June 19, 2024 – At the 2024 annual shareholder meeting of Tesla, CEO Elon Musk discussed a significant challenge facing the company’s Fully Self-Driving (FSD) system. This challenge, which aligns with speculations made by industry insiders two years ago, could potentially impact Tesla’s goal of achieving autonomous taxi services.
According to Musk, as the FSD system continues to advance, determining which Artificial Intelligence (AI) model is superior becomes increasingly difficult. The challenge lies in how to quickly evaluate the performance of a new model when interventions are only required after thousands of miles of driving.
“These different AI models don’t perfectly solve all problems,” Musk admitted. “One model might solve problem A, but it might introduce problem B.”
Zachary Shahan, an editor at Cleantechnica, had anticipated this “seesaw problem” in Tesla’s FSD system two years ago. He speculated that the system might introduce new issues while fixing existing ones, hindering overall progress.
To address this challenge, Musk explained that Tesla is employing various strategies, including simulation testing and shadow mode operation. “Having some vehicles not enabled with FSD can actually be helpful,” he said. “We can use shadow mode to compare the driving behavior of new models with user behavior, revealing the strengths and weaknesses of different models.”
Tesla can test its AI models using its fleet of millions of cars, comparing the predicted behavior of the AI with actual user driving behavior to determine which model performs better.
“Currently, the biggest limitation is not training data but the efficiency of testing AI models,” Musk stated. “And figuring out how to cleverly determine if a new model is better. For instance, we know that certain intersections are very complex, and all models perform well on good roads without complex intersections. So, we have selected thousands of complex intersections in the US specifically to test the performance of new models.”
It’s worth noting that Musk also mentioned that Tesla has not yet fully unleashed the potential of its Hardware 4 autonomous driving hardware and plans to upgrade it later this year.
In summary, overcoming the “seesaw problem” and rapidly iterating FSD models remain crucial challenges for Tesla in achieving fully autonomous driving and autonomous taxi services.