Summary
In episode 387 of the Lex Fridman Podcast, the host had an insightful conversation with tech entrepreneur and computer hacker, George Hotz. Commencing with an exploration of consciousness and artificial intelligence, Hotz unveils his theory of AI's potential to be used destructively by humans, rather than posing a direct threat itself. Contextualizing his argument, Hotz further outlines the stark difference between the swift progression of computing technology, driven by the Moore's Law, and the considerably slower development of biological life forms.
In a deeper dive into forecasting an AI-driven apocalypse, Hotz introduces the idea of 'wireheading', highlighting the crippling human dependence on technology that could lead to societal self-destruction. However, in a potentially optimistic view of an 'end of the world', he predicts the rise of a reconstituted society formed by survivors with a strong taboo against technological interference. The discussion further meanders into the realms of time, human subjectivity, and the possible AI dominance over human intelligence. Hotz critiques the human tendency to attribute consciousness to AI models, even suggesting that our perception of large language models like GPT-3 is grounded in such attribution.
Turning to the topic of virtual reality, Hotz shares his aspiration to live within it but concedes that the technology is currently underdeveloped, lacking the ability to provide a truly immersive experience. He remains hopeful though, asserting that the human craving for escape from reality will propel the advancement of virtual reality.
The discourse continues with their perspectives on the revival of the human experience through technology, ranging from the simple pleasures of savoring a pizza to the complex engagement in human-like AI relationships. The podcast also explores the difficulties in writing software for AI and new ML accelerators, touching on the challenges of creating a PyTorch stack on Nvidia GPUs. The conversation also emphasizes the criticality of decentralization of computational power to curb monopolistic tendencies, highlighting their preferable approach in the form of TinyGrad and their custom PC, TinyBox.
Hotz also sheds light on his aspirations for Comma AI, advocating for a movement towards learning a human driving policy from data while also acknowledging that Tesla is considerably ahead in this race. The dialogue veers into an analysis of AI developments, particularly around looping retrieval systems with the potential to redefine search engines. The discussion raises concerns about the misuse of AI and the false information or "hallucinations" that such misuse can incite.
The conversation eventually navigates through the redesigning of Twitter's existing sponsorship model, advocating for improved user experience and the simplification of Twitter's complex codebase. Hotz, while expressing a predilection for Scala as a programming language, divulges his vision of AI-powered coding aids and 'prompt engineering' as the highest level of abstraction achievable in coding.
Lastly, the discussion dives into the realm of gaming, exploring the possibilities of AI in gaming and virtual reality. Exploring his relationship with computation, Hotz emphasizes his passion for his 'first love', his computer, and wraps up by underlining his belief that God would want His creations to achieve their potential, just as he wishes for his own creations in AI and robotics.
