Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
It is now more than a decade since I took my first ride in a self-driving car. As a Google prototype zipped me around Mountain View in Silicon Valley, what promised to be an exhilarating taste of the future quickly became quite boring — the robot car was just such a smooth driver. A 2017 deadline for bringing autonomous vehicles to market, which was set in 2012 by Google co-founder Sergey Brin, seemed aggressive but not implausible.
Of course, Google failed that deadline, just as Elon Musk’s Tesla has done over the years on autonomous vehicles.
I was reminded of these shaky timelines when reading Sam Altman’s new year’s “reflections”. The OpenAI chief is “now confident we know how to build” artificial general intelligence — the point at which AI’s capabilities surpass those of most humans — and predicts that “in 2025, we may see the first AI agents join the workforce”.
AI agents that can plan and carry out a series of tasks on behalf of a user — such as buying items or booking trips — are the industry’s latest holy grail. Many see general-purpose agents as a necessary component to reaching AGI [artificial general intelligence], though they will start out focused on specific tasks. But as the autonomous driving industry has found over the past decade, there is a big difference between AI that works well in a constrained environment, like a chatbot window, and a “free range” agent let loose in the real world.
There is no doubt that 2024 was an inflection point for Waymo, as the Google self-driving car project was renamed in 2016. Before last year, Waymo had ferried public passengers for 1mn miles since the initiative began in 2009. In 2024, that total rose by a further 4mn as Waymo extended its service to the public from Phoenix to San Francisco and Los Angeles.
Tekedra Mawakana, Waymo’s co-chief executive, said at the Consumer Electronics Show this month that Waymo’s robotaxis were now doing 1mn miles a week — “more than a human drives in a lifetime”. According to Waymo’s own studies on more than 25mn miles of autonomous testing, its system is also safer than humans, with vehicles involved in far fewer serious collisions than when a person is behind the wheel.
Waymo’s acceleration was all the more striking in a year that saw General Motors shut its Cruise robotaxi unit after pouring $10bn into the venture since 2016, as well as Apple abandoning its Project Titan car project.
The autonomous vehicle optimism of the mid-2010s has returned. “The AV revolution has arrived — after so many years,” said Nvidia chief Jensen Huang in his own CES keynote, predicting that autonomous vehicles would be “the first multi-trillion-dollar robotics industry”.
Waymo is winning by resisting, not riding, the hype. Cognisant of the backlash that forced Uber out of AVs after one of its vehicles killed a pedestrian, Waymo has prioritised safety over growth. Even after Waymo moved its vehicles out of the “lab” and into public use, it still spent more than three years testing its robotaxis in Phoenix — a less complex environment than a city like San Francisco.
Cracking the big city still required ironing out some very weird bugs, such as the bizarre 2am outbreaks of honking that woke up San Francisco residents near a Waymo parking lot last summer. It is still figuring out other complications that arise when AI meets humans, such as how messy its cars can get when there’s no driver to keep an eye on passengers.
The company’s long robotaxi journey stands in contrast to ChatGPT’s breakneck growth, with the chatbot attracting more than 300mn users in just two years. Scaling an app and deploying a fleet of taxis are very different propositions. Nonetheless, even if generative AI agents operate in a purely digital world, there is a huge gulf between co-pilots that work alongside people and fully autonomous bots that can reliably roam the open web.
The safety issues may not be life and death with online agents but the risks are no less real. Will AI developers or credit card companies, for example, shoulder the cost when a hallucinating agent goes on an unauthorised spending spree? And even with current AI tools, many businesses are finding that uncooperative employees and mismatched organisational structures present greater barriers than shortcomings in the technology itself.
Altman might not be a whole decade out in his prediction that AI agents will reach their potential. But making a free-range AI that is as reliable as a human may yet prove just as challenging as Waymo’s long road to robotaxis.
tim.bradshaw@ft.com
