Waymo CEO: L2 to L4 Upgrade Possible in 2026? End-to-End AI Alone Isn't Enough

In a rare and candid interview, Waymo Co-CEO Tekedra Mawakana has addressed one of the most debated questions in the autonomous vehicle industry: Can a Level 2 driver-assistance system be upgraded to a full Level 4 self-driving system? Her answer is a qualified yes — but with a critical caveat that challenges the current industry trend of relying solely on end-to-end neural networks.

The Limitations of End-to-End AI for L2 to L4 Upgrade

According to a conversation published by Sixth Street, Mawakana explained that while the technological path from L2 to L4 exists, it cannot be achieved through end-to-end learning alone. “End-to-end models are powerful, but they lack the interpretability and safety guarantees required for L4 operations,” she said. Instead, Waymo is pioneering a hybrid approach that combines cloud-based foundation model distillation with language-aligned world models.

Why End-to-End Alone Falls Short for L4 Upgrade

The autonomous driving industry has seen a surge in companies advocating for end-to-end AI systems, where a single neural network maps raw sensor inputs directly to driving commands. However, Mawakana argues that this approach, while impressive in demos, fails to meet the rigorous safety standards required for commercial robotaxi services.

According to a New York Times report from March 2026, Mawakana made a forceful case for the safety of driverless cars, stating that “safety cannot be an afterthought; it must be engineered into the system from the ground up.” She pointed out that end-to-end models often behave as black boxes, making it difficult to diagnose failures or ensure consistent performance across rare edge cases.

Cloud Distillation: A Path to Safer Self-Driving Technology

Waymo’s alternative strategy involves distilling knowledge from large, cloud-based foundation models into on-vehicle systems. This process, known as cloud-based foundation model distillation, allows the car to leverage the vast reasoning capabilities of a large model while maintaining the low latency and computational efficiency needed for real-time driving.

Language-Aligned World Models: The Missing Piece for L4 Safety

Beyond distillation, Mawakana emphasized the importance of language-aligned world models. These models use natural language as a bridge between human understanding and machine perception. By aligning the car’s internal representation of the world with human language concepts — such as “a child running toward the street” or “a construction zone ahead” — Waymo aims to create a system that is both more interpretable and more robust.

How Waymo Builds Trust Through Transparency

As noted on Waymo’s official leadership page, Mawakana’s background in policy and safety has shaped the company’s approach. She previously served as Waymo’s Chief Safety Officer and has been a vocal advocate for transparency in autonomous vehicle testing. Her co-CEO role, shared with Dmitri Dolgov, reflects Waymo’s dual focus on technology and trust.

Real-World Results: Waymo's Safety Record

The company’s approach is already yielding results. Waymo’s fleet, operating in Phoenix, San Francisco, and Los Angeles, has logged millions of miles without a single at-fault accident. Mawakana attributes this safety record directly to the company’s multi-layered architecture, which includes traditional planning and control modules alongside machine learning components.

The Future of Autonomous Driving: L2 to L4 Upgrade Path

Industry analysts have taken note. While Tesla and other automakers push for a purely vision-based, end-to-end upgrade path from L2 to L4, Waymo’s more conservative, sensor-rich strategy is gaining credibility. “The industry is realizing that scaling autonomy isn’t just about data volume — it’s about data quality and system design,” Mawakana said.

Looking ahead, Mawakana envisions a future where L2 systems can indeed evolve into L4 systems, but only if manufacturers invest in the foundational infrastructure — cloud computing, language models, and safety validation — that Waymo is building today. “The upgrade path exists, but it’s not a shortcut. It’s a complete rethinking of how cars understand the world,” she concluded.

Key Takeaways for L2 to L4 Upgrade

• End-to-end AI alone is insufficient for L4 safety
• Cloud distillation enhances reasoning without latency
• Language-aligned world models improve interpretability
• Waymo's approach combines AI with traditional control systems