Термин «hill-climbing» из мира искусственного интеллекта становится новой модной фразой в бизнес-лексиконе

Silicon Valley loves jargon—and ‘hill-climbing’ is its favorite new phrase A term that describes a computer-science algorithm is the metaphor du jour among AI types—and coming soon to a boardroom near you. SAN FRANCISCO—An industry that loves its buzzwords has a new favorite and, like a rogue artificial-intelligence model, it has broken out of its sandbox and is colonizing new domains. “Hill-climbing” is the term that is suddenly on lips across Silicon Valley, for reasons that have nothing to do with San Francisco’s dramatic topography or techies’ love of hiking. A niche term of art in machine-learning research, it is proving a handy metaphor for talking about growth and threatening to cross over into wider business-speak. In an interview with The Wall Street Journal in June, Microsoft Chief Executive Satya Nadella said companies should use many different AI models, “all hill-climbing inside of a machine you control.” Google DeepMind co-founder Demis Hassabis, Google Cloud CEO Thomas Kurian and Palantir technology chief Shyam Sankar are among other tech executives who have invoked it in recent public remarks. The phrase dates back to the early days of AI and cognitive science, when researchers used it to describe a methodical step-by-step algorithm that yields predictable improvement through trial and error—think of the game “hot and cold.” In the context of AI, it has a specific technical meaning. Like rats in a Skinner box being conditioned to get food or stop pain, AI models hill-climb in a process similar to reinforcement learning, tweaking their parameters based on feedback. But in optimize-everything San Francisco—a city where tech strivers rate romantic partners on spreadsheets and feed their biodata to AI for life coaching—the term has become an irresistible shorthand for talking about everything from professional achievement to parenting. “Hill climbing has become, basically, if you have some metric, how do you get better at it?” said Ashwin Sreenivas, co-founder of Decagon, a company that builds AI customer service agents. Nick Heiner, head of reinforcement learning environments at data training company Surge AI, put it more simply: “It’s just a universal term for the act of improving.” Paul Graham, co-founder of the influential startup incubator YCombinator, employs the term to encourage founders “to do whatever will make you grow at that moment” rather than worry about hitting a “local maximum,” or peak that leaves nowhere higher to climb. Hassabis, who recently relinquished his day-to-day role running Google’s DeepMind lab in favor of a chairman job, mused in a podcast in July about whether the AI industry itself has found itself on just such a peak. Is hill-climbing, he asked, “all that is needed from here, or do we actually need one or two more big breakthroughs?” Tech workers who spend their days training models and prompting agents have found the term infiltrating their daily speech, whether they are talking about work projects—“What are you hill-climbing today?”—or personal pursuits. Charlie O’Neill, co-head of model training at AI inference startup Baseten, thinks of quantitatively minded new hires as “hill-climbers” or “number-go-up people,” he said. Praising a group of trainees on X, he wrote: “These Waterloo interns at baseten are hill climbing machines.” Shira Eisenberg, founder of an AI startup for early childhood, thinks the popularity of the phrase reflects a culture of bristling ambition stoked by the AI boom. “San Francisco is a city of hill-climbers,” she said, adding that she has heard others make the same observation. Hannah Gao, an intern at OpenAI, finds the constant use of “hill-climbing” among co-workers and friends endearing. Still, the underlying ethos rankles somewhat. “Why feel the need to optimize everything? People need to hill-climb some swag,” she said. For AI researcher Alex Dong, hill-climbing is a philosophy. While the real world is complicated and messy, Dong values the “mathematical purity” of his reinforcement-learning work: “If you can create an environment where you can hill-climb, it’s a guaranteed return,” he said. He has found the concept helpful in getting his 12-year-old daughter over her hatred for math homework. When she returns home from school, he asks, “Would you like to climb the geometry hill today?” Dhruv Mahajan had less success using the term for parenting. When his daughter lost a tennis match 0-6, the chief AI scientist at startup Resolve AI tried to comfort her by saying the sport is a “hill-climbing process.” She was baffled. Digital technology has long been fertile ground for new English idioms, with terms such as “reboot,” “bandwidth” and “downtime” entering the popular lexicon. Michael Adams, provost professor of English at Indiana University, said instances of tech jargon that go mainstream tend to be “fairly straightforward terms that can be used metaphorically in tech but are familiar enough that they can then emerge out of tech into everyday speech.” Businesspeople who interact with more technical collaborators often serve as the vectors who bring such jargon into the wider corporate vernacular, he said. Hanlin Tang, the CTO of neural networks at Databricks, said he has spoken with clients not in technology who have dropped the term “hill-climbing” in calls—mainly AI-forward ones in consulting. Adams likened “hill-climbing” to “flywheel,” a popular bit of engineering-bred MBA-speak, in its easy-to-grasp physicality. “You can pick that up and make it into effective jargon fast because it’s related to the original idea. But it’s also not some weirdly intimidating word, right?” Write to Tina Li at [email protected]