Raghuram Rajan, former governor of the Reserve Bank of India and professor of finance at the University of Chicago Booth School of Business, explains in an article for Project Syndicate how to understand the fears and risks associated with slowing AI, and why models can solve the most complex mathematical problems but are not always able to handle everyday requests.

More computing power, less bang for your buck

A professor of mathematics I know recently decided to take a year off to work on AI safety. Given that OpenAI has published 722 papers claiming to solve 377 previously unsolved mathematical problems, and considering that AI models keep getting better, he has started to worry about the future of his profession. If even such a brilliant mathematician can so easily be put out of work, why aren’t we all worried?

This is partly because, although AI can solve one of the Navier–Stokes equations, my bank has yet to create a reliable enough chatbot to resolve a money transfer issue without my having to angrily demand to speak to a human agent. AI labs are using recursive self-improvement techniques to give their still-unreleased models even more incredible scientific capabilities, while pioneers in some industries are already using bots to negotiate contracts and reconcile invoices with deliveries. But AI adoption across traditional businesses is generally moving slowly, and the tasks it could take on remain modest.

So why are so many AI firms pouring so much computing power into developing increasingly complex models if only a few will be able to make full use of their capabilities? It’s not simply the thrill of creating a new, faster form of intelligence. The belief is that the first lab to create a fully self-improving model will be able to use it to improve its capabilities exponentially and leave its competitors behind forever.

But it’s unlikely that competitors won’t catch up. A leader may gain an advantage for a month or two (even with the sought-after exponential growth in capabilities), but fast followers may well find methods that let them leap ahead. That’s what’s happening now in the AI capabilities race: the model at the top keeps changing. It seems incredible that one winner could dominate forever (especially when companies can poach each other’s employees), unless that winner literally paralyzes or destroys its competitors. Let’s hope no one has such intentions.

There may be another motive: none of the messianic leaders of the top AI labs is simply willing to entrust someone else with the power that comes with AI’s extraordinary capabilities. Each is motivated not only by the desire to win, but also by the sense that only they can wield this power responsibly.

Another reason is the fear that China will catch up and overtake the US. Chinese AI models consistently trail the leading American models by only a couple of months. As with the capabilities race within the US, it’s unlikely that the Chinese will suddenly fall further behind unless America mounts a more effective campaign to block China’s access to the most advanced chips, both within China and around the world.

Whatever the reason, the vast computing resources being poured into training models are putting pressure on AI labs. They’re being asked to deliver more returns—and sooner rather than later. That’s why Anthropic created a joint venture, Ode, to help non-tech companies adopt AI with the support of Anthropic engineers who can be assigned to them. This means that some banks may soon have much more sophisticated chatbots, and all the others will have to follow suit. How long before customer support jobs join the ranks of math teachers?

Social carnage or stagnation

There are many ways to adopt AI. You can simply use it to replace jobs, or you can reshape work and skills training so that existing employees’ capabilities are dramatically expanded. The first option will bring in revenue sooner, but the result will be social carnage; the second will take more time and forethought, but will preserve jobs and stability.

The businesses targeted by AI expansion projects are well positioned to adopt AI in ways that put employees first. But these companies need to take more responsibility for their employees’ futures, even as governments provide new incentives to resist the temptation to boost short-term profits through layoffs.

There are many practical ideas, including taxes on the use of corporate tokens and tax breaks for employee training, but only a few countries have implemented them directly. The conclusion seems obvious: society isn’t ready for AI adoption at scale.

Still, we’re now seeing an effort to slow the rapid expansion of computing capacity. Many US political leaders are calling for a pause on permits for new data centers, citing local residents’ concerns and drawing attention to the broader consequences of unchecked AI adoption. A potential reduction in the computing capacity available to monetize AI in the medium term will force investors to limit the allocation of expensive resources to model development.

Moreover, slowing the pace of construction will raise the cost of using AI, affecting decisions about replacing employees. This will give the supply of chips and computer memory time to catch up with demand, helping avoid a relentless rise in costs in the future. It will also help regulators catch up with AI, so that AI lab founders aren’t the only ones deciding how much existential risk (or risk to the financial system) the world must endure. The goal of a pause should not be to stop AI development or use, but to make time for a more considered adoption of the technology.

As for the race with China, it remains manageable. US labs won’t halt development altogether, and judging by what we’ve seen so far, the risk that anyone could gain a lasting advantage is small. Moreover, the Chinese are also concerned about AI risks, and they may see reason to slow their own development by taking a cue from America. A collective slowdown in AI development could even prompt a long-overdue discussion between China and the US about how to use AI for the benefit of humanity, and eliminate the risk that, willingly or unwillingly, we could become its slaves.

#HighTechNews