Sam Altman: OpenAI, GPT-5, and AGI
A 2-hour long interview with Lex Fridman
Morning y’all!
Today I’ve got the much talked about interview between Sam Altman and Lex Fridman. I’ve got a summary of it as well for those who are more inclined to read and also a few links to check out as well.
Have a good Tuesday friends!
※\(^o^)/※
— Summer
Here’s the summary of the first hour:
In the Lex Fridman Podcast #419, Sam Altman, CEO of OpenAI, discusses his experiences with the tumultuous events surrounding OpenAI's board saga in November 2022. He reflects on the painful nature of the experience, filled with negative emotions but also moments of support and love.
Altman expects the road to Artificial General Intelligence (AGI) to be a power struggle and believes that the experience helped build resilience for future challenges. He also emphasizes the importance of learning from the experience to understand power dynamics, company structures, and the tension between research, product development, and money.
During the podcast, Altman also discusses the importance of having a well-rounded board with various expertise, including technical savvy, nonprofit experience, and expertise in running companies. He emphasizes the need for a group of individuals rather than hiring board members one at a time. Altman also shares his experience during a challenging weekend for OpenAI and expresses his respect for the team and partners.
The conversation then shifts to the economic implications of AI and the potential impact on creators and artists. Altman believes that people deserve compensation for their valuable data and that a new economic model will need to be established as we transition from traditional methods to AI-generated content.
He also touches on the potential of OpenAI's language model, GPT-4, and its ability to understand and represent the world, acknowledging its limitations and potential dangers. Overall, the podcast covers a range of topics related to OpenAI, AGI, and the role of a board in a high-stakes organization. Altman shares his experiences and insights, providing valuable perspectives on the future of AI and its implications for society.
More details on each major talking point in the first hour:
00:00:00 In this section of the podcast, Sam Altman, CEO of OpenAI, discusses his experience with the tumultuous events surrounding the company's board saga that took place in November 2022. He describes it as the most painful professional experience of his life, filled with negative emotions, but also with moments of great support and love from colleagues. Altman expresses that he expects the road to Artificial General Intelligence (AGI) to be a power struggle, and the experience helped him and OpenAI build resilience for future challenges. Despite the difficult past, they are now back to focusing on their mission, but Altman acknowledges the importance of reflecting on the experience to learn about power dynamics, company structures, and the tension between research, product development, and money.
00:05:00 In this section of the Lex Fridman Podcast #419, Sam Altman discusses the importance of learning from the high-stress experience of nearly losing OpenAI and the role of a board in such a situation. He reflects on the value of structuring and incentives for a board, especially as they relate to building a resilient organization capable of handling the pressure of developing Artificial General Intelligence (AGI). Altman shares that the board members were well-meaning but that the nonprofit's board, which answers to no one but itself, may need to answer to the world as a whole. He also discusses the process of selecting new board members, which included adding Larry Summers and Brett Weinstein, during a tense weekend. The selection was made in the heat of the moment to help move forward with a board of three members.
00:10:00 In this section of the podcast, Sam Altman of OpenAI discusses the importance of having a well-rounded board with various expertise, including technical savvy, nonprofit experience, and expertise in running companies. He emphasizes the need for a group of individuals rather than hiring board members one at a time. Altman also highlights the importance of having technical experts on the board to understand the details of running the business and the impact of their products on society. He mentions that while technical understanding is crucial, it's not the only factor, and the board needs to consider the societal impact of their products as well. Altman shares that during a challenging weekend for OpenAI, he considered stepping down and starting anew but eventually found excitement in the possibility of focusing on AI research.
00:15:00 In this section of the podcast, Sam Altman shares his experience during a tumultuous weekend when OpenAI, the company he co-founded, faced instability and potential change in leadership. Despite the lack of sleep and food, Altman describes the experience as surreal and even painful, but predominantly filled with love and support from the team and partners. He admires Mera Meradi's leadership during the chaos and values her quiet moments of decision-making. Altman also expresses his respect for Ilia, another co-founder, who shares concerns about the safety and societal impact of AGI. Despite the dramatic weekend, Altman emphasizes that OpenAI's progress and the future of AGI are the real focus.
00:20:00 In this section of the podcast, Sam Altman discusses his conversations with Ilya Sutskever, a co-founder of OpenAI, regarding the long-term implications of artificial general intelligence (AGI). Ilya, known for his thoughtful and long-term perspective on technology, has been quiet lately, and Altman speculates that he might be in a reflective mood. Altman shares an anecdote about Ilya's unexpected playful side during a dinner party. The conversation then shifts to the board structure at OpenAI and the lessons learned from a past experience. Altman expresses his trust and gratitude for the team at OpenAI and acknowledges the importance of surrounding oneself with wise and trustworthy individuals as the stakes get higher. The conversation also touches upon Elon Musk's criticism of OpenAI, but no definitive conclusion is reached.
00:25:00 In this section of the podcast, Sam Altman discusses the early days of OpenAI and the challenges they faced in developing language models before they became a big deal. He recalls how they started with research without a clear product in mind and had to make assumptions that often turned out to be wrong. Elon Musk's involvement and eventual departure from OpenAI is also touched upon, with Altman acknowledging some mischaracterization in Musk's blog post about the history of their partnership. Altman emphasizes that OpenAI's mission is to put powerful technology in the hands of people for free as a public good, and that this openness is a crucial aspect of their mission.
00:30:00 In this section of the podcast, Lex Fridman and his guest discuss Elon Musk's lawsuit against OpenAI and the implications of open-source models in the field of artificial general intelligence (AGI). Musk's lawsuit is seen as a way to make a point about the future of AGI and the company leading the way, rather than a legally serious matter. The conversation then shifts to the debate between open-source and closed-source models, with OpenAI being the first to open-source a model, albeit not the state-of-the-art one. The pros and cons of open-source models are discussed, with the potential for tax incentives being mentioned but deemed unlikely for most startups. The speakers express hope for a friendly relationship and competition between Musk and OpenAI in the future, as they both excel at building cool technologies. The conversation then moves on to discuss Sora, a product developed by Musk's Neuralink, and its philosophical and technical differences from GPT-4.
00:35:00 In this section of the podcast, Sam Altman discusses the advancements of AI models, specifically OpenAI's Sora, and their ability to understand and represent the world. He acknowledges the limitations of these models, such as their inability to fully grasp occlusions and the occasional generation of inaccurate results, like cats sprouting extra limbs. Altman believes that these weaknesses are not fundamental flaws but rather areas for improvement through larger models, better technical details, or more data. He also touches upon the involvement of humans in the data labeling process and the potential for self-supervised learning using internet-scale data. The conversation then shifts to the possibility of LLMs (Language Learning Models) moving towards visual data and the dangers of releasing such a powerful system, with efficiency and deep fakes being among the concerns.
00:40:00 In this section of the podcast, Sam Altman discusses the economic implications of AI and the potential impact on creators and artists. He believes that people deserve compensation for their valuable data and that a new economic model will need to be established as we transition from traditional methods to AI-generated content. Altman uses the example of the evolution from CDs to Napster to Spotify as an analogy. He also emphasizes that humans have an inherent desire to create and be rewarded, and while the nature of that reward may change, it is not going away. Altman suggests that the focus should be on what percentage of tasks AI can do over time, rather than what percentage of jobs it will replace. He believes that humans will continue to be the driving force behind content creation, even if AI tools are used in production. Altman expresses his belief that humans have a deep-rooted desire to connect with other humans and that this will continue to be a significant factor in the entertainment industry. He also touches on the impressive capabilities of GPT-4 and looks forward to its potential impact.
00:45:00 In this section of the podcast, Sam Altman discusses his perspective on the advancements of OpenAI's language model, GPT, specifically GPT-4. He acknowledges the impressive achievements of GPT-4 but also emphasizes the need to improve upon it. Altman believes that the gap between GPT-4 and future models, such as GPT-5, will be similar to that between GPT-3 and GPT-4. He encourages the audience to imagine the potential of these future models and how they can be used as creative brainstorming partners or assistants in tackling complex problems. Altman also highlights the importance of the post-training steps and the product development around these models, which are crucial in making them accessible and effective for a large user base.
00:50:00 In this section of the podcast, Sam Altman discusses the expansion of context length in language models, comparing the use of 8K to 128K tokens in GPT 4 and GPT 4 Turbo. He envisions a future where models will have the ability to access vast amounts of information, leading to a better understanding of individuals. Altman draws a parallel to the early days of computing and the exponential growth of technology. He also shares his experiences with using GPT 4 for knowledge work and finds it particularly useful as a starting point for tasks. However, he acknowledges the need for fact-checking to ensure accuracy and expresses concerns about the potential for users to rely too heavily on the model without verification. Altman believes that people are becoming more sophisticated in their use of technology but expresses concern about journalists who may not fully understand the limitations of language models.
00:55:00 In this section of the podcast, Sam Altman discusses the potential of AI models, specifically GPT, to remember conversations and learn from experiences. He expresses his desire for a model that gets to know him and becomes more useful over time. Altman also touches on the issue of privacy, stating that user choice and transparency from companies are essential. He shares a personal experience of a traumatic event and the importance of moving forward despite it. Altman acknowledges the complexity of trust and the need to find a balance between paranoia and trust in people and technology.
Now, you should probably take a break before hitting the second hour! Here’s the high-level of the 2nd hour:
Sam now discusses various topics related to artificial intelligence (AI), including the capabilities and potential developments of OpenAI's GPT technology, the importance of compute as the future's most valuable commodity, the competition in the field of AGI, OpenAI's business model, and the potential capabilities and implications of AGI.
Altman shares his concerns about the limitations of current AI systems and the need for AI to be able to think harder about complex problems and more quickly about simpler ones. He also discusses the importance of safety in advanced AI models, the future of programming, and the role of humans in it. Altman expresses his excitement and fear about the potential of AGI and its impact on society, while acknowledging the challenges and risks associated with its development.
He also touches upon the importance of governance and balance of power in the development of AGI and the potential for AGI to significantly increase scientific discovery rates and have novel intuitions.
Greater detail for reading:
01:00:00 In this section of the podcast, Lex Fridman engages Sam Altman in a discussion about the capabilities and potential developments of OpenAI's GPT (Generative Pre-trained Transformer) technology. Altman shares concerns about the limitations of current AI systems, particularly in high-stress and distrustful environments, and the need for AI to be able to think harder about complex problems and more quickly about simpler ones. Fridman asks about the possibility of slower, sequential thinking in AI and whether it will be similar to current language models or a new paradigm. Altman suggests that more compute may be required for harder problems and that there could be ways to implement slower thinking, but the specifics are yet to be determined. The conversation also touches on the mysterious QAR project at OpenAI and the importance of better reasoning in AI systems. OpenAI's iterative deployment strategy is also discussed, with Altman explaining that the world needs time to adapt to the progress of AI and that OpenAI's approach has been effective in getting the world to take AGI seriously.
01:05:00 In this section of the podcast, Sam Altman of OpenAI discusses the release strategy for their upcoming language model, GPT-5. He acknowledges the importance of milestones and celebrating achievements but expresses concerns about potential shock updates to the world. Altman suggests a more iterative approach, possibly releasing GPT-5 in a different way. He also mentions the challenges in bringing together various components to create a giant transformer model, emphasizing the importance of having a broad understanding of the entire project. Altman also touches upon the need for significant resources, mentioning a figure of $7 trillion, although it was not his original statement.
01:10:00 In this section of the podcast, Sam Altman discusses the importance of compute as the future's most valuable commodity, comparing it to energy. He believes that the world will require a massive amount of compute, and the challenges include energy, data centers, supply chain, and chip fabrication. Altman expresses optimism about the potential of nuclear fusion and new nuclear reactors to address the energy puzzle. He also acknowledges the human fear of nuclear technology and the need to win people over. Regarding AI, Altman expresses concerns about potential theatrical risks and the possibility of it being politicized, which could lead to left versus right wars. He emphasizes the importance of understanding the actual risks and benefits of AI and the need for truth and balance in the conversation.
01:15:00 In this section of the podcast, Sam Altman discusses the potential benefits and drawbacks of the ongoing competition in the field of artificial general intelligence (AGI). He acknowledges the advantages of competition, such as faster, cheaper innovation, but expresses concerns about the possibility of an arms race that could lead to safety issues. Altman emphasizes the importance of collaboration on safety and prioritizing a slow takeoff for AGI. He also touches upon Elon Musk's involvement in the space and the potential benefits of collaboration between different entities. Altman also shares his thoughts on Google's dominance in the information access space and OpenAI's potential to offer new ways of helping people find and act upon information, rather than just building a better search engine.
01:20:00 In this section of the podcast, Sam Altman discusses OpenAI's business model and the potential use of advertisements. He expresses his dislike for ads as an aesthetic choice but acknowledges their historical importance in the industry's growth. Altman believes that OpenAI's current business model, which relies on users paying for access to its AI models, is sustainable. However, he also acknowledges the potential for AI to improve the ad experience by showing relevant and necessary ads without manipulating the truth. Altman also touches upon the issue of safety and bias in AI models, mentioning recent controversies and the importance of having clear guidelines for desired model behavior.
01:25:00 In this section of the podcast, Sam Altman of OpenAI discusses the challenges and complexities of ensuring safety in advanced AI models, such as GPT-5. He emphasizes the importance of defining clear principles and expectations for AI behavior, using Trump vs. Biden as an example. Altman also touches upon the ideological pressures in the tech industry, particularly in San Francisco, and how OpenAI has managed to avoid the culture wars that he's heard about in other companies. The conversation shifts to the topic of safety, with Altman expressing that as AI becomes more powerful, safety will become the primary focus for the company. He also mentions the various aspects of safety, including technical alignment and societal impacts. Altman is excited about the progress of GPT, noting that it's not just getting better in one area but across the board, creating a deeper intellectual connection between the user and the AI.
01:30:00 In this section of the Lex Fridman Podcast #419, Sam Altman discusses the future of programming and the role of humans in it. He believes that programming will continue to exist but in a different form, with some people programming entirely in natural language. The skill set required for programming may change, moving away from traditional programming languages like Fortran. Altman also expresses his hope for the development of humanoid robots to handle physical tasks, as he finds it depressing to think that AGI may be limited to only mental tasks. Regarding AGI, Altman expects significant progress by the end of this decade, with systems becoming quite capable, but not yet reaching a Singularity-level transition. He sees GPT-3.5 as an important step towards taking AGI seriously and changing the world's expectations.
01:35:00 In this section of the podcast, Sam Altman discusses the potential capabilities and implications of Artificial General Intelligence (AGI). He expresses excitement about the possibility of AGI significantly increasing scientific discovery rates and having novel intuitions. If given the opportunity to interact with the first AGI, Altman would ask it questions about fundamental scientific concepts, such as the existence of alien civilizations and the Theory of Everything in physics. However, he acknowledges that the first AGI may not be able to answer these questions and might require more data or assistance to do so. Altman also emphasizes the importance of a robust governance system for AGI, sharing his experiences with OpenAI's board drama and his opposition to having total control over the organization.
01:40:00 In this section of the podcast, Sam Altman, the co-founder of OpenAI, discusses the importance of governance and balance of power in the development of Artificial General Intelligence (AGI). He expresses his belief that no single person should have control over AGI and emphasizes the need for a balance of power. Altman acknowledges that he has made mistakes in the past but does not think any one person is in control of the AGI movement. He also shares his perspective on existential risks, stating that while he is not currently worried about the AI itself, there have been times when he has been. The conversation then shifts to Altman's use of lowercase letters in his tweets, which he explains as a personal preference and a rejection of following rules for the sake of it.
01:45:00 In this section of the podcast, Sam Altman discusses his writing habits, specifically the use of capitalization, and wonders if people still capitalize their searches or messages to themselves. He mentions that he rarely thinks about it and that it's a matter of convention. Altman then shifts the conversation to Sora, OpenAI's ability to generate simulated worlds, and the simulation hypothesis. He shares that the capability to generate novel, photo-realistic worlds increases the probability that we live in a simulated reality, but it's not the strongest evidence. Altman also reflects on the idea that simple concepts can lead to profound insights, using the square root function as an example, and how AI can serve as a gateway to new realities.
01:50:00 In this section of the podcast, Sam Altman expresses his excitement and fear about an upcoming trip to the Amazon jungle. He reflects on the natural world and the possibility of intelligent alien civilizations, expressing a deep appreciation for the advancements of human civilization and the collective progress we have made. Altman also discusses the potential nature of AGI (Artificial General Intelligence) and the role it may play in our future. He shares his thoughts on death and the importance of being grateful for life, and expresses his gratitude for the opportunity to discuss these topics with the podcast host. Altman's reflections touch on the themes of nature, intelligence, and human progress.
01:55:00 In this section of the podcast, Sam Altman discusses his work with OpenAI, specifically mentioning GPT-5 and Sora. He shares that GPT-5 is a large language model that has been trained on a vast amount of data and can generate human-like text. Sora, on the other hand, is an open-source research platform for AGI (Artificial General Intelligence). Altman also talks about his collaboration with Elon Musk and Ilya Sutskever on OpenAI, and how their shared vision for AGI led to the creation of the organization. He emphasizes the importance of AGI and its potential impact on society, but also acknowledges the challenges and risks associated with its development.
Hope you found that interesting!
A few tools and links to click:
Entelligence — Engineering mentor and codebase visualization tool.
Flowsage — Flowcharts and AI.
AI Blueprint — Open source for scaling AI models.
AI-powered decks — Precisely what you think it is.
Browser automation via AI with an SDK too.
\( ﾟヮﾟ)/