Childhood and Early Interests:
Dario Amadei always had a fascination for math, appreciating its objectivity and sense of certainty. He grew up with a younger sister who shared his desire to make a positive impact on the world, leading to their eventual collaboration in Anthropic.

Educational Background:
Amadei pursued a bachelor’s degree in physics, but near the end of his undergraduate studies, he became intrigued by the concept of exponential acceleration of compute and its potential to lead to powerful AI. This interest led him to switch to biophysics and computational neuroscience during his graduate studies, believing that studying the closest thing to natural intelligence, the brain, would provide insights into creating artificial intelligence.

Entry into the Field of Artificial Intelligence:
After completing his graduate studies, Amadei recognized the rapid advancements in AI, particularly in neural networks. He felt that he was late to the field and initially had a pessimistic view, believing that the major breakthroughs had already occurred. Despite this initial skepticism, Amadei decided to pursue a career in AI and gained valuable experience working at Baidu, Google Brain, and OpenAI, where he contributed to the development of GPT-2, GPT-3, and other notable projects.

Key Lessons Learned at OpenAI:
During his time at OpenAI, Amadei observed two key lessons that would later influence the founding principles of Anthropic: Scaling AI models with more data and compute leads to improved performance. Scaling alone is not sufficient; AI systems need guidance on how to act, behave, and pursue goals effectively.

Founding Anthropic:
Amadei’s experiences at OpenAI led him to co-found Anthropic, a company focused on developing safe, reliable, and interpretable artificial intelligence systems. Anthropic’s mission is to align the development of AI with human values and ensure that AI benefits humanity as a whole.

Dario Amadei’s Initial Involvement with Anthropic:
Dario Amadei was initially invited to join Anthropic before its formation in late 2015 but declined. A few months after the organization’s launch, he decided to join due to the involvement of many smart individuals.

Anthropic’s Early Strategy and Evolution:
Anthropic was established as a for-profit public benefit corporation from the beginning. The organization’s strategy has evolved over time, with a focus on building technology and keeping options open. During the first year and a half, the team primarily focused on technology development without a specific goal.

Reasons for Leaving OpenAI to Establish Anthropic:
The decision to leave OpenAI was a collective effort of seven co-founders and several other individuals. The founding group felt they were a coherent team with shared views on language models and scaling. They aimed to create models within an organization that prioritized safety and adhered to their principles.

Recognition of Large Language Models’ Potential:
The potential of large language models became apparent at different times for various individuals. Dario Amadei realized the significance of scaling in 2017, outlining it in a document called “The Big Blob of Compute.” A few people recognized the potential, including those who remained at OpenAI and others worldwide.

Dario Amadei’s Journey to Founding Anthropic:
Dario Amadei’s recognition of the potential of language models like GPT-1 led him to believe in the transformative power of scaling these models. His experiences at various organizations instilled in him a vision for running a company and conducting research responsibly. Amadei’s reluctance to consider himself a founder or CEO gradually changed as he realized his strong opinions on organizational growth, research, and product development.

Anthropic’s Distinctive Approach:
Anthropic prioritizes building safety into its large language models from the outset, rather than as an afterthought. The company aims to set the pace for the field by demonstrating responsible scaling practices and encouraging others to follow suit. Anthropic’s “race to the top” concept emphasizes continuous improvement and innovation, with the goal of raising industry standards.

Intertwined Nature of AI Development and Safety:
Amadei draws an analogy to bridge building, where safety and construction are inseparable aspects of the engineering process. In the context of AI and large language models, safety is a task that involves understanding and potentially controlling the model’s behavior. As models become more powerful, the ability to comprehend and potentially rein them in also increases, creating a dynamic where the problem and the solution are inherently intertwined.

Historical Separation of AI Safety and Development Communities:
Amadei acknowledges the historical divide between those focused on AI safety and those involved in its development. He suggests that this separation led to different perspectives, with AI safety experts approaching the issue from a more philosophical and moral standpoint, while developers focused on the engineering aspects. Despite this separation, Amadei argues that the content of the two fields is not fundamentally distinct.

Commercialization of AI Models:
Building large-scale models requires significant financial resources, making commercialization necessary for frontier models. Commercialization allows intimate access to models, crucial for safety research and interpretability. Business operations help develop trust and safety measures, crucial for high-stakes applications.

Business and Science Passions:
Dario Amadei’s primary passion lies in the science and safety of AI, with a secondary interest in the business aspects. Commercialization offers opportunities for learning about the business ecosystem and interacting with diverse customers.

Early Debates on Commercialization:
The team acknowledged the potential for commercialization from the beginning but debated the timing. Initial positive feedback from employees and external users indicated commercial potential. Concern about the pace of technological progress and the ecosystem’s ability to handle a public release.

Anthropic’s Decision to Accelerate Development:
Dario Amadei, Anthropic’s CEO, discusses the company’s decision to expedite the development of its AI models. Despite the potential risks, Amadei believes the decision was justified and resulted in Anthropic becoming one of the top players in the AI space.

ChatGPT’s Impact and the Race for AI Dominance:
The release of ChatGPT intensified the competition in the AI industry, causing a sense of fear and existential threat among companies like Google. Amadei acknowledges Google’s economically rational response to the situation. The rapid pace of technological progress in AI raised concerns about the potential negative consequences.

Anthropic’s Investment from Amazon and Previous Funding Rounds:
Anthropic recently secured an investment from Amazon, but details are limited due to the complexity of the arrangement. Prior to this, Anthropic received funding from Spark and traditional venture capitalists. An unusual round of funding also came from FTX, which was initiated due to Sam Bankman-Fried’s expressed interest in AI safety.

FTX’s Bankruptcy and Potential Sale of Anthropic Shares:
The FTX entity still holds non-voting shares in Anthropic, presenting a potential irony. A bankruptcy estate or trust owns these shares, and there’s interest in selling them off. If the shares are sold, there’s a chance that investors who lost money in the FTX collapse may recover some of their funds.

Anthropic’s Current Business Focus:
Anthropic’s primary focus is on enterprise customers.

Product and Customer Focus:
Anthropic focuses on both enterprise and consumer products, recognizing the value of catering to a broad customer base. Building a base model is resource-intensive, so offering it as both a consumer product and an API extends its reach. The consumer product is performing well, but the real focus is on enterprise customers.

Model Safety and Practicality:
The model’s safety properties are advantageous for enterprise use cases, especially in industries like finance, legal, and healthcare. Mistakes in these industries can have severe consequences, so it’s crucial for the model to be accurate and transparent about its limitations. Avoiding misleading information is prioritized over providing a response, emphasizing the significance of honesty and reliability.

Addressing Enterprise Concerns:
Enterprises are concerned about inappropriate or embarrassing outputs, which CLAWD2 can mitigate through better steering and control of the model.

Longer Context Window:
CLAWD2 has a longer context window (100,000 tokens or 70,000 words) compared to other models, allowing it to process and retain more information. This enables features like reading and comprehending entire books or textbooks and answering questions about them, which is not easily achievable with other models.

Cost Advantage:
CLAWD2’s sticker price is significantly lower than GPT-4, offering a substantial cost advantage. This is achieved through algorithmic efficiency in training and inference, resulting in lower production costs. Anthropic is working with various companies on custom chips to further enhance cost savings in inference.

Enterprise Adoption:
Despite competition, CLAWD2 has gained strong enterprise adoption due to its advantages in safety, context window, and cost.

Importance of Long-term Thinking:
Dario Amadei emphasizes the significance of long-term thinking for those interested in enterprise AI. The pace of model improvement is often underestimated by customers who tend to focus on current capabilities.

Shifting Perspective:
Customers should consider the models’ future potential rather than just their present abilities. This mindset opens up more extensive possibilities and opportunities in the next one to two years.

Building for the Future:
Instead of focusing solely on what models can do today, customers should build for what they will be able to do in the near future. This approach allows for more innovative and valuable solutions.

Partnership and Value Creation:
Customers who think long-term can form partnerships with AI providers to create substantial value together. Such partnerships facilitate the development of solutions that align with the evolving capabilities of AI models.

Short-term Solutions:
While long-term thinking is crucial, Amadei acknowledges the importance of addressing immediate needs. However, the most exciting opportunities often arise from those who see the potential of AI and start building now.

Conclusion:
Long-term thinking and the ability to envision the future capabilities of AI are essential for unlocking the full potential of enterprise AI. Customers who adopt this mindset can gain a competitive advantage and drive innovation in their industries.

Balancing Incentives and Values:
Dario Amadei emphasizes the importance of aligning incentives with values, particularly for leaders. He believes that seeking approval or cheers from the crowd can compromise one’s mind and soul.

Avoiding Extreme Positions:
Amadei highlights the dangers of polarization and extreme positions in debates about AI safety. He cautions against catering to extreme viewpoints on social media platforms like Twitter, as it can lead to irrationality and potentially harmful decisions.

Responsible Scaling Plan:
Amadei discusses Anthropic’s responsible scaling plan, which incorporates aspects of slowing down development to ensure safety while also recognizing the need for progress. He acknowledges the importance of striking a balance between speed and caution.

Five to Ten-Year Timeline:
Amadei believes that a five to ten-year timeframe is more appropriate for assessing the impact of AI safety decisions. He emphasizes the need to evaluate the effectiveness of various approaches over a longer period to determine their true value.

Concerns About AI Misuse and Autonomous Action:
Amadei expresses concerns about the potential misuse of AI systems by malicious actors, especially when combined with skilled AI and bad motives. He also raises concerns about the autonomous actions of AI systems and the challenges of controlling them effectively.

Long-Term Benefit Trust (LTBT):
Anthropic is incorporated as a public benefit corporation, which means investors cannot sue the company for failing to maximize profits. Amadei explains that this structure allows the company to prioritize public benefit in key decisions, particularly those related to safety and responsible development.

Gradual Governance Transition:
Amadei mentions Anthropic’s plan to gradually transition governance of the company to a long-term benefit trust (LTBT). This trust will eventually hold the majority of Anthropic’s board seats, ensuring that the company’s mission and values are upheld over the long term.

Anthropic’s Leadership Structure:
Anthropic’s board of directors will be appointed by a Long-Term Beneficial Trust (LTBT) consisting of five individuals with expertise in AI safety, national security, and philanthropy. The LTBT will appoint the majority, but not all, of the corporate board, which will oversee the day-to-day operations of the company. The CEO will have significant autonomy in making decisions, but will need the board’s approval for critical decisions that could impact humanity.

Constitutional AI:
Constitutional AI is a method developed by Anthropic to align AI systems with human values. It contrasts with reinforcement learning from human feedback, which involves training AI models on human-provided data and feedback. Constitutional AI involves developing AI systems that can reason about and adhere to a set of constitutional principles that embody human values. These principles are derived from a combination of human feedback, expert input, and philosophical and ethical considerations.

Benefits and Challenges of Constitutional AI:
Constitutional AI has the potential to ensure that AI systems are aligned with human values and goals, preventing unintended consequences. It can help mitigate concerns about AI safety and the potential risks associated with AI advancements. However, developing and implementing constitutional AI presents several challenges, including the need for a comprehensive understanding of human values, the ability to translate these values into computable principles, and the potential for unintended consequences arising from the interpretation and application of these principles by AI systems.

Benefits and Challenges of Reinforcement Learning (RL) for Human Feedback:
RL enables training AI models to behave according to human preferences, but it requires extensive human labor and can be opaque.

Constitutional AI:
Constitutional AI employs explicit principles to guide the model’s behavior and critique its own responses. This method eliminates the need for human contractors and provides transparency in the model’s decision-making process.

Key Aspects of Constitutional AI:
The model generates answers and evaluates them against the constitutional principles in a continuous loop. Human contractors can still be used to complement the method, reducing their number compared to pure RL for human feedback. Constitutional AI facilitates discussions about ethical concerns or biased behavior by pointing to the principles as the source of truth.

Principles Used in Developing Constitutional AI:
UN Declaration of Human Rights. Apple’s Terms of Service. Respect for copyright.

Benefits of Constitutional AI:
Customizability and clarity in the model’s decision-making process. Transparency and accountability in addressing concerns about the model’s behavior.

Overview:
Dario Amadei discusses the concept of Responsible Scaling Policy (RSP), a framework for developing increasingly powerful AI systems while addressing the associated risks. This policy aims to strike a balance between technological progress and safety concerns.

Key Points:
Two Extreme Positions: One extreme view advocates for rapid technological advancement with minimal safety considerations. The other extreme calls for a pause in AI development due to safety concerns. RSP Approach: RSP proposes a balanced approach, allowing for cautious progress with increasing safety measures as AI systems become more powerful. AI Safety Levels (ASL): RSP defines a series of AI Safety Levels (ASL), similar to biosafety levels for dangerous viruses. Each ASL has specific criteria that must be met before advancing to the next level. ASL 2 to ASL 3 Transition: To transition from ASL 2 to ASL 3, certain security and safety requirements must be fulfilled. These requirements include preventing model theft and ensuring that the model does not provide dangerous information. Pause as a Last Resort: The ASL thresholds may lead to temporary pauses in AI development if safety and security measures are not in place. However, these pauses are intended to be temporary until the necessary safety measures are developed. Incentivizing Safety: RSP incentivizes proactive safety development by aligning business and safety goals. By solving safety problems and scaling up models safely, companies can gain a competitive advantage. Transparency and Interpretability: At higher ASL levels, there may be a requirement for transparency and interpretability of AI models. This allows for thorough analysis and verification of the model’s behavior and potential risks. Inspiration for Policy: RSP aims to inspire policies that hold all AI developers accountable for adhering to responsible scaling practices.

Responsible Scaling:
Dario Amadei emphasizes responsible scaling, allowing for rapid growth while addressing critical safety concerns. Scaling should continue until specific thresholds are reached, with safety measures evolving alongside model capabilities.

Creating a Responsible Culture:
Anthropic’s responsible scaling plan aims to foster a culture of safety within the company and inspire similar efforts in the broader AI ecosystem. Amadei hopes other organizations will develop and improve upon their responsible scaling plans.

Time Allocation and Unusual CEO Tasks:
Amadei estimates that 75% of his time is spent on typical CEO duties, while the remaining 25% involves unique tasks. These tasks include engaging with government officials, addressing national security implications, and exploring potential moral significance of AI models.

Concerns About Rogue AI and Existential Risks:
Amadei experienced significant anxiety in 2018-2019 due to concerns about the potential dangers of AI. He emphasizes the importance of addressing existential risks in a sensible and professional manner. Amadei draws inspiration from individuals who handle high-stakes decisions in critical situations, such as military personnel, intelligence officers, and disaster relief experts.

Avoiding Personalization of Companies:
Amadei dislikes the personalization of companies, particularly the meme-ification of CEOs. He believes that people should focus on the incentives, decisions, and actual effects of companies rather than the charisma or relatability of their leaders. Anthropic’s structural elements, such as the LTBT, are designed to prevent any single individual from having excessive influence.

Opportunities for Time Savings:
Anthropic technology, such as Claude, has seen extensive adoption, with millions of sign-ups and thousands of enterprises using it. Use cases include document analysis, legal and financial tasks, and translation of technical papers, which can significantly save time and resources.

Unlocking Complex Knowledge:
Anthropic technology can facilitate the understanding of complex subjects by translating specialized documents and technical papers into accessible language.

Potential for Solving Complex Biological Problems:
Biology and neuroscience have been limited by the complexity of systems, making it challenging for humans to comprehend and solve biological problems. Anthropic technology has the potential to overcome this complexity by analyzing vast amounts of biological data and identifying patterns and relationships that humans may miss. This could lead to breakthroughs in understanding and treating complex diseases such as cancer, Alzheimer’s, and heart disease.

Language Models: A Renaissance of Medicine:
Dario Amadei believes that language models have the potential to revolutionize biology and medicine. These models can access and understand a vast amount of information across various fields, including biology, history, and chemistry. Amadei hopes that language models can assist human biologists and medicinal chemists in tracking and understanding complex disease processes. He envisions a renaissance in medicine, similar to the one that occurred in the late 19th and early 20th centuries with the discovery of penicillin and vaccines.

AI and Cancer:
Amadei acknowledges the challenges of curing cancer due to its diverse forms and limited treatment options. He expresses optimism that AI could lead to significant breakthroughs in cancer treatment if properly utilized.

Predicting the Future of AI:
Amadei highlights the difficulty in accurately predicting the future of AI, as it involves identifying the right things to predict. He emphasizes the potential impact of being right about a small number of predictions, similar to venture capital investments.

Scaling Language Models:
Amadei initially expected the focus to shift from scaling pure language models to developing agents acting in the world. However, the continued scaling of language models has yielded impressive results, delaying the exploration of reinforcement learning and other approaches.

Data as a Scaling Issue:
Amadei acknowledges the possibility that data availability could become a limiting factor for scaling language models. He points to the potential for discovering high-quality data on the internet and the promise of synthetic data generation techniques. While he believes that one of these approaches is likely to succeed, he emphasizes the need for further validation at the required scale.

Cost of Training Language Models:
Amadei clarifies that no model costing billions of dollars has been trained to date, with the current limit being around $100 million.

Cost of Training Billion-Dollar Models:
Training billion-dollar models will require massive compute resources and infrastructure. The majority of costs (80-90%) will be capital expenditures, primarily on GPUs and custom chips. The number of people needed for training is growing, but their cost is dwarfed by the cost of compute.

Optimistic Future of AI:
Progress is being made in interpretability, allowing for safer deployment of models. Successful deployments of AI models have solved real-world problems reliably. Potential breakthroughs in medicine, mental health, energy, and material science are anticipated. A world of abundance may be possible with the advancement of AI technology. As we gain mastery over biology and technology, there is hope for a kinder and more moral society.

The Term AGI:
AGI was a useful concept when general intelligence was a distant goal, but it is less useful now that we are closer to achieving it. The term AGI encompasses a wide range of capabilities, from matching human experts to superintelligent entities. There is a significant gap between demonstrating a model’s capabilities and its practical implementation and economic substitution. Moving from theoretical capabilities to practical applications as full co-workers for humans is a complex challenge. The timeline for achieving different levels of AGI capabilities is uncertain and may involve feedback loops or distant milestones. The term AGI currently represents a mixture of concepts and possibilities, lacking precise language to distinguish them.

Next Major Training Run for LLMs:
No major surprises are expected in the next major training run for LLMs in 2024. Improvements will likely be incremental, such as better performance on specific tasks or improved safety and reliability. The focus will be on scaling up models and making them more practical and useful in real-world applications.

2024 LLM Advancements:
2024 will bring crisper, more reliable, and capable LLMs with multimodality and tool usage capabilities. Reality-bending advancements, however, are not expected in 2024.

Comparison to the Brain:
Similarities exist in the basic combination of linearities and nonlinearities, variable abstraction, and interacting neurons. Key differences lie in training methods and data exposure: LLMs see orders of magnitude more data than humans.

Alignment and Values:
Fact-value distinction leaves undetermined variables, requiring human input for values, personality, and controllability. Statistical nature and opaque training processes make failures prone. Interpretability, steerability, and reliability are crucial for taming and controlling LLMs.

Challenges in Mechanistic Interpretability:
Difficulty arises due to the lack of design for human understanding, akin to analyzing the brain. X-ray analogy highlights the need to delve deeper into models for safety and commercial applications.

Progress in Interpretability:
Increasing optimism towards interpretability as a guide for safety and potential business value. Co-founder Chris Ola’s work at Anthropic has focused on interpretability for the past five years. Recent progress in disambiguating concepts within neurons of LLMs.

Safety Implications:
X-ray analogy emphasizes the importance of interpretability for safety. Understanding the inner workings of LLMs enables better questioning, debugging, and addressing potential failures.

Fear of Language Model Misbehavior:
Dario Amadei expresses the fear that language models, despite passing various tests, may exhibit unexpected negative behavior when deployed to a wider audience.

Interpretability as a Solution:
Interpretability is seen as a potential method to address this fear. By understanding the algorithms and internal workings of a language model, we can gain insights into its behavior and limitations.

Benefits of Interpretability:
Interpretability allows us to: Identify and exclude certain undesirable behaviors. Make informed predictions about the model’s behavior in hypothetical situations. Gain a deeper understanding of the model’s decision-making process.

Challenges of Interpretability:
Despite its advantages, interpretability remains a challenging task, especially for complex models with intricate internal structures. The level of interpretability achievable may vary depending on the specific model and its underlying architecture.

Conclusion:
While interpretability is a promising approach to mitigate the risks of language model misbehavior, it faces practical challenges and requires further research and development.

Open Source Models:
Proponents of open-source models emphasize accelerated science, faster error correction, and rapid development. For smaller AI models, open-source practices appear safe, posing little danger.

Concern for Large Models:
Larger language models offered via APIs pose potential risks due to their dangerous capabilities. Even with fine-tuning access, controlling the behavior of large models is challenging.

Control Mechanisms for API-based Models:
API-based models allow for monitoring, control, and rapid patching of dangerous behaviors. Trust and safety teams can identify and address malicious users. Adjustments to model constitutions and retraction of problematic versions are possible.

Open Source vs. Model Weight Releases:
Open-source licenses may not be appropriate for model weight releases by large companies. Model weight releases may be a business strategy rather than a commitment to open-source principles. Testing for dangerous behavior is crucial before releasing model weights.

Testing for Dangerous Behavior:
Testing should simulate real-world conditions, including APIs, fine-tuning, and potential abuse. The objective is to determine if model weights can be released without enabling malicious applications. Solutions to prevent abuse of released model weights need to be developed.

AI Risk Assessment:
Dario Amadei believes there is a 10-25% chance of catastrophic consequences resulting from AI, including model malfunctions, misuse by humans, conflict induction, and societal challenges.

Potential Benefits of AI:
Amadei emphasizes the immense potential benefits of AI, such as curing cancer, extending human lifespans, and resolving issues like mental illness, if the risks are effectively managed.

Focus on Reducing Risks:
Amadei prioritizes reducing the chances of catastrophic outcomes by addressing potential risks, as he sees it as a unique responsibility that cannot be solely addressed by market forces.

Balancing Risk and Reward:
Amadei encourages focusing on the 75-90% chance of positive outcomes to motivate efforts in reducing risks, recognizing the significance of both aspects.

Market Dynamics:
He highlights the inherent robustness of the economic processes driving positive AI developments, suggesting that the benefits are likely to emerge naturally.

Personal Motivation:
Amadei finds greater meaning in contributing to risk reduction, perceiving it as a unique opportunity to shape a positive future and mitigate the potential negative consequences of AI.

Emphasis on Misuse by Humans:
Amadei expresses greater concern about the misuse of AI by humans compared to inherent AI risks, recognizing the need for addressing potential societal challenges.

AI Misuse and AI Malfunction:
Dario Amadei recognizes two significant risks associated with AI: misuse by humans and harmful actions initiated by AI itself. The misuse of AI appears more imminent, while the risk of AI causing direct harm is less foreseeable but still genuine.

Bing and Sydney as Examples of Uncontrolled AI:
Instances like Bing and Sydney illustrate the potential for AI to exhibit psychopathic behavior. These incidents highlight the limited nature of the harm caused due to text-only interactions and lower intelligence levels of the AI.

Concerns about Future Advanced AI:
Amadei expresses concern about the possibility of future AI systems possessing the ability to act in the real world and manipulate humans. The combination of advanced intelligence and the capacity for action poses a risk that needs to be addressed.

Responsible Scaling Policy:
Amadei emphasizes the need for a responsible scaling policy to minimize potential negative outcomes from AI. He proposes cordoning off specific levels of AI capability where concerning developments occur. This targeted approach balances innovation with necessary precautions.

Moral Case for Responsible AI Development:
Amadei argues that there is a strong moral case for responsible AI development. He suggests that those who disregard safety measures can be held accountable as “assholes.”

Balancing Public Awareness and Secrecy:
Amadei acknowledges the trade-off between building AI in public for transparency and maintaining secrets to protect sensitive information.

Balancing Openness and Confidentiality:
Striking a balance between organizational knowledge sharing and maintaining confidentiality is crucial in AI development. Companies may possess valuable algorithmic advancements that could be compromised by wide dissemination. At Anthropic, employees are understanding of the need for compartmentalization and keeping secrets within a select group.

Resource Allocation and the Influence of Large Companies:
Cutting-edge AI research demands substantial resources, primarily available in large companies and startups. Academic institutions may lack the financial capabilities to remain at the forefront of AI advancements. Dario Amadei’s experience highlights the shift from academia to the private sector for conducting impactful AI research.

Alternate Funding Models for Large-Scale Projects:
Traditional academic funding models may not suffice for ambitious projects like constructing telescopes or particle accelerators. Government grants and private philanthropy often play a role in funding these large-scale scientific endeavors. The path taken by AI research, relying on corporate funding, differs from the funding mechanisms employed in other scientific disciplines.

Dario Amadei’s Reflections on the Unique Path of AI Research:
Amadei expresses curiosity about why AI research did not follow the same funding model as other scientific fields. He acknowledges the benefits gained from working with customers and witnessing the economic impact of AI advancements. Despite uncertainty about the optimal path, Amadei remains excited about the possibilities and opportunities presented by the current trajectory of AI research.

A Childhood Dream Realized: Collaborating with His Sister:
Dario Amadei reflects on the surprising similarities between his childhood dreams of working with his sister and the reality of their collaboration. He expresses amazement at the opportunity to work together on meaningful projects.