The Ultimate AI "Killer App"

Will Waymo hold its lead over Tesla in the race to an autonomous future?

An intense partisan divide exists not only in the political sphere but also in the realm of autonomous vehicles. Industry professionals, AI scientists, venture capitalists, and investors have formed opposing camps regarding who is on the right technology path and leading the race to deploy robotaxi services at scale. This debate can be best delineated into two camps: Tesla and Waymo, with Nvidia acting as an autonomous technology enabler and Mobileye as a potential third-party supplier of autonomous systems.

Autonomous transportation is the ultimate AI killer app. Success in solving this highly complex problem will unlock an enormous market opportunity. However, nothing this promising is easy. It requires enormous technological breakthroughs and a risk appetite to challenge conventional wisdom.

Waymo’s Competitive Edge

The prospects in the robotaxi market currently favor Waymo over Tesla, largely due to Waymo’s pragmatic strategy to autonomous driving technology. Waymo's "compound AI system" approach enables a more robust and commercially viable platform.

This multi-faceted approach, combining high-definition mapping, rigorous safety protocols, and modular AI frameworks with extensive use of neural networks, ensures greater reliability—a crucial factor for commercial deployment. This layered system not only reflects the adoption of the most viable technology today but also provides a flexible foundation for future advancements, including the eventual adoption of end-to-end AI systems as the technology matures.

Waymo's technology strategy is already delivering tangible results, as demonstrated by the continuous expansion of its commercial robotaxi services. The company has successfully launched fully autonomous operations in Phoenix, San Francisco, and Los Angeles, with plans to expand into Atlanta, Austin, and Miami. Additionally, Waymo has announced plans to test in 10 more cities in 2025, further solidifying its leadership in autonomous mobility. Moreover, the company has begun testing autonomous vehicles on Los Angeles freeways, showcasing its readiness to handle complex, high-speed environments.

Despite Waymo's early lead in the commercial deployment of fully autonomous robotaxi services, Alphabet's stock price arguably does not ascribe any value to Waymo. Conversely, Tesla's stock valuation already embeds a substantial option value tied to the future potential of its robotaxi service and autonomous vehicles. This leaves Tesla’s share price exceptionally vulnerable to valuation compression if there’s any slippage relative to expectations in its robotaxi timeline, particularly given the current state of its EV business.

Tesla’s Ambitious Robotaxi Strategy

Tesla’s pursuit of a pure end to end neural network system is technically more challenging and represents the holy grail of autonomous vehicle systems. If successful, this approach would offer scalability and further reduce the need to integrate rules based systems. The promise of an end-to-end system is that it can generalize across different environments, enabling deployment of robotaxi services to new markets with fewer adjustments.

While this approach holds significant promise, it still faces inherent technical challenges in handling real-world variability and edge case scenarios, which could take considerable time to overcome. If Tesla remains entrenched in this technical approach, it risks a slower-than-expected expansion of its robotaxi service, potentially allowing Waymo to further capitalize on its first-mover advantage.

Although the latest version of Tesla's Full Self-Driving (Supervised) system demonstrates noticeable improvements and impressive driving performance under certain conditions, the lack of objective data makes it difficult to quantify its rate of improvement. In assessing improvement in FSD performance, there is a heavy reliance on anecdotal accounts and online videos from Tesla enthusiasts. These often understate the gap between a well-performing advance driver assisted system and a truly unsupervised, driverless system.

On Tesla's most recent earnings call, Elon Musk expressed strong optimism about the outlook for FSD and the company's robotaxi service, projecting a 2026 expansion beyond the initial rollout. Tesla announced plans to launch an unsupervised robotaxi service in Austin in June 2025, with ambitions to expand into several additional cities by year-end. This rollout will utilize an internal fleet of Tesla vehicles, with no initial inclusion of individually-owned vehicles—a distinction from Musk's shared fleet vision, where Tesla owners could deploy their vehicles in the robotaxi network.

Despite reiterating the 2025 target, the scope of Tesla’s initial robotaxi launch appears somewhat modest compared to expectations. Like Waymo’s service, this launch will begin in a predefined geographical area. However Elon Musk addressed this on the earnings call -

It’s not that it doesn’t work beyond Austin. In fact, it does. We just want to put our toe in the water, make sure everything is okay.

The launch of its robotaxi service in a fixed geographical area is reasonable given the practical challenges of deploying a unsupervised robotaxi service. Furthermore, unsupervised FSD will likely perform better in familiar, frequently travel roads, making this type of environment ideal for the initial rollout.

In the past, the debate surrounding autonomous vehicle (AV) technology leadership and the timeline for widespread robotaxi deployment was often clouded by overly confident, unsubstantiated hype. Many projections were driven more by opinion than fact, resulting in subjective perspectives and excessively optimistic forecasts about the imminent arrival of Level 5 autonomy—vehicles capable of operating in all conditions without the need for a human driver or occupant.

This wave of unfounded optimism epitomized the classic Silicon Valley mantra of "fake it until you make it." However, the industry has now shifted from theoretical promises to tangible progress. The early-stage commercial deployment and expansion of robotaxi services are finally upon us.

Solving Closed Domain vs. Open Domain Problems

To better understand the different approaches to vehicle autonomy its helpful to have a clear framework for distinguishing between closed-domain and open-domain problems.

Closed-domain problems are more fixed and tractable. They often involve a finite set of possible scenarios and solutions. A classic example is a chess game, where the rules are fixed, and the possible moves are limited. Essentially solving a closed-domain problem leverages AI as a sophisticated process of mathematical approximation.

Another example is Netflix's recommendation system. The system operates within a defined domain: Netflix's content library. The items (movies and shows) and their attributes (genre, actors, etc.) are structured and controlled. This allows the system to leverage AI for sophisticated pattern recognition and prediction within this specific domain.

In contrast, open-domain problems are characterized by their vastness and unpredictability. They involve an infinite number of potential scenarios and solutions, making them far more challenging for AI systems to master. Solving a complex open-domain problem may require capabilities associated with intelligence, not just mathematical function approximation.

Driving a car is an open-domain problem with an infinite long tail of difficulty and unknown scenarios. The real world presents an endless array of unpredictable situations, from changing weather conditions to unexpected obstacles to the unpredictable behavior of other drivers. Solving this complex problem requires capabilities beyond simple pattern recognition. It arguably necessitates elements of general intelligence, such as the ability to reason abstractly, adapt to new situations, and understand the physical environment in a nuanced way.

Some AI applications blur the lines between closed and open-domain problem-solving, taking a hybrid approach. For instance, Meta's recommendation engines leverage both domain-specific knowledge and real-time adaptation. The system primarily operates within the closed domain of Meta's platforms and content. However, it also utilizes machine learning and generative AI to adapt to evolving content and user preferences in real-time, a characteristic more commonly associated with solving open-domain problems.

Taking Different Paths to Autonomy

Waymo is taking a hybrid approach to autonomy. While primarily solving for an open-domain problem, it also incorporates aspects of a closed-domain strategy to enhance safety and performance. The system is optimized to perform within a defined geographical area but can also operate in unfamiliar environments with unforeseen edge cases, leveraging advanced neural networks to adapt in real time.

Tesla's approach prioritizes solving for a open-domain problem. Although Tesla may still incorporate some structural elements, its approach leans heavily on an end-to-end neural network system.

Waymo's vehicles operate in fixed, mapped-out environments, reflecting some attributes of solving for a closed-domain problem. This enhances safety, performance, and operational efficiency. Using maps is somewhat analogous to a driver's memory, allowing for safer navigation in familiar areas.

Waymo’s strategy is grounded in tangible proof points. The company’s commercial robotaxi service has shown rapid growth in ride volume and geographical expansion. With services now operational in multiple cities and plans for further expansion, Waymo demonstrates a consistent trajectory toward broader deployment. The Waymo camp also points to a more robust sensor suite, which provides a more comprehensive picture and understanding of the vehicle’s environment, as a system advantage.

On the Tesla side are Elon Musk's dedicated advocates, who overlook years of unrealized forecasts that unsupervised Full Self-Driving (FSD) was just around the corner.

Tesla provides limited FSD performance metrics and data, making it difficult to quantify its progress. However, the company does provide a vehicle safety report that offers quarterly figures on miles driven per accident when Tesla’s Autopilot technology is engaged. This report appears to encompass miles driven for both standard Autopilot and FSD. However, this data lacks granularity—there is no breakdown of the proportion of miles driven on highways versus more challenging city miles.

While this data shows steady improvement over the years, the trend appears to have stalled since 2022. Although there may be some seasonal environmental factors to consider, oddly in first and third quarter 2022 Tesla recorded one crash every 6.57 million and 6.26 million miles driven, respectively, compared to 5.94 million miles in fourth quarter 2022.

Tesla management also often makes more general comments about FSD progress. For example, during the October 2024 earnings call, Ashok Elluswamy, VP of AI Software, stated that in terms of miles driven between critical interventions there had been a 100-fold improvement in FSD performance with version 12.5 since the start of 2024, with a further step change in performance expected with version 13. However, without a reliable baseline and historical data, it is difficult to assess what a 100x improvement truly means. Understanding the current and past frequency of critical interventions is crucial for evaluating FSD's progress.

While the numerous videos posted by Tesla owners suggest consistent improvement in performance and highlight prolonged periods or trips with few or in some cases no critical interventions, these anecdotal accounts still reflect a system that may fall short of being imminently ready for unsupervised operations. Although imperfect and a limited data set, crowdsourced data from the FSD Community Tracker also points to a continued need for frequent critical interventions.

Ultimately, the pubic is left relying on Tesla management's public pronouncements about the impending exponential improvement in FSD's performance. While a technological breakthrough is possible, the question remains: if the system is improving so rapidly, why doesn't Tesla provide more detailed, objective data to substantiate these claims?

Nevertheless, there is a possibility that Tesla's path to fully autonomous vehicles is set to accelerate to a point that finally delivers a large-scale network of driverless vehicles. Ultimately, it is Musk's track record and determination that give his advocates the confidence that he is on the right path.

Compound AI System vs. End-to-End Neural Networks

The debate between compound AI systems and end-to-end neural networks highlights a critical choice in autonomous driving technology, each approach having its own set of trade-offs, advantages, and challenges.

Waymo’s compound AI system uses neural networks to handle complex, dynamic data, integrated with rules based systems. The Waymo team has adopted a flexible approach, remaining open to alternative technical paths, including exploring the potential of a pure end-to-end system. Building upon the Waymo Foundation Model, the Waymo team recently introduced an End-to-End Multimodal Model for Autonomous Driving (EMMA) in a published paper highlighting the advantages and challenges of such a system and how it can contribute to creating a more generalizable and adaptable driving system.

As AI technology advances, utilizing a pure end-to-end neural network system may eventually prove to be the optimal approach. However, the available data continues to support the notion that a compound AI system still results in a better performing system, and this may remain the case for sometime.

A compound AI system is an approach dictated by a willingness to make trade-offs to achieve reduced generalization error. As an article posted by Mobileye's CEO, Professor Amnon Shashua, and Professor Shai Shalev-Shwartz explains:

The grand question, driving any school of thought for building a data-driven system, is the "bias/variance" tradeoff. Bias, also known as "approximation error," means that our learning system cannot reflect the full richness of reality. Variance, also known as "generalization error," means that our learning system overfits to the observed data and fails to generalize to unseen examples.

The total error of the learned model is the sum of the approximation and generalization errors, so in order to reach a sufficiently small error, we need to delicately control both.

Mobileye and Waymo's use of compound AI systems deliberately introduces bias to reduce generalization error. Generalization refers to an AI model or system's ability to perform well on new, previously unseen data. This approach is predicated on the assumption that this process will still result in a smaller total error rate than an end-to-end system.

In contrast, an end-to-end network directs incoming sensor data directly to vehicle control decisions, potentially avoiding this bias. However, it may overfit to the observed data, resulting in a higher generalization error that could prompt more frequent critical disengagements. While an end-to-end system theoretically aims to eliminate all "edge cases," its performance may also be constrained by the size of the onboard computer and memory. This limitation arises because compensating for the higher generalization error inherent in end-to-end systems requires vast amounts of data.

Tesla's commitment to an end-to-end system offers potential advantages, including better driving performance, greater scalability, and adaptability to diverse conditions. However, relying solely on ever-larger datasets and compute power does not guarantee overcoming the inherent challenges of end-to-end systems when encountering unforeseen edge cases.

For example, these systems can exhibit “casual confusion”. This occurs when a model learns incorrect cause-and-effect relationships due to biases in the dataset. In other words, the model confuses correlation with causation, that could lead to incorrect generalization when encountering new data or edge cases.

Two additional challenges for end-to-end systems are overfitting and “catastrophic forgetfulness.” Overfitting occurs when the system overfits to observed data and fails to generalize when encountering previously unseen, new corner cases. Catastrophic forgetfulness happens when a neural network forgets previously learned information as new data essentially overwrites existing knowledge. This can be problematic when continuous adaptation to new environments is required.

As AI technology evolves, the balance may shift. Researchers are actively working on ways to make end-to-end systems more reliable, and future innovations may overcome the issues of overfitting, catastrophic forgetfulness, and casual confusion. However, for now, the compound AI system seems to offer the most reliable approach for deploying a robotaxi service.

The Race Continues

The final chapter of the autonomous vehicle technology race has yet to be written. Waymo continues to execute with little fanfare, while Elon Musk makes optimistic forecasts.

In the near term, Waymo appears to be in the lead. However, Musk's incredible force of will suggests that it would be unwise to count him out.