The Allure of Mathematics:
Greg Brockman’s journey into AI began before college, driven by a desire to dedicate his life to mathematical pursuits.

Chemistry and Programming:
Adam D’Angelo’s initial passion for chemistry led him to create a chemistry textbook with a mathematical focus. To share his knowledge, he ventured into programming and became captivated by its ability to translate complex ideas into tangible outcomes.

Alan Turing’s Influence:
D’Angelo encountered Alan Turing’s 1950 paper on the Turing test, which highlighted the need for machines capable of learning beyond human comprehension.

The Pursuit of Machine Learning:
Inspired by Turing’s vision, D’Angelo resolved to dedicate his life to developing machines that could learn.

Challenges and Obstacles:
Despite his enthusiasm, D’Angelo faced setbacks in his early attempts, as AI technology was still in its infancy.

Entrepreneurial Endeavors:
D’Angelo founded Stripe in 2010 and grew it from a small team to over 250 employees. The company’s success allowed him to witness the transformative potential of AI firsthand.

Deep Learning Breakthroughs:
D’Angelo recognized the emergence of deep learning as a breakthrough that aligned with his long-held aspirations in AI. He became determined to contribute to the advancement of AI for the benefit of society.

OpenAI’s Progress and Investment: The conversation begins with a recognition of OpenAI’s significant growth, underscored by a substantial investment from Microsoft. The speaker questions the use of this investment, given its magnitude, particularly in a research-oriented organization.

Comparative Analysis with Physics Research: Adam D’Angelo compares the funding in AI with that in physics, highlighting the enormous costs associated with building structures like particle accelerators (e.g., LHC). He emphasizes that unlike certain physics investments, AI technology promises a direct Return on Investment (ROI) by benefiting people more tangibly.

Strategic Use of Investment: The primary use of OpenAI’s investment is revealed to be in building massive supercomputers. This decision stems from a finding that scaling up systems to unprecedented sizes has been the most effective strategy in advancing AI.

Cost Allocation and Research Focus: It’s clarified that most of the funding will go towards computing infrastructure rather than researcher salaries. This is because, in comparison, the computational costs are significantly higher. This approach represents a deviation from traditional AI research, which often emphasizes algorithmic improvements.

Exponential Growth in Compute Usage: D’Angelo references a graph showing an exponential increase in compute usage for landmark AI achievements since 2012. This growth rate far exceeds Moore’s Law, illustrating the rapid advancement in AI capabilities largely driven by increased computational power.

The Bitter Lesson of AI Research: The discussion shifts to a concept called “The Bitter Lesson,” articulated by Rich Sutton, a pioneer in reinforcement learning. This lesson suggests that while new ideas are valued in AI research, the most significant advances have consistently been tied to scaling up computational power. Small tweaks to existing methods, rather than groundbreaking new ideas, have enabled significant progress when applied at much larger scales.

Pragmatism in AI Development: Concluding, D’Angelo characterizes OpenAI’s approach as pragmatic. The organization did not initially intend to focus on building large supercomputers, but this direction emerged as the most effective in advancing the field. This strategy is seen as a culmination of a 60-year progress wave in AI research.

Introduction to GPT-2 and Release Strategy: The discussion begins with an acknowledgment of GPT-2, a highly capable language model developed by OpenAI. The speaker notes that unlike previous AI research, OpenAI chose not to release this model immediately, citing potential dangers and societal impacts.

AI Technologies as More Than Toys: Adam D’Angelo emphasizes that AI technologies have significant societal impacts, necessitating a shift from viewing them merely as technological advancements. He suggests that AI researchers must consider the broader implications of their creations from the outset, rather than leaving these considerations to those commercializing the technology.

GPT-2’s Potential Risks: The decision to withhold GPT-2 from immediate public release stemmed from concerns about its potential misuse, such as generating fake news on a superhuman scale. D’Angelo highlights the internal debates within OpenAI regarding the ethical implications of releasing such a powerful tool, leading to a policy of caution in the face of uncertainty.

Changing Norms in AI Research: The decision not to release GPT-2 immediately was seen as an opportunity to change norms within the AI field. D’Angelo explains that while the aim is not to halt progress, there is a need to balance it with responsibility for future consequences. This approach challenges the traditional norm of open publication in AI research.

Controversy and Miscommunication: D’Angelo acknowledges that the nuanced message behind OpenAI’s decision sparked controversy and was often oversimplified in headlines. He notes the difficulty in conveying complex ideas in a media landscape driven by sensational headlines.

Impact and Acceptance of New Norms: Despite initial controversy, D’Angelo finds encouragement in the fact that OpenAI’s cautious approach worked. Other organizations have replicated OpenAI’s results and adopted similar staged release patterns. This shift indicates a growing acceptance in the scientific community that considering the implications of AI research is a fundamental responsibility.

Introduction to OpenAI 5: The discussion opens with a focus on OpenAI 5, a project where OpenAI developed an AI system to play Dota, a highly popular video game. This game was chosen for its complexity, large player base, and the high level of skill and strategy involved, including elements of teamwork and coordination among its five-player teams.

OpenAI’s Achievement in Dota: Adam D’Angelo highlights OpenAI 5’s significant progress. Initially, the AI lost to professional players at the world championships, but after extended training, it succeeded in defeating the reigning world champions. This achievement is noteworthy not just as a gaming accomplishment but as an indication of the generality and adaptability of AI systems.

Broader Goals Beyond Gaming: The primary excitement around OpenAI 5 isn’t rooted in gaming achievements. Instead, the focus is on developing general-purpose technologies. The training system used for Dota was also successfully applied to robotics, leading to a significant breakthrough in controlling a robotic hand.

Addressing the Critique of AI in Gaming: Responding to the critique that AI advantages in gaming might largely be due to reaction times rather than strategic intelligence, D’Angelo points out that in Dota, unlike games like StarCraft, rapid clicking (actions per minute) is not a crucial skill. This distinction underscores that OpenAI 5’s success is more about strategy than mere reaction speed.

Public Perception of AI’s Gaming Capabilities: D’Angelo notes a shift in public perception. Initially, when AI lost to humans, it was perceived as lacking strategic depth. However, upon winning, critics suggested the victory was due merely to faster reaction times. This change in perception illustrates the evolving understanding and expectations of AI in complex tasks.

Influence of AI on Human Gameplay: Intriguingly, the human team that OpenAI 5 defeated and that won subsequent championships started adopting strategies similar to those used by the AI. This suggests that AI can not only compete with human strategies but also influence and potentially improve them.

Hype and the Potential for an AI Winter:
Hype surrounding AI’s potential and the promise of general AI could lead to unrealistic expectations. Failure to meet these expectations could result in an AI winter, similar to what occurred in the past.

Commercial Value Driving AI Research:
Funding for AI research has shifted from academic institutions to commercial entities like Google, Facebook, and Microsoft. Commercial value and potential profits drive research efforts and funding decisions.

Importance of Results:
Investors recognize the potential value of successful AI applications, leading to increased funding. However, funding should be based on results and value creation, not solely on hype.

Calibration and Honesty:
Researchers should be honest about AI’s capabilities and limitations. Avoiding fear of excitement over results is essential to communicate the benefits of AI.

Conclusion:
Balancing hype with realistic expectations is crucial to avoid an AI winter. Commercial value drives research, but funding should be tied to results. Honest communication about AI’s capabilities is vital for progress and public support.

Question on Openness and Release of Work: The discussion starts with an audience question regarding OpenAI’s commitment to openness and the apparent contradiction in not releasing certain AI technologies. This question addresses the tension between OpenAI’s foundational mission and its current practices.

OpenAI’s Founding Goal: Adam D’Angelo explains that OpenAI was founded with the aim of ensuring that advancements in general intelligence and AI would benefit everyone. The initial strategy involved inclusive development, inspired by the successes of the internet and open-source software.

Realization of Dual-Use Implications: Over time, OpenAI recognized that AI technologies could have dual-use implications. While AI can be used for beneficial purposes, it also holds the potential for negative applications by malicious actors. This understanding led to a reevaluation of their strategy.

Responsibility in AI Development: Addressing concerns about AI misuse, D’Angelo asserts that developers of powerful AI technologies have a responsibility to keep these technologies safe and secure. This responsibility has shifted OpenAI’s focus from inclusive development to ensuring that everyone benefits from AI advancements.

Structural Changes in OpenAI: In alignment with this new strategy, OpenAI transitioned from a nonprofit to a capped profit structure. This change was designed to ensure that the benefits of their AI advancements are not confined to a single institution but are distributed more broadly.

Conclusion: The conversation concludes with a reaffirmation of OpenAI’s commitment to its mission, balancing the need for openness with the responsibility of ensuring safe and beneficial use of AI technologies.