
Peter Norvig explores design thinking in architecture as a potential model for interface design. Architects use specific design patterns and iterative processes to generate possibilities and explore design options.

Using Lego Blocks for Knowledge Expression:
Norvig suggests providing designers with Lego-like tools to express their knowledge and explore design possibilities. Automating the process of generating diagrams based on constraints and feedback could facilitate the design process.

The Importance of Expressing Preferences and Desires:
Machine learning systems require clear specifications of user preferences, utilities, and desires to optimize outcomes. Cautionary tales like King Midas and the genie’s lamp highlight the consequences of imprecise specifications.

Unexpected Consequences of Language Use:
Norvig emphasizes the importance of considering the potential unexpected consequences of using natural language in interface design. The Forbidden Planet example illustrates how language can be manipulated to achieve unintended results.

Bug Fixes for Unforeseen Consequences:
The Krell civilization’s software product allowed users to conjure anything they desired, resulting in nightmares and monsters that killed them all. Bug fixes for such scenarios include adding an undo function, conducting simulations and small-scale tests, performing adversarial tests, and implementing shutdown and rollback capabilities. Prompt specifications should include options for unsafe actions and continued evolution and improvement.

Reliance on Language and the Need for Clarity:
Language alone cannot guarantee accurate interpretation and execution of desires. Laws and regulations, along with case law examples, are necessary for better understanding and implementation. Over-reliance on language can lead to misleading outcomes.

The Role of Leaders in Machine Learning:
Leaders should provide objectives and let their teams determine the best approach to achieve them, allowing for surprises and innovative solutions. This approach aligns with General Patton’s philosophy of letting people know what to do rather than dictating how to do it.

The Paradox of Needs and Wants:
Mick Jagger’s statement, “Can’t always get what you want, but you get what you need,” is inaccurate. Human behavior is often driven by short-term wants, while true needs are often overlooked. The marketplace rewards clicks and wants, leading to a vicious cycle of short-term gratification that does not necessarily align with long-term needs.

Aligning the Marketplace with Human Needs:
The tragedy of the commons can be avoided by learning from successful historical examples of shared resource management. Design patterns exist to help create reliable systems and share resources equitably within the marketplace.

Combining Different Programming Forms:
Traditional programming, machine learning model building, conversation, and design thinking can complement each other. Machine learning can optimize traditional programs and help determine optimal parameter values for operating systems.

Composability of Widgets:
Infinite possibilities in user interfaces are limited by the need for composability. A constraint language is necessary to predict how adding new elements will affect the overall UI. Providing users with recovery options in case of errors is important.

Under-Invested Areas of AI:
Fairness, understanding data sources and usage, and robotics are under-invested areas in AI. Robotics is similar to personal computers in the 1970s, with potential for significant advancements. Conversational systems and their programming are also intriguing areas to watch.

Computing’s Role in Human Society:
Computing should enhance human lives by simplifying tasks, providing information, and enabling creativity. The extent of computing’s role should be determined by its ability to improve human lives and society as a whole.

The Role of Programming in Empowering Individuals:
Peter Norvig highlights two aspects of programming: how one programmer can help others and how much everyone should be a programmer. He emphasizes the potential of programming in enhancing people’s understanding of the world and dispelling misconceptions, such as the Flat Earth Theory. Norvig stresses the importance of programming for personal use, allowing individuals to create customized solutions and develop problem-solving skills. He believes that programming empowers individuals to explore their interests, seek answers, and model the world, leading to a deeper understanding of it.

Machine Learning and AI as a Unique Design Material:
In response to a question about the uniqueness of machine learning and AI as a design material, Norvig provides his perspective. He suggests that ML and AI differ from other design materials due to their ability to learn and adapt based on data and patterns. This dynamic nature of ML and AI allows designers to create systems that can continuously improve and evolve over time. Norvig emphasizes the need for designers to understand the limitations and biases of ML and AI systems to ensure responsible and ethical design practices.

Machine Learning:
Peter Norvig emphasizes that the life-changing moments in machine learning and AI occur when a created program surpasses human capabilities, leading to deep understanding and curiosity about the process.

Quantum AI:
Quantum AI’s potential impact is significant, but its practical applications are still uncertain. Norvig acknowledges the progress made by Google’s quantum team and sees potential benefits for optimizing neural net parameters with fewer parameters and improved algorithms.

Debuggability:
Debuggability is crucial, especially in critical software components and operating systems. Norvig acknowledges the challenges of debugging machine learning systems and suggests the need for better tools and techniques.

Debuggability and Proof in Machine Learning:
Machine learning poses challenges in terms of debuggability and proving correctness compared to traditional programs due to the complexity of the algorithms and the non-amenable nature of some domains to proof.

Limitations of Proof in Machine Learning:
Proving the correctness of machine learning algorithms is often difficult or impossible due to the inherent uncertainty and complexity of the models and the domains they are applied to.

Envelopes and Control Theory:
In control theory, the concept of envelopes is used to define boundaries within which a system is expected to stay. This approach can be applied to machine learning to establish boundaries on the expected behavior of the system.

Tiered Testing for Continual Updating:
With online learning and continual updating, testing becomes more challenging as new data points are continuously incorporated. A tiered testing approach is necessary, involving different levels of testing at different frequencies to ensure system reliability.

Conversational Systems and the Spectrum of Natural Language Interaction:
The benefits of conversational systems can be realized across a spectrum of natural language interaction, not just limited to colloquial, natural conversations. Specialized languages and programming languages also play important roles in human-computer interaction.

Specialized Solutions for Particular Tasks:
Peter Norvig believes that the future of interaction with AI will involve specialized solutions for particular tasks. Instead of using natural language, users may use specialized notations or learn programming concepts to interact with AI for specific tasks.

Generalized Training:
The generalized part of AI interaction will be the training process. Users will learn how to create models, debug them, and see the results and implications of their actions.

Model Building and Problem-Solving for Everybody:
Norvig prefers the phrase “model building or problem-solving for everybody” over “programming or coding for everybody.” He believes that programming is just one way to express models and solve problems, and there are other ways as well.

Emphasis on Model Creation, Debugging, and Results:
The important aspect of AI interaction is the ability to create models, debug them, and see the results and implications. The specific programming language used is less important.