How self-driving startups Wayve, Waabi, and Autobrains are using end-to-end AI learning, hoping to leapfrog market leaders like Cruise and Waymo
The mainstream approach to driverless cars is slow and difficult. These startups think going all-in on AI will get there faster. Tweets: @mat Tweets: Mat Honan / @mat : This is pretty fascinating: “Throw enough data at the AI and it learns to convert input (camera or lidar data about the road ahead) into output (turning the wheel or hitting the brakes), much like a kid learning to ride a bike.” https://www.technologyreview.com/ ...
Context & Ripple Effects
Waymo has been building its stack around deep learning since at least 2018, when it detailed its collaboration with Google Brain researchers — but its approach still layers learned perception under heavily engineered planning. Wayve staked out the purist alternative back in 2019 with its world-first autonomous drive on roads it never saw during training, mapping camera input straight to steering and braking.
Now Wayve, Waabi, and Autobrains are positioning that end-to-end method as a leapfrog play against Cruise and Waymo, betting cheaper sensors and less hand-engineering can close the gap faster. The timing matters: by late 2022, Bloomberg tallied an estimated ~$100B invested in self-driving startups with little progress to show, so any thesis promising a cheaper path is also a fundraising argument.
First-order effects
- Cruise and Waymo now face challengers whose core claim is structural, not incremental: that end-to-end learning replaces the costly mapping, rules, and sensor suites the leaders spent years and billions perfecting.
- Wayve, Waabi, and Autobrains get a sharper pitch to skeptical investors — a smaller bill for compute, sensors, and engineering headcount per mile of capability gained.
Second-order effects
- The leaders are pushed into defending their architecture publicly rather than just their results — a debate that surfaced years later when Waymo co-CEO Dmitri Dolgov answered claims that Waymo's software relies on hand-coded rules and can't handle freeway driving.
- With capital tightening after the ~$100B bet produced thin returns, funders can split portfolios between incumbent-style scale spend and these leaner end-to-end bets instead of defaulting to the leaders.
Third-order effects
- If end-to-end systems reach parity with far less infrastructure, the industry's moat shifts from mapped cities and fleets toward data pipelines and model training — commoditizing what made Waymo and Cruise expensive to copy.
- If it doesn't, the field consolidates around the fallback Ars Technica identified earlier: commercializing the least demanding applications first, leaving full autonomy to whoever survives the capital drought.
The trend: Autonomous driving is splitting into two camps — incumbents scaling engineered modular stacks and startups betting that end-to-end AI reaches the same goal with less capital — with investor patience after ~$100B of thin returns deciding which camp wins.