What happened
Axios reports that Srikanth Thirumalai, Waymo’s vice president of onboard software, said pure end-to-end AI systems do not meet the company’s safety standard for large-scale autonomous-vehicle deployment. Waymo says it reached that conclusion after testing similar approaches and drawing lessons from more than 200 million driverless miles.
Axios reports that Srikanth Thirumalai, Waymo’s vice president of onboard software, said safely deploying autonomous vehicles at scale requires more than improving an AI model. According to Axios, Waymo has described its objective from the beginning as building “demonstrably safe AI,” and Thirumalai said the company has tested the same general tools used by newer autonomous-driving developers. Waymo’s conclusion, as reported by Axios, is that “pure end-to-end” systems are not able to meet its safety bar at the scale at which it operates. The source does not specify the numerical threshold or testing protocol behind that safety bar.
Axios reports that Waymo’s criticism is aimed at systems that process raw sensor data and directly produce steering commands. The company says such architectures lack enough safety guardrails for its deployment needs. Thirumalai also told Axios that even very large AI models can hallucinate, creating a special problem in physical systems: a vehicle cannot simply be refreshed or restarted after making a bad decision. In the source’s account, Waymo is arguing that an autonomous vehicle must manage the consequences of an error in the real world, where people, vehicles and infrastructure are immediately affected.
According to Axios, Waymo’s approach has changed over time. The company began with many specialized models, including separate systems for tasks such as detecting pedestrians and recognizing when a traffic signal turns green. It later moved toward fewer, larger foundation models while continuing to rely on a broader system of controls. Axios says Waymo published a blog post outlining 10 lessons from more than 200 million autonomous miles. The report describes Tesla, Wayve and Waabi as pursuing more extensive end-to-end systems, which those companies say can support humanlike reasoning. Waymo’s comments are therefore part of an active industry debate, not a settled technical finding.
Axios also reports that Waymo found visibility was better when cameras, lidar and radar were used together than when one or more of those sensor types were removed. Thirumalai told the outlet that AI can only interpret what its sensors detect. Waymo made a similar argument about high-definition maps, which some competitors say are unnecessary. The report does not give comparative test results, specify the environments used, or establish whether the finding applies equally across all autonomous-driving systems.
Read the primary source: axios.com ↗
Why it matters
The position challenges a central premise of the autonomous-vehicle race: that better models and more data could let newer companies catch up quickly. It also highlights an unresolved technical and policy question about how much autonomy should rely on a single neural system versus multiple models, sensors and safety controls.
The practical stakes are unusually high because autonomous-driving software controls physical machines around members of the public. If an AI system misinterprets a scene, the consequence is not merely an incorrect answer that can be regenerated. Axios reports that Thirumalai used this distinction to explain why Waymo believes physical AI needs stronger safeguards than a general-purpose model. The company’s argument implies that reliability must be demonstrated across long-running real-world operation, not inferred from model size, benchmark scores or the apparent fluency of a system’s reasoning.
The report also frames the dispute as a question about competitive timing. Axios says advances in data collection and smarter models have encouraged hopes that newer companies could find a shortcut to self-driving. If those hopes are justified, rivals could reduce the importance of Waymo’s long head start. If Waymo is right that safe deployment requires extensive testing, multiple layers of control and broad sensor coverage, autonomous vehicles may remain a slower and more expensive engineering project. The source does not independently establish which side of that debate is correct.
Waymo’s position matters for how safety standards may be defined. A requirement for several sensor types, detailed maps or layered model architectures could improve redundancy, but it could also raise development and operating costs. Conversely, accepting a single end-to-end system could simplify a vehicle’s software stack while placing more responsibility on one learned model. Axios reports that Waymo’s safety argument also serves its competitive interests because the company has accumulated years of testing and real-world driving data. That incentive does not invalidate the claim, but it is relevant context for evaluating it.
The story is significant beyond the companies named because it concerns how regulators, insurers and the public may judge autonomous systems. A vehicle that performs well in demonstrations may still face difficult edge cases, sensor limitations or failures that are hard to reproduce. The report offers no independent assessment of Waymo’s record, no comparison with rival systems, and no evidence that one architecture is universally safer. Its main contribution is to document Waymo’s current technical position and the reasoning the company is presenting publicly.
What to watch next
The key test will be public evidence from Waymo and its competitors about safety performance, deployment conditions and failure handling. The source does not provide exact safety thresholds, comparative failure rates, or independent verification of Waymo’s claims.
The most important follow-up is measurable evidence. Readers should look for publicly documented information about where each autonomous-driving system operates, how often human intervention or remote assistance is needed, how failures are classified, and how safety performance is compared across conditions. Axios reports Waymo’s experience in terms of more than 200 million driverless miles, but the source does not provide a denominator for incidents, a standardized comparison with competitors, or an independent audit of that figure’s safety implications.
Sensor strategy will be another dividing line. Axios says Waymo’s training showed better visibility when cameras, lidar and radar worked together, while also reporting that Tesla is pursuing a camera-only approach. Further reporting should establish whether these results hold across weather, lighting, road layouts and unusual objects, and whether the additional sensors create meaningful improvements relative to their cost and maintenance requirements. The source does not say that cameras alone are impossible, only that Waymo says better AI cannot fully compensate for inadequate sensing.
Mapping requirements also deserve scrutiny. Axios reports that some rivals consider high-definition maps unnecessary, while Waymo argues that mapping remains important. The unresolved issue is how each system handles road changes, construction, unusual traffic patterns and locations outside its normal operating area. Public documentation from the companies or regulators could clarify whether maps are safety-critical, optional support tools, or a constraint on deployment. None of those details is established in the source.
Finally, watch whether AI advances alter the architecture debate. Axios reports that Waymo believes there is no single “silver bullet,” but the article also notes that rapid progress could still produce a faster route to autonomy. Meaningful unknowns include how Waymo defines its safety bar, what specific guardrails its systems use, how its tested end-to-end components are integrated, and whether Tesla, Wayve or Waabi can demonstrate comparable safety at scale. Until those questions are answered with comparable public evidence, Waymo’s claim should be treated as an attributed industry position rather than a definitive verdict.
