A deflating A.I. bubble may be just what the tech industry needs: As funding dries up, companies will have to build models that do more with fewer chips and less power. Gen A.I. needs its own course correction - both for the sake of energy efficiency and for its own advancement. Large language models, for all their wonders, can only predict the next thing a human would say. Train one on texts from the late 1800s and it won’t invent airplanes or rockets. If we only scale up our current approach, wasting money on fast-obsolete chips and energy-guzzling data centers, we won’t progress beyond current technology, which still yields limited results. A course-corrected A.I. would help tackle new challenges - scientific discoveries, enabling medical breakthroughs - rather than refining what we already know.

“Bubbles are great. May the bubbles continue,” Eric Schmidt, Google’s former chief executive, recently said. For artificial intelligence to advance, companies must continue to pour record-breaking investments into A.I. infrastructure — or so the thinking goes. Build more data centers, and A.I. will find a cure for cancer, reach artificial general intelligence and beat China.

But progress usually happens under pressure. When energy gets expensive, people invent energy-saving methods. When there’s a worker shortage, they invent labor-saving machines. A deflating A.I. bubble may be just what the tech industry needs: As funding dries up, companies will have to build models that do more with fewer chips and less power.

Economists have a name for innovation in times of scarcity: directed technical change. In 1977, as Americans stood on long, winding gas lines, President Jimmy Carter likened the energy crisis to a war, and businesses responded accordingly, developing technologies we now take for granted: more-efficient engines, better-insulated homes, a wave of electric and hybrid vehicle technologies and early forms of renewable energy.

Similar circumstances transformed agriculture. In the early 20th century, abundant, low-wage labor dulled the incentive to mechanize. Then, in the spring of 1927 the Mississippi River burst through the levees, turning cotton country into an inland sea. Many residents took refuge in Red Cross camps; in some counties, up to four-fifths of families left. With fewer hands for planting and harvesting, landlords turned to machines: Tractors replaced teams, and mechanical tools spread faster there than in neighboring counties. Generative A.I. needs its own course correction — both for the sake of energy efficiency and for its own advancement. Large language models, for all their wonders, can only predict the next thing a human would say. Train one on texts from the late 1800s and it won’t invent airplanes or rockets. It will channel ideas from that period, when leading scientists thought human flight was impossible. If we only scale up our current approach, wasting money on fast-obsolete chips and energy-guzzling data centers, we won’t progress beyond our current technology, which still yields limited, mediocre results.

Better A.I. would remember what it learns, just as humans do, and squeeze more work from each watt. Tech companies spend billions of dollars running large language models that don’t learn while they run. A tool that does both simultaneously would come closer to approximating the human brain, allowing it to innovate more readily.

The boom and bust pattern has been central to A.I.’s advancement. In the 1980s, the blossoming A.I. industry tried to replicate human reasoning by feeding computer systems thousands of “if-then” rules written by programmers. The approach proved expensive and limited. But the resulting shock pushed researchers toward models that learned from examples and dealt better with uncertainty, while neural networks, then unfashionable, kept getting better.

Progress became easier to measure, and the field stopped betting everything on a single big approach. Jobs were lost, and labs closed, but that slowdown taught scientists and developers better habits — more empirical, flexible and results-focused — that set the stage for modern A.I.

Scarcity is still pushing A.I. forward, as companies with fewer resources learn to do more with less. In 2018, Europe’s data regulations imposed strict rules and heavy fines on how personal data could be collected and stored. In response, tech companies adapted tactics to fine-tune existing models and use artificially generated data instead of real records. And since 2018, DeepSeek, a Chinese company, has worked around the U.S. export constraints. Its models — trained with a small fraction of the computing power of Western rivals yet comparable on many performance benchmarks — show how scarcity breeds ingenuity. When there’s no incentive for energy-efficient innovation, technology risks settling on the wrong track. Around 1900, electric vehicles had promise; New York and London even ran electric taxi fleets. But underinvestment in the electricity grid, coupled with cheap oil, led to a system that favored internal combustion for generations. It isn’t hard to imagine a different century had we priced carbon early and kept building the grid. Without changing course, A.I. could be bound for the same fate — a technology that had immense promise but that is trapped in an outdated paradigm that saps our resources.

Humans are astonishingly energy-efficient. A child can pick up cause and effect, how the physical world behaves and basic social norms with a brain that runs on only about 20 watts of power. Today’s A.I. models burn through mountains of data and electricity to approximate that same performance, yet still misfire the moment they must handle unfamiliar problems. A course-corrected A.I. would help us tackle new challenges — making scientific discoveries, enabling medical breakthroughs — rather than merely refining what we already know.

Bubbles are noisy while they inflate. When they burst, the froth clears and you can see which ideas hold up without subsidy. If the A.I. boom cools, what survives will be the systems that do more with less.