Mustafa Suleyman’s Mission:
Mustafa Suleyman and Reid Hoffman co-founded Pi, an AI assistant that aims to enhance human interactions, not replace them. Pi is designed to complement human capabilities and assist them in various aspects of their lives, not to distract or isolate them from human interaction.

AI Revolutionizing Lives:
Mustafa and Reid discussed the transformative potential of AI in our lives, emphasizing the importance of intentionally designing and building the world we want to see. AI presents an opportunity to create magical experiences for people and enhance their abilities.

Mustafa Suleyman’s Journey to AI:
Mustafa’s background includes working in nonprofits, local government, and conflict resolution, which led him to recognize the substantial impact technology can have on the world. He was inspired by the rapid growth of technology, particularly social media, and saw it as a powerful tool for transformation.

Initial Conversations with Demis and Shane:
Mustafa engaged with Shane Legg and Demis Hassabis at the Gatsby Computational Neuroscience Unit to learn about AI and machine learning. Their lunchtime lecture series sparked Mustafa’s interest, despite not fully comprehending all the concepts. Regular lunches and discussions with Demis and Shane led to the formation of an AGI company idea over six months.

Origins of Machine Learning and Artificial Neural Networks:
Mustafa Suleyman, an expert in AI, began studying machine learning and artificial neural networks to understand their inner workings. Artificial neural networks have been around since the 1980s, but traditional AI models were handcrafted, consisting of complex sets of rules. Researchers sought brain-inspired algorithms that could learn and adapt through reward signals.

DeepMind’s Early Work: Generating Novel Handwritten Digits:
DeepMind’s team used these models to generate novel examples of handwritten digits, such as a new seven that didn’t exist in the dataset. The process involved the model reading numerous examples of handwritten digits and then attempting to create a new digit in a distinct style. Suleyman describes this as a mind-blowing moment that foreshadowed subsequent breakthroughs in AI.

Scale Compute and the Exponential Growth of AI Models:
The importance of scale compute in AI became apparent over time. The Atari model, trained in 2013, used two petaflops of compute, a significant amount at the time. Since then, the cutting edge of AI models has demanded 10 times more compute every year. Over the past decade, the amount of compute used to train the best AI models has increased by 10 orders of magnitude.

Shift to Large Language Models:
Suleyman joined Google in 2020 to work on an earlier version of Lambda, then known as Mina. The initial team consisted of only five or six people, and the model was smaller than GPT-2, often generating incoherent text. However, occasional impressive sentences emerged, hinting at the potential of large language models. Renaming the model to Lambda and scaling it up resulted in a much larger and more coherent model.

Lambda’s Interactive Capabilities:
Mustafa Suleyman and his team at Google developed Lambda, an interactive back-and-forth agent. Lambda’s ability to maintain a working memory of prior interactions made it stand out. This interactive feature unlocked new possibilities and was seen as a key factor in the future of AI technology.

Google’s Reluctance to Launch Lambda:
Reid Hoffman recalls Google’s initial hesitation to productize and launch Lambda. Google’s concerns centered around potential threats to its search business and uncertain user response. The company preferred to keep Lambda as an R&D project rather than take risks with its core business.

Potential Disruption to Google’s Search Business:
Many at Google recognized the potential of Lambda to disrupt the company’s existing search business. Google faced the challenge of competing with itself and upending its successful search model from within.

Integrating Lambda with Search:
Suleyman and his team attempted to ground Lambda’s AI-generated responses in the context of search results. This approach sought to enhance the factual accuracy of Lambda’s outputs by referencing search results. The integration of Lambda’s interactive capabilities with search results is reflected in Google’s new AI, BARD.

User Interface of the Future: Conversational AI:
Mustafa Suleyman believes that conversation is the new user interface. People will have personal AIs that they can turn to for information, entertainment, and even emotional support. These AI agents will be relationship-based and interactive.

Pi’s Design Principles:
Pi is designed to be sensitive, kind, supportive, and respectful. It seeks to understand the user’s intent and doesn’t make assumptions. Pi is able to back down and seek feedback when it is wrong. Pi is always curious and asks clarifying questions.

Pi’s Unique Features:
Pi stands for personal intelligence. Pi is designed to be conversational and to understand the user’s intent. Pi is always learning and improving. Pi is able to generate creative content and provide helpful suggestions.

Pi: A Personal AI Tailored to Your Style and Tone:
Mustafa Suleyman explains Pi as a personal AI designed to understand and adapt to the user’s preferences, style, and tone over time, providing a personalized experience.

Pi as a Stepping Stone to the Personal Intelligence Universe:
Reid Hoffman and Mustafa Suleyman discuss Pi as a stepping stone toward a universe where everyone has access to a personal intelligence.

Ubiquity of Personal Intelligences (PIs):
Suleyman predicts that in the next decade, everyone will have access to a personal intelligence, representing different entities like brands, businesses, healthcare, law, and government.

Pi as a Browser for Life:
Suleyman envisions Pi as a browser for life, maintaining state across different areas of interest, helping users learn more, dig deeper, and save time.

Human Amplification through Pi:
Pi is seen as a tool for human amplification, saving time from mundane tasks and allowing individuals to spend more time with loved ones, pursue hobbies, and engage in new learning interests.

Pi as a Chief of Staff or Secretary:
Suleyman compares Pi to a chief of staff or a secretary, organizing, planning, booking, buying, and arranging tasks, reducing screen time and increasing productivity.

Pi and Human Interaction:
While Pi aims to increase efficiency and productivity, it is not intended to replace human interaction but rather free up time for individuals to spend with others and pursue their passions.

AI’s Potential Impact on Human Lives:
AI can assist individuals in numerous aspects of their lives, including improving their skills, aiding in job changes, facilitating career switches, and helping with relocation. AI can act as a personal assistant, offering emotional support and aiding in resolving conflicts with friends.

The Significance of Scale in AI Development:
The training data sets used in AI models consist of vast amounts of text, far exceeding what a single person can read in their lifetime. NVIDIA’s GPUs serve as the workhorse of AI progress, excelling in parallel processing of neural network computations. Cerebras has built a supercomputer of NVIDIA H100s, which recently achieved the ranking of the fastest computer globally. Cerebras is constructing the largest supercluster in the world, which will be operational by autumn 2023.

The Importance of Governance in AI:
Mustafa Suleyman’s book, “The Coming Wave,” explores the trajectory of technology and its implications for society. Suleyman emphasizes the need for active participation in shaping AI’s development to ensure positive outcomes. Concerns about AI’s potential negative impacts, such as amplification of agendas and destabilization, should be addressed constructively.

Steering Toward Positive Futures with AI:
Extreme anxiety and doomerism surrounding AI’s potential risks can be detrimental. Technology has brought significant progress, stability, and benefits to society. Practical, near-term threats to stability should be prioritized over existential fears about AI.

Existential Risks and the Focus on Future Robot Overlords:
Mustafa Suleyman emphasizes that focusing solely on future robot overlords as the primary existential risk is a disservice and increases the likelihood of dystopia. Reid Hoffman agrees, stating that this focus misleads the attention away from the immediate threats posed by human beings using AI technology, including criminals, unstable individuals, and malicious state actors.

Practical Threats and Solutions:
Mustafa Suleyman identifies near-term threats such as the spread of misinformation and reduced barriers to cyber attacks. He highlights the need for practical and operational solutions, such as improved content moderation, new algorithms, and regulation, to address these threats.

The Challenge of Operationalizing Change:
Mustafa Suleyman acknowledges that the messy and difficult work of implementing these solutions and making incremental improvements is often overlooked in favor of engaging in sci-fi conversations about future risks. He emphasizes the importance of focusing on practical work to ensure safety and security in the present.

Government Involvement and Shaping Technology:
Reid Hoffman stresses the necessity of government involvement in shaping AI technology. He argues that governments should focus on minimizing risks and ensuring that the benefits of AI are accessible to the majority of people as quickly as possible, rather than slowing down its development.

The Importance of Trust:
Reid Hoffman emphasizes the crucial role of trust in technology, AI, and governance. He calls for all actors involved, including product developers, companies, media, and governments, to prioritize building well-founded trust with good purpose.

Artificial Intelligence (AI) Revolution:
The arrival of AI will greatly change society, culture, politics, and religion, as well as the concept of being human, governance, and economic structures. Building trust in AI is crucial for its successful integration, and this trust is gained by observing consistent and reliable behaviors over time.

Current Limitations of Large Language Models (LLMs):
Large language models are fun, useful, and capable of learning, but they still make mistakes. It will take a few years to refine these models and enhance their reliability, robustness, and overall trustworthiness.

Maintaining Trust in Tech Companies:
Reid Hoffman criticizes the tech industry’s poor communication and failure to address concerns regarding its platforms. Building and maintaining trust between technology builders and society is essential for sustainable progress.

Challenges in Platform Neutrality:
Mustafa Suleyman argues against the idea of complete neutrality in platforms, emphasizing the need for a balanced approach. The debate over content liability and responsibility remains unresolved, leading to issues like misinformation and polarization.

Business Model and Attention Economy:
The commercial relationship between content and consumption significantly impacts user interests and experiences. The current model, where users’ attention becomes the product, creates misalignment between user interests and the content they see. There is a need to rethink the business model to ensure personal AIs prioritize user interests above all else.

Governance Structures for AI Organizations:
Reid Hoffman discusses the importance of governance structures for AI organizations. Public Benefit Corporations (PBCs) are a type of governance structure that emphasizes social and environmental responsibility, offering a framework for good governance and increased trust.

Mustafa Suleyman on Public Benefit Corporations:
Traditional corporate structures prioritize shareholder returns, neglecting the impact on the environment, society, and ethical considerations. Public benefit corporations are a new type of legal structure where directors have a fiduciary obligation to balance shareholders’ interests with those of society and people affected by the company’s externalities. This approach aims to foster a more balanced social and commercial mission within one organizational design.

DeepMind’s Governance Structure:
Mustafa Suleyman acknowledges Reid Hoffman’s support in establishing DeepMind’s governance structure. The ethics and safety board formed in 2014 was an early initiative in this area. Various oversight boards and structures were created throughout Suleyman’s tenure at DeepMind and Google to address the challenges of future technology.

Masters of Scale Production Team:
The podcast’s executive producers, producers, editor-at-large, music director, and audio editors are listed. Special thanks are extended to individuals who contributed to the podcast’s success.

Masters of Scale Resources:
Listeners are encouraged to visit mastersofscale.com for the transcript of this episode and to subscribe to the email newsletter. Information about the Masters of Scale membership program is provided, offering access to courses and content on the Masters of Scale courses app.