On January 12th, Microsoft co-founder and former CEO Bill Gates released a podcast conversation with OpenAI CEO Sam Altman on his self-media platform.
They discussed various topics, including artificial intelligence, ChatGPT, and the entrepreneurial journey of OpenAI. The conversation between these two tech giants provided insights into their perspectives on artificial intelligence and revealed some details about OpenAI.
It’s worth noting that Bill Gates mentioned early on that shortly after recording the podcast, Sam Altman was removed from the position of OpenAI CEO, went through a series of events, and was eventually rehired. Therefore, the podcast content does not address these recent developments, and Altman’s perspective might be different if he were to reflect on the conversation now.
At the beginning of the discussion, Gates expressed his surprise at ChatGPT’s performance and curiosity about how AI could “encode” Shakespeare’s works for data processing. Altman explained that, similar to the human brain, OpenAI doesn’t fully understand which “neurons” are at play, but he believes that over time, there will be more progress in the interpretability of AI.
“You will see this in many empirical discoveries of technological development history. Although they don’t know what happened, it clearly works. Then, as scientific understanding deepens, they can make it better,” Altman said.
Looking ahead to the next two years, Altman predicted that multimodal AI, capable of understanding text, image data, and more, will be a significant trend. He also emphasized the need for improvement in AI reliability, citing an example where GPT-4 might provide a good answer to a question among 10,000, but it might not know which one. Customizability and personalization of AI will also be crucial due to diverse user needs for GPT-4.
Regulation in the AI field was a major topic of their conversation.
Gates mentioned government failures in regulating social networks, and Altman agreed, stating that AI-related challenges faced by governments are even more complex.
Regarding the regulation of artificial intelligence, Altman stated, “We are beginning to figure it out.” He highlighted the risk of excessive regulation in technology, as it can easily occur, and assessing the impact of technology on society and geopolitical power balance is challenging for technology inventors. Altman then provided an example of a regulatory model under discussion, similar to the International Atomic Energy Agency’s approach, involving a global organization overseeing large-scale AI systems. However, he anticipated encountering short-term challenges and being influenced by factors such as copyright and geopolitical situations.
Altman further expressed that demanding AI systems to slow down is exceptionally difficult. Nonetheless, he mentioned potential solutions, such as regulating a few of the world’s most powerful systems, setting specific and high-power thresholds for computing clusters, subjecting any cluster to inspections by international weapons inspectors, and ensuring large AI models undergo safety audits, tests during training, and audits and tests before deployment.
Discussing the productivity boost from artificial intelligence, Altman expressed excitement about programming as the most promising field. Currently, AI-assisted programming has been widely deployed, and he also anticipated rapid developments in healthcare and education as two other areas.
Gates, when considering the potential achievement of AGI+ (superintelligent general artificial intelligence), voiced concerns about three things: the control of the system by malicious actors, the possibility of the system controlling everything, and the philosophical question of human value after achieving superintelligent general artificial intelligence. He focused on the last point, questioning how society would find purpose and organize itself.
Altman responded by acknowledging the psychological challenges of working in the technology field and finding value in those challenges. He stated that finding value is the most difficult aspect and, in a sense, might be the last difficult task he undertakes.
Altman addressed the cost issues related to ChatGPT, stating that GPT-3, the model with the longest launch time and most extended optimization from OpenAI, had its costs reduced by 40 times over the three-plus years since its introduction. As for GPT-3.5, they have cut its costs by nearly 10 times. Altman explained that the GPT-4, being a new product, hasn’t had as much time for cost reduction, but this process will continue. He expressed confidence that OpenAI’s cost reduction curve for AI systems is even better than Moore’s Law. This improvement is attributed to finding more efficient ways to model and gaining a better understanding through research.
Altman explained that the fundamental models in the world currently consist of intelligence costs and energy costs, the two most significant factors influencing the quality of life. By simultaneously reducing costs in these two aspects, humans can have access to more. Currently, OpenAI has managed to sufficiently lower costs, evident from the $20 monthly subscription fee for ChatGPT, surpassing its value.
Regarding the average age of OpenAI’s employees, Altman noted that it is not a company where all employees are very young. The 500 employees are mostly in their thirties, forties, and fifties. Gates commented that it is different from the early days of Apple and Microsoft. Altman mentioned that the company has generally aged, which, to some extent, is not a positive sign for society. He observed that over time, the age of the most outstanding founders is increasing.
In terms of entrepreneurship, Altman stated that OpenAI did many things contrary to the standard Y Combinator (YC) recommendations. It took them four and a half years to launch their first product. At the company’s inception, they had no concept of the product and did not engage with users. However, Altman mentioned that his learning and experiences at YC taught him how to break the rules.
Altman expressed that OpenAI is not a typical startup. Due to high requirements for funds and computing power, Silicon Valley investors might not provide the necessary support, possibly exceeding the limits of what venture capital can bear. This is because OpenAI cannot provide investors with predictable productization timelines and values. The investment from Microsoft ultimately played a crucial role.
Discussing the team building and advice behind OpenAI’s achievements, Gates emphasized the importance of establishing a team with the right skill set, where mathematics and science are cool, but composite skills should not be overlooked. Altman encouraged people to step out of their comfort zones and do what they genuinely want to do. “Clearly state what you want to do and make requests to others; there will be unexpected gains. Many people are trapped in spending time on things they don’t want to do, and my most common advice is probably to find a way to solve this problem.”
Finally, Gates and Altman exchanged thoughts on commonly used software. Altman mentioned that he uses Slack the most when it comes to office collaboration software, although he “wishes he could say that ChatGPT is his most frequently used software.” Gates, on the other hand, stated that his most commonly used software is the Outlook email application, considering himself an old-fashioned email enthusiast.