Frontier Labs All Talk About Safety—Why Hasn't Anyone Copied Waymo's Clunky Approach?

In his tweet, Michael R. Bock raised an overlooked issue: there is a lot of safety discussion in frontier labs, but few people talk about how companies build a safety culture. He worked at Waymo from 2017 to 2019 and saw a different organizational model. Waymo is an autonomous driving company under Alphabet; it grew out of Google's consumer product system and later became a safety-first hardware company. The results are data-backed: about 95% safer than human drivers, with hundreds of millions of rider-only miles.

He notes that many frontier labs today look more like software startups than safety-critical companies. Safety at Waymo permeates every part of the process:

• Layered testing: unit test coverage, subsystem tests, real-driving replay, synthetic simulation, closed-course testing, progressive road testing, and even buying a fake city in the desert to drive in.
• Testing and simulation investment of roughly $1 billion per year in people and compute.
• Gradual deployment: new software first goes onto a small number of cars.
• Safety case construction: using complex statistics to compare robot and human drivers.
• Whistleblower protection: anyone can raise their hand to pause testing or deployment.

These practices are not mysterious. The hard part is that they are expensive and slow, and they require giving pause authority to the front line. Safety culture comes down to budget, testing layers, and who can call a stop. If you only issue statements of principles, don't invest in testing infrastructure, and don't protect whistleblowers, safety culture is just a PR phrase. Frontier labs can keep talking about model capabilities, but without these hard constraints in the organization, safety discussions will stop at the press release.