The breakthrough that AI needs | Mint

Large language models have a huge appetite for electricity. The energy used to train OpenAI’s gpt-4 model could have powered 50 American homes for a century. And as models get larger, costs quickly rise. By one estimate, training today’s largest models costs $100 million; the next generation could cost $1 billion, and the next, $10 billion. On top of this, asking a model to answer a query has a computational cost: between $2,400 and $223,000 to summarize the financial reports of the world’s 58,000 public companies. Over time, those “inference” costs, when added up, may exceed the cost of training. If they did, it’s hard to see how generative AI could ever become economically viable.

This is spooking investors, many of whom have bet big on AI. They have flocked to Nvidia, which designs the chips most commonly used for AI models. Its market cap has risen by $2.5 trillion over the past two years. Venture capitalists and others have invested nearly $95 billion in AI startups since the start of 2023. OpenAI, the maker of ChatGPT, is reportedly seeking a $150 billion valuation, which would make it one of the largest private tech companies in the world.

There is no need to be alarmed. Many other technologies have faced limits and have thrived on human ingenuity. The difficulty of getting people into space led to innovations that are now used on Earth as well. The oil price crisis of the 1970s spurred energy efficiency and, in some countries, alternative means of generation, including nuclear. Three decades later, fracking made it possible to reach oil and gas reserves that were previously unprofitable to extract. As a result, the United States now produces more oil than any other country.

Advances in AI are already demonstrating how constraints can spur creativity. As our quarterly technology report this week notes, companies are developing chips specifically for the operations needed to run large language models. This specialization means they can run more efficiently than more general-purpose processors like those from Nvidia. Alphabet, Amazon, Apple, Meta, and Microsoft are all designing their own AI chips. More money has gone into funding AI chip startups in the first half of this year than in the past three semesters combined.

Developers are making changes to AI software, too. Larger models that rely on brute force computing power are giving way to smaller, more specialized systems. OpenAI’s newest model, o1, is designed to reason better, but not to generate text. Other manufacturers are employing less burdensome calculations to make more efficient use of chips. Through clever approaches, such as using a mix of models, each suited to a different type of problem, researchers have dramatically reduced processing time. All of this will change how the industry works.

Investors and governments have grown accustomed to the idea that among tech companies, the first-place company has a natural advantage. In the case of AI, that assumption can no longer be taken for granted. Today, Nvidia sells four-fifths of the world’s AI chips, but more specialized rivals could easily steal its share. Google’s AI processors are already the third most widely used in data centers around the world.

OpenAI may have launched the pioneering large language model, but as resource constraints have set in, other big model makers, such as Anthropic, Google, and Meta, are catching up. While there is still a gap between them and second-tier models, such as France’s Mistral, it may be closed. If the trend toward smaller, more specialized models continues, the AI ​​universe could contain a constellation of models, rather than just a few superstars.

That means investors face a tough road ahead. Their bets on the current leaders look less certain. Nvidia could lose ground to other chipmakers; OpenAI could be supplanted. Big tech companies are snapping up talent, and many of them make the devices through which, they hope, consumers will communicate with their AI assistants. But competition among them is fierce. Few companies yet have a strategy for profiting from generative AI. Even if the industry ends up belonging to a single winner, it’s not clear who that will be.

Governments will also have to change their thinking. Their penchant for industrial policy is all about handouts, but progress in AI is as much about having the right talent and a thriving ecosystem as it is about accumulating capital and computing power. Countries in Europe and the Middle East may find that the hard work of cultivating ingenuity matters as much as buying computer chips. The US, by contrast, is blessed with chips, talent and entrepreneurial spirit. It has many of the world’s best universities and, in San Francisco and Silicon Valley, an enviable and established talent pool.

Husking

But the US attempt to rein in China is failing. Hoping to prevent a strategic rival from taking the lead in a crucial technology, it has sought to restrict China’s access to cutting-edge chips. In doing so, it has unwittingly spurred the growth of a research system in China that excels at circumventing limitations.

When ingenuity counts more than brute force, a better way to ensure America’s leadership would be to attract and retain top researchers from elsewhere — for example, through more flexible visa rules. The AI ​​era is still in its early days and there are many uncertainties, but the breakthroughs AI needs will come from giving ideas and talent the space to flourish at home, not from trying to take down rivals abroad.

© 2024, The Economist Newspaper Ltd. All rights reserved.

From The Economist, published under license. The original content can be found at www.economist.com