George Hotz is the founder of Comma.ai, a machine learning based vehicle automation company. He is an outspoken personality in the field of AI and technology in general. He first gained recognition for being the first person to carrier-unlock an iPhone, and since then has done quite a few interesting things at the intersection of hardware and software. This conversation is part of the Artificial Intelligence podcast. If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter, LinkedIn, Facebook, Medium, or YouTube where you can watch the video versions of these conversations. If you enjoy the podcast, please rate it 5 stars on iTunes or support it on Patreon.
In this episode of the Artificial Intelligence podcast, Lex Friedman interviews George Hotz, the founder of Comm. AI, a machine learning-based vehicle automation company, on his views on AI and technology in general, as well as his thoughts on living in a simulation.
The podcast discusses the possibility of humanity being in a simulation, the idea that the simulation may be designed to keep the hardware in spec, and whether it would be possible to prove that we are in a simulation or not.
A developer talks about receiving feedback from respected individuals regarding code quality, and the difficulty of debugging a processor without a debugger or low-level tools.
The speaker expresses his bias towards the peak years of 2011 to 2015 in the CTF competitive scene, and mentions its similarity to modern competitive programming. He notes the quality of the players and the amount of innovation during those years.
A discussion on the possibility of good hackers being lured into unethical hacking for monetary gain and the distribution of good and bad people in the hacking community.
The guest shares her admiration for the speed and efficiency of the host's programming, likening it to watching a skilled gamer. They discuss Project Zero's approach to API documentation.
In this episode, the host discusses distributed file systems like Ceph and Seaweed FS, and reflects on his recent embrace of programming languages like JavaScript and PHP, which are often criticized as being "bad" languages with poor performance and functionality.
In this podcast, the guest discusses their conversation with Elon Musk regarding autonomous driving, including correcting Musk's misunderstandings and struggles to solve all aspects of autonomous driving.
Lane keeping is a valuable safety feature for autonomous vehicles as it alleviates stress from the driver and improves overall safety by removing the need to constantly stay vigilant while driving. However, the technology is still reliant on sensors that could potentially disengage if lane markings are faded or absent.
The different types of lane keeping assist systems that cars use includes small torque application after crossing a lane line and the Tesla system which keeps the car centered in the lane. The speaker mentions a new car with support for 45 different car makes and emphasizes the importance of genuine customer interaction.
The need for frictionless driver monitoring systems for consumer autonomous cars is critical. Driver monitoring needs to be implemented in a way that cannot be cheated, similar to Tesla's current system.
The need for driver monitoring is discussed in relation to the state of the driver and the car to ensure safe driving, despite Elon Musk's recent claim that it's not necessary.
Researchers are developing reinforcement learning models to train pedestrians to move in more natural ways using real data. The focus is on creating ultra-realistic behavioral models to improve simulations.
Waymo's self-driving cars have a perception stack and planner, but the line between the two is not always clear. The company relies on a combination of hand-engineering and machine learning to make its vehicles more autonomous.
Turning autonomous driving into a learning problem allows for a more efficient and effective approach, rather than relying on hand-coded hacks and modules.
The use of data and machine learning could enable cars to automatically do lane changes and learn appropriate driving policies from good drivers.
The possibility of achieving level five self-driving vehicles without any machine learning is discussed, with questions raised about whether or not it is a feasible approach. The motivations behind other companies' promises to deliver L5 vehicles within the next few years are examined in comparison to Google's Waymo.
The focus on winning self-driving cars as a goal doesn't always align with other opportunities presented, such as selling consumer products or access to data. The prioritization of these opportunities can depend on how they contribute to the end goal of winning self-driving cars.
The use of perception systems, based on both vision and radar, is being considered as a substitute for relying heavily on LIDAR technology in autonomous vehicles.
A first-mover advantage exists in carefully balancing a market with self-driving cars, as switching costs increase for consumers. Chinese knockoff self-driving cars may be attractive for blanketing a city, but the commoditization of the technology could resemble the current state of scooter sharing.
The static driving problem, or the ability to navigate roads assuming you're the only car, is not an easy feat for current technology. This is where Tesla is failing in their lane-keeping system.
Lateral's approach for reinforcement learning on the world will scale to handle both the static and dynamic elements. The future plan is to perform reinforcement learning on the world once all the cars come in.
The human element of catching the system when it fails is key to making reinforcement learning work effectively.
The speaker compares torque limits in their company to Tesla's and analyzes the language used in the Misera documentation, noting it reflects the software's origins in computer science.
The speaker discusses the importance of having real skills and being known for them. They also express admiration for those who contribute to building infrastructure.
The speaker talks about their interest in observing pedestrian behavior and understanding it from a computer vision perspective. They also discuss the potential business value of using technology to monitor employee behavior and reduce accidents.
The discussion covers the trolley problem and the level of technical expertise required to have an opinion on developing reliable and secure systems.
The guest discusses the meaning of work, and the potential merging of virtual reality and AI with sex work.
The speaker discusses their perspective on winning, personal goals, and building a compressive model of the world based on the total number of human and machine flops.
The technical goal of AI is to maximize uncertainty while reducing it at the same time, and the universal reward function is an important aspect of AI.