<h1>Elon Musk Tests Tesla’s Full Self Driving Technology with a Twist</h1>
<p>On a late-August Friday, Elon Musk decided to embark on a unique adventure. He got into his trusty Model S at Tesla headquarters in Palo Alto, set a random destination on his navigation screen, and let the car take the wheel using its remarkable Full Self Driving (FSD) technology. As his car glided along for 45 minutes, with the soothing music of Mozart playing in the background, he livestreamed the journey for his followers. Along the way, he even drove past the residence of Facebook’s Mark Zuckerberg, playfully challenging him to a cage-match fight. “Maybe I should knock on his door and ask if he’s up for some hand-to-hand combat,” Musk joked before continuing on his autonomous adventure.</p>
<p>Musk chose to test the latest version of FSD, known as FSD 12, on August 25, 2023. This new iteration of the technology marked a significant departure from its predecessors. FSD 12 was built on a revolutionary concept that Musk believes will not only revolutionize autonomous vehicles, but also serve as a monumental step towards achieving artificial general intelligence in real-world settings. Unlike previous self-driving software versions, which relied on hundreds of thousands of lines of code, FSD 12 taught itself how to drive by analyzing billions of video frames showing human driving behavior. Think of it as a large language model chatbot, but for cars.</p>
<p>Surprisingly, this groundbreaking approach was conceived just eight months prior to Musk’s test drive.</p>
<h2>Introducing the “ChatGPT for Cars”</h2>
<p>In a meeting with Tesla’s autopilot team in December, Dhaval Shroff, a young member of the team, described the idea they were working on as “like ChatGPT, but for cars.” Shroff was referring to the newly released chatbot developed by OpenAI, a lab co-founded by Musk in 2015. He explained that their approach involved processing extensive data on how real human drivers navigate complex driving situations and then training a computer’s neural network to imitate that behavior.</p>
<iframe src=”https://www.example.com/video” allowfullscreen></iframe>
<p>Until that point, Tesla’s Autopilot system relied on a rules-based approach. The car’s cameras would identify lane markings, pedestrians, vehicles, signs, and traffic signals. The software would then apply a set of rules, such as stopping at red lights, moving on green lights, staying within lane markers, and proceeding through intersections when it’s safe. Tesla engineers manually wrote and updated hundreds of thousands of lines of code to handle complex scenarios.</p>
<p>The new approach, called the “neural network planner,” took a different path. Shroff elaborated, “Instead of relying on rules to determine the car’s path, we now rely on a neural network that uses millions of examples of human driving behavior.” In essence, it is imitation of human driving. When faced with a situation, the neural network selects a path based on how humans have handled similar situations in the past. This mirrors how humans learn various skills like language, driving, chess, or even eating spaghetti. Although we may receive a set of rules, we mostly learn by observing others. This approach to machine learning was conceptualized by Alan Turing in his 1950 paper, “Computing Machinery and Intelligence,” and gained widespread attention with the release of ChatGPT a year ago.</p>
<h3>Training the Neural Network</h3>
<p>By early 2023, the neural network planner project had analyzed a staggering 10 million clips of video collected from Tesla customers. However, it wasn’t just any video data. The team made a deliberate choice to only use data from humans who demonstrated excellent driving skills, as assessed by human labelers. These labelers, many based in Buffalo, New York, reviewed the videos and provided ratings. Musk instructed them to look for behaviors displayed by “a five-star Uber driver,” and these were the videos used to train the neural network.</p>
<p>Musk made it a habit to wander through the Autopilot workspace in Palo Alto, frequently engaging in impromptu discussions with engineers. As he delved into the human-imitation approach, he couldn’t help but question its necessity. Wasn’t it a bit of overkill? Musk was a firm believer in utilizing the simplest solution for any given problem. Shroff presented him with scenarios where the neural network planner outperformed the rules-based approach. In one demonstration featuring a road cluttered with obstacles like fallen traffic cones and trash cans, the car guided by the neural network planner skillfully navigated through, occasionally crossing lane lines and bending the rules as needed. “Look at the transition from rules-based to network-path-based,” Shroff highlighted. “With this approach, the car can navigate unstructured environments without any collision risk.”</p>
<p>This glimpse into the future delighted Musk. Excitedly, he exclaimed, “We need a James Bond-style demonstration with bombs exploding and a falling UFO, while the car flawlessly zooms through without hitting anything!”</p>
<h4>Measuring Success with Mileage</h4>
<p>Machine learning systems often rely on a metric as they train themselves. Musk, who had a penchant for setting parameters, provided the team with their guiding star: the number of miles driven by cars equipped with Full Self-Driving technology without human intervention. “Make sure that the starting slide of every meeting displays the latest data on miles driven without intervention,” Musk decreed. He wanted it to be like a video game where they could see their score increase every day. Musk enthusiastically declared, “Video games without a score are boring, so it will be motivating to watch our miles-per-intervention numbers grow.”</p>
<p>To create a sense of excitement and accomplishment, the team installed massive 85-inch television monitors in their workspace. These screens displayed real-time updates on the average number of miles driven without human intervention by FSD-equipped cars. Additionally, they placed a gong nearby. Whenever the team successfully resolved a problem leading to intervention, they joyfully banged the gong, celebrating their achievement.</p>
<h5>A Test Drive with Musk at the Helm</h5>
<p>In mid-April 2023, the time had come for Musk to experience the neural network planner firsthand. He sat in the driver’s seat alongside Ashok Elluswamy, Tesla’s director of Autopilot software. Three members of the Autopilot team joined them in the back seats. As they prepared to leave the parking lot at Tesla’s Palo Alto office complex, Musk selected a destination on the map and relinquished control to the self-driving system. The first challenge arose when a bicyclist approached their path. Without any human intervention, the car yielded, mirroring how a human driver would respond.</p>
<p>For the next 25 minutes, the car flawlessly navigated fast roads, neighborhood streets, complicated turns, and skillfully avoided cyclists, pedestrians, and even pets. Musk never had to touch the wheel. Only a few times did he take control by gently tapping the accelerator when he felt the car was excessively cautious, such as hesitating too long at a four-way stop sign. At one point, the car executed a maneuver that even Musk himself admitted he couldn’t have done better. “Oh, wow,” he exclaimed, “even my human neural network failed here, but the car made the right decision.” Filled with delight, Musk began whistling Mozart’s “A Little Night Music” in G major.</p>
<iframe src=”https://www.example.com/livestream” allowfullscreen></iframe>
<h6>Impressed and Ready to Forge Ahead</h6>
<p>As the drive came to an end, Musk expressed his admiration, declaring, “This is truly remarkable, guys. Amazing work.” They all proceeded to the weekly Autopilot team meeting, where the verdict would be delivered. Around the conference table sat 20 individuals, almost all dressed in black T-shirts, who eagerly awaited Musk’s assessment. Many had been skeptical about the neural network project’s feasibility, but Musk converted them into believers. He urged them to redirect resources toward further advancement of the technology.</p>
<p>During the discussions, Musk seized upon a crucial discovery made by the team: the neural network’s performance significantly improved after training on at least one million miles’ worth of driving data from FSD-equipped cars.</p>
Denial of responsibility! Vigour Times is an automatic aggregator of Global media. In each content, the hyperlink to the primary source is specified. All trademarks belong to their rightful owners, and all materials to their authors. For any complaint, please reach us at – [email protected]. We will take necessary action within 24 hours.
