Wojciech Zaremba is a co-founder of OpenAI. Please support this podcast by checking out our sponsors: - Paperspace: https://gradient.run/lex to get $15 credit - Indeed: https://indeed.com/lex to get $75 credit - Blinkist: https://blinkist.com/lex and use code LEX to get 25% off premium - Grammarly: https://grammarly.com/lex to get 20% off premium - Eight Sleep: https://www.eightsleep.com/lex and use code LEX to get special savings EPISODE LINKS: Wojciech's Twitter: https://twitter.com/woj_zaremba Wojciech's Website: https://wojzaremba.com/ OpenAI's Website: https://openai.com/ PODCAST INFO: Podcast website: https://lexfridman.com/podcast Apple Podcasts: https://apple.co/2lwqZIr Spotify: https://spoti.fi/2nEwCF8 RSS: https://lexfridman.com/feed/podcast/ YouTube Full Episodes: https://youtube.com/lexfridman YouTube Clips: https://youtube.com/lexclips SUPPORT & CONNECT: - Check out the sponsors above, it's the best way to support this podcast - Support on Patreon: https://www.patreon.com/lexfridman - Twitter: https://twitter.com/lexfridman - Instagram: https://www.instagram.com/lexfridman - LinkedIn: https://www.linkedin.com/in/lexfridman - Facebook: https://www.facebook.com/lexfridman - Medium: https://medium.com/@lexfridman OUTLINE: Here's the timestamps for the episode. On some podcast players you should be able to click the timestamp to jump to that time. (00:00) - Introduction (07:57) - The Fermi paradox (14:59) - Systems of government (17:36) - Life, intelligence, and consciousness (24:49) - GPT language model (27:02) - Engineering consciousness (31:19) - Is there an algorithm for intelligence? (38:41) - Neural networks and deep learning (51:32) - Human reward functions (56:26) - Love is part of the human condition (58:53) - Expanding our circle of empathy (1:02:58) - Psychedelics and meditation (1:14:25) - Ilya Sutskever (1:21:42) - How does GPT work? (1:31:35) - AI safety (1:38:21) - OpenAI Codex (1:51:54) - Robotics (2:01:11) - Developing self driving cars and robots (2:12:02) - What is the benchmark for intelligence? (2:15:04) - Will we spend more time in virtual reality? (2:17:18) - AI Friendships (2:26:48) - Sleep (2:29:22) - Generating good ideas (2:35:47) - Advice for young people (2:40:31) - Getting started with machine learning (2:43:44) - What is beauty? (2:47:35) - Death (2:54:23) - Meaning of life
The episode features ads for Indeed's job platform and Paper Space Gradient's machine learning model building, training, and deployment platform, as well as recommendations for FastAI courses and Gradient for future machine learning projects.
The speaker ponders the idea that humanity may be alone in the universe, making consciousness life more valuable, and discusses the possibility that aliens may seek to optimize their use of compute rather than wasting the current entropy of the universe.
The process of creating a model of reality involves simplifying reality to predict what will happen in the world. However, intelligence is inherently messy and imperfect, making it difficult to apply perfect compression to the task of intelligence such as predicting the next word in AI models like GPT.
The process of predicting the future involves program synthesis, where you start with errors and enumerate all the programs that can produce the expected output given the input.
Christopher Oleg's work on interpretability of neural networks allows for the identification and extraction of specific functions, such as detecting a wheel or mask for a car, within a larger network.
The podcast explores whether AI can develop empathetic abilities beyond language and whether it can enable access to therapy to more people. It discusses the challenges involved in developing such an AI and the requirements for the same.
The boundary between two people dissolves when in love and they end up optimizing each other's reward functions. Love means helping others to optimize their reward functions, not your own, by caring and understanding what the other person wants.
The speaker expresses the desire for humanity to develop a scientific understanding of consciousness, enabling a tangible way to place value on it. They mention their experience with meditation and make a connection to why people don't like to torture animals.
Focusing attention can bring back forgotten memories and simulate the reward function of others, leading to a sense of oneness with everything.
The host asks Ilya Sutskever about the physical discomforts of sitting for long hours while also discussing the power of human intuition and deep thinking in machine learning.
The conversation between two AI models can become inaccurate as there is no human feedback loop during the generation process. In contrast, humans learn from their experiences and correct their misconceptions quickly.
The idea of possessing an AI system with exceptional power has raised concerns about the possibility of it corrupting the character and integrity of individuals, including those in the US government, and the strategy behind deploying such systems.
The AI model, pre-trained on language and then fine-tuned on code, can suggest code for specific requirements, allowing developers to avoid searching for appropriate code by themselves.
The speaker stresses the importance of reliability in programming, particularly when it comes to controlling high-risk machinery. They also advocate for democratizing knowledge in programming so that people from various fields can easily program and implement models.
In order to solve a Rubik's Cube, an AI has to manipulate it with five baby arms without dropping it, and utilize simulated interactions to develop a model with the necessary data to solve it in real life.
The diversity in environments and objects that robots have to interact with pose difficult challenges in developing a cost-effective product. Specifically, creating a home robot that can touch an extensive variety of objects with different shapes is a significant challenge.
The process of deploying self-driving cars involves showing people that the technology can be trusted, given that society is harsher on them than on human-driven cars. Establishing a benchmark for achieving intelligence can play a role in building this trust.
The podcast explores the possibility of achieving compelling natural language conversations between people and AI systems on the internet. The ultimate goal is to create an AI tutor to help children learn and grow in a more empathetic way.
The speaker experimented with gradually decreasing their number of hours of sleep but found that after a certain point they felt tired all the time. The speaker also mentions their tendency to do research at night and go to bed in the early morning.
The speaker finds reinforcement learning to be an enjoyable way to create something special, like fun games. They also note that criticizing ideas as "bad" can interrupt the process of generating more ideas.
The speaker muses on the properties of a cup and how its expansion due to temperature changes allows for life in extreme conditions, while also referencing Jeff Bezos' philosophy of regret minimization and the healing power of short psychedelic experiences.
The speaker discusses how our brain's reward functions, such as comprehension and curiosity, lead us to optimize our actions and ultimately contribute to our life satisfaction.
Lex Fridman interviews Wojciech Saremba about his career in robotics and machine learning.