#421: Google’s Waymo Robotaxi Service Is Progressing, & More

Last week, Google announced¹ that, in fewer than two years after limited public launch, its Waymo subsidiary is commercializing autonomous ride-hail for all in San Francisco and has completed more than 20 million autonomous rider-only miles.² While impressive, our research suggests that Tesla is generating ~66X that amount of driving data, as shown below.

Source: ARK Investment Management LLC, 2024. This ARK research is based on a range of external data sources as of June 2024, which may be provided upon request. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.

Currently, Waymo is the dominant robotaxi service in the US, as Cruise is facing regulatory setbacks,³ and Tesla has yet to launch. Waymo has only 600-700 vehicles in its autonomous fleet and is relying on partners like China’s Zeekr to scale. Waymo also relies on lidar for precision driving but recently ran into a telephone pole in Phoenix. In contrast,⁴ Tesla is vertically integrated and should be able to leverage the data that its ~6 million vehicles have been generating on the road.

We will continue to monitor the progress of robotaxi contenders who are competing for an opportunity that ARK’s research estimates will approach $28 trillion in enterprise value by the end of this decade.⁵

Last week, AI hardware startup Etched.ai raised $120M⁶ to produce chips that inference transformer-based models, the neural network architecture powering large language models like GPT-4o and Llama 3.⁷ Etched is betting that transformer architecture will continue to dominate and that designing hardware to inference transformer models will boost performance significantly relative to Nvidia’s general-purpose graphics processing unit (GPU) hardware. Interestingly, Etched expects that, to inference Llama 3 70B, its hardware will surpass the speed of Nvidia's H100 GPUs by 20x.⁸

That said, Etched's chips face an important tradeoff: although architecture-specific hardware should be much more efficient, if model architectures shift away from transformer technology, its chips would be rendered obsolete. Novel architectures⁹ are evolving and could improve performance significantly.

As AI becomes deeply integrated into our daily lives, hardware customized for specific architectures or models should reduce costs. At the early stages of this revolution, however, who knows which architectures and models will dominate?

Last week, researchers at the Arc Institute published¹⁰ research on a new bridge RNA-based genome editing system—a compact and programmable molecular system that enables the direct insertion of new DNA sequences into the genome. Potentially, bridge RNA-based editing could offer significant improvements over CRISPR gene editing. Already delivering on the promise of curing disease, gene editing based on CRISPR changes only one base pair in the targeted DNA or relies on natural “cell repair mechanisms” to complete the editing.

The new bridge RNA system derives from IS110 family elements, one of the “jumping genes” that can cut and paste itself seamlessly, enabling more precise editing in microbial genomes like bacteria. IS110 elements consist of a gene encoding the recombinase and its non-coding flanking DNA segments. The researchers discovered that those segments can produce a “bridge RNA” that folds into two loops—one that binds to the IS110 element itself and the other to the target DNA in which the element is inserted. Because each of those loops can be altered independently, the researchers were able to program one loop to target any genomic site while programming the other loop to carry any sequence, including a functional copy of the entire disease-causing gene, to be inserted at the target site.

The researchers demonstrated that the new gene-editing system can achieve more than 60% insertion efficiency in E. coli with ~94% specificity at the correct genomic site. In addition to its versatility, this compact system can accomplish all three fundamental DNA rearrangements—insertion, deletion, and inversion.

Having demonstrated that the new editing system works in bacteria, the researchers are aiming to adapt it to function in mammalian cells. If successful, the technology could develop cures for diseases that require highly complex genome-level manipulation.