After DeepSeek, the AI giants still have plenty of work left to do

11/29/2024, 7:56 AM
fastcompany.com

The tech industry has long been infatuated with metaphors relating to the original space race. Its favorite one is moon shot, a term it applies to any undertaking of atypical ambition. But Chinese startup DeepSeek’s release of a reasoning AI model that may be a peer of OpenAI’s GPT-o1—despite having been created on the cheap, without access to Nvidia’s best chips—has everyone reaching back to 1957’s original Sputnik moment as a point of comparison.

It somehow took most people a week to pay attention to DeepSeek’s R1, which the company released on January 20. Once they did, it spawned an insta-frenzy whose shockwaves ranged from the technological to the geopolitical. They include a stock market beating for Nvidia and other chipmakers, new questions about whether vast resources actually provide an edge in AI after all, and shock that the Biden administration’s bans on shipping the most powerful U.S.-designed chips to China didn’t prevent that country’s researchers from making a possibly epoch-shifting breakthrough with the stuff they had on hand.

DeepSeek’s abrupt impact has undeniable similarities to the panic set off more than 67 years ago when the U.S.S.R. successfully put a satellite into orbit before the U.S. did. But as former Reddit CEO Yishan Wong pointed out in a post this week, the parallels are shallow. For one thing, the Soviets worked in deep secrecy. By contrast, DeepSeek is publishing code and research relating to its techniques for creating AI that does more with less. That gives the entire world the opportunity to quickly build upon what the company has created, potentially accelerating AI’s use everywhere rather than preserving a daunting competitive advantage for one company or country.

To be sure, the sudden commodification of AI could have profound implications for the handful of powerful U.S. companies that have hitherto propelled the technology forward. But while the details and timing of such an inflection point were unpredictable, its inevitability was not. For example, an internal Google document leaked in early 2023 was titled “We Have No Moat and Neither Does OpenAI.” Or, as Microsoft CEO Satya Nadella put it when I talked to him later that year, “As far as I’m concerned, early leads in technology don’t matter.”

DeepSeek’s R1—and other AI technologies modeled upon its approach—may well force AI’s incumbent giants to reassess everything about their future. Yet that’s hardly an end game for the industry as we’ve known it. Artificial intelligence isn’t anywhere near hitting an insurmountable wall that prevents further progress, and it’s tough to imagine that companies with access to vast resources won’t be able to unlock some advances that those operating under greater constraints cannot.

Most importantly, the dizzying improvements we’ve seen in LLMs over the past few years have yet to be matched by the real-world AI in applications we use. As generative AI’s novelty wears off, tools such as Microsoft’s Copilot look like rougher and rougher drafts of something that needs further ingenuity to live up to its potential. The work of hooking up AI to all the processes we use to get stuff done has barely begun, and a lot of money stands to be made by the companies who get it done.

That’s the underyling fact behind the industry’s obsession with so-called agentic AI—a slightly annoying buzzword that encompasses forms of the technology that can perform complex tasks without constant human oversight. There are some decent early stabs at the idea out there, such as Asana’s “AI teammates,” which already shoulder some of the grunt work of wrangling tasks in the project-management app. But those examples are outnumbered by instances of agentic AI that mostly prove the technology isn’t ready to do much on its own.

Last week, for example, OpenAI released Operator, a “research preview” available to users of its $200/month ChatGPT Pro tier. Operator can type into a web browser and control a mouse pointer, a theoretical first step toward letting it handle all the tasks we humans perform on the web. Over at Platformer, Casey Newton reported on his hands-on experience with the service, which included asking it to perform tasks such as writing a high school lesson plan for The Great Gatsby. It took minutes to achieve results that were no better than what the non-agentic ChatGPT came up with almost instantly. And when Newton tried to use Operator to order groceries—something a stock chatbot can’t do—it turned out that the current version of Operator is pretty hopeless at the job, too.

In December, I got a demo of Google’s experimental agentic AI, “Project Mariner,” that also involved grocery ordering and was too glacially slow to look like progress. That Operator and Mariner aren’t yet ready to handle a humble task such as buying a gallon of milk isn’t evidence that they’re exercises in futility—just that the goal of making AI usefully agentic remains largely aspirational, even at OpenAI and Google.

DeepSeek and other feats of LLM optimization yet to come won’t get in the way of further development of agentic AI. Indeed, they’ll surely help by making the underlying infrastructure more accessible to more people with good ideas. Even then, the U.S.’s AI kingpins will maintain some distinct advantages, from the money and engineering talent they can throw at tomorrow’s challenges to their ability to market new products to big, established customer bases. Maybe DeepSeek-R1’s arrival marks a turning point for these companies. But only a failure of imagination would doom them to irrelevance.

You’ve been reading Plugged In, Fast Company’s weekly tech newsletter from me, global technology editor Harry McCracken. If a friend or colleague forwarded this edition to you—or if you’re reading it on FastCompany.com—you can check out previous issues and sign up to get it yourself every Wednesday morning. I love hearing from you: Ping me at [email protected] with your feedback and ideas for future newsletters. I’m also on Bluesky, Mastodon, and Threads.