The multi-trillion-dollar artificial intelligence (AI) boom was built on certainty that generative models would keep getting exponentially better. Spoiler alert: they are not.
In simple terms, “scaling laws” said that if you threw more data and computing power at an AI model, its capabilities would continuously grow. However, a recent flurry of press reports suggests that is no longer the case, and AI’s leading developers are finding their models are not improving as dramatically as they used to.
OpenAI’s Orion is not that much better at coding than the company’s last flagship model, GPT-4, Bloomberg News said, while Alphabet Inc’s Google is seeing only incremental improvements to its Gemini software. Anthropic, a major rival to both companies, has fallen behind on the release of its long-awaited Claude model.
Illustration: Mountain People
Executives at OpenAI, Anthropic and Google all told me without hesitation in recent months that AI development was not plateauing. However, they would say that. The truth is that long-held fears of diminishing returns for generative AI, predicted even by Bill Gates, are becoming real.
Ilya Sutskever, an AI icon who popularized the bigger-is-better approach to building large language models, recently told Reuters that it had leveled off.
“The 2010s were the age of scaling,” he said. “Now we’re back in the age of wonder and discovery once again.”
“Wonder and discovery” puts quite the positive spin on “we have no idea what to do next.” It could also, understandably, spark anxiety attacks for investors and businesses, who are expected to spend US$1 trillion on the infrastructure needed to deliver on AI’s promise to transform everything.
Wall Street banks, hedge funds and private equity firms are spending billions on funding the buildout of vast data centers, a recent Bloomberg News investigation showed.
Does this all add up to a terrible gamble? Not exactly.
There is no question that the main beneficiaries of the AI boom have been the world’s largest tech companies. Quarterly cloud storage revenue for Microsoft Corp, Google and Amazon Inc has been growing at a steady clip and their market capitalizations, along with those of Nvidia Corp, Apple Inc and Meta Platforms Inc, have soared by US$8 trillion in aggregate over the past two years. Returns on investment for everyone else — their customers — are taking longer to show up.
Yet a break in the market hype around AI could be useful, just as it has been for previous innovations. That is because technology typically does not hit a brick wall and die, but goes through an S-curve. The idea of the S-curve is that initial progress takes years before rapidly accelerating, as we have seen over the last two years with generative AI, before it starts to slow again and, crucially, evolve.
Critics over the years, for instance, regularly declared Moore’s Law dead just before a manufacturing breakthrough for chips pushed it forward again. The development of airplanes progressed at a glacial pace until the transition from propellers to jets in the late 1950s led to a leap forward — before the technology seemed to plateau. Just like chip manufacturing, aviation’s development did not stall, it transformed; passenger planes have become far more fuel efficient, safer and cheaper to operate, even if they are only nominally faster than they were in the 1960s.
A similar plateau for AI and its scaling laws might also mean a new approach to development and measuring success, which until now has focused too much on capability, and not enough on other areas such as safety. Some of the most advanced generative AI models fall short on critical areas such as security and fairness, according to a recent academic study that measured how well they followed Europe’s upcoming AI law.
For much of this year already, AI researchers have been looking at new paths for improving their models that do not just involve throwing more data and computing power at them. One approach is to focus on enhancing a model after it has been trained, in the so-called inference phase. This can involve giving a model extra time to process multiple possibilities before settling on an answer, and it is why OpenAI described its most recent model, o1, as being better at “reasoning.”
The beauty of the S-curve is that it can give everyone else some breathing room, instead of clamoring for the latest tech that would give them an edge over their competitors. Companies that have been experimenting with generative AI and grappling with ways to boost their productivity, now have some time to redesign their workflows and business processes to better capitalize on current AI models, which are already powerful. (Remember, it took years for businesses to reorganize themselves around computers in the 1980s.)
Stanford University professor Erik Brynjolfsson’s writing on the “productivity paradox” said that output often appears to stall or drop when major new technologies arrive, before surging. A pause for AI gives businesses more space in that all-important investment phase.
It also gives regulators time to design more effective guardrails. The EU’s AI Act, which companies would be subject to from 2026, needs to be more specific in how it defines harms. As standards bodies do that work, it helps that newfangled models leading to a batch of unexpected problems are not about to flood the market.
Generative AI has been on a bullet train during the past two years, and the momentum has clearly been lucrative for tech giants. A slowdown at the station offers a much-needed break for everyone else.
Parmy Olson is a Bloomberg Opinion columnist covering technology. A former reporter for the Wall Street Journal and Forbes, she is author of Supremacy: AI, ChatGPT and the Race That Will Change the World.
Taiwan’s higher education system is facing an existential crisis. As the demographic drop-off continues to empty classrooms, universities across the island are locked in a desperate battle for survival, international student recruitment and crucial Ministry of Education funding. To win this battle, institutions have turned to what seems like an objective measure of quality: global university rankings. Unfortunately, this chase is a costly illusion, and taxpayers are footing the bill. In the past few years, the goalposts have shifted from pure research output to “sustainability” and “societal impact,” largely driven by commercial metrics such as the UK-based Times Higher Education (THE) Impact
History might remember 2026, not 2022, as the year artificial intelligence (AI) truly changed everything. ChatGPT’s launch was a product moment. What is happening now is an anthropological moment: AI is no longer merely answering questions. It is now taking initiative and learning from others to get things done, behaving less like software and more like a colleague. The economic consequence is the rise of the one-person company — a structure anticipated in the 2024 book The Choices Amid Great Changes, which I coauthored. The real target of AI is not labor. It is hierarchy. When AI sharply reduces the cost
The inter-Korean relationship, long defined by national division, offers the clearest mirror within East Asia for cross-strait relations. Yet even there, reunification language is breaking down. The South Korean government disclosed on Wednesday last week that North Korea’s constitutional revision in March had deleted references to reunification and added a territorial clause defining its border with South Korea. South Korea is also seriously debating whether national reunification with North Korea is still necessary. On April 27, South Korean President Lee Jae-myung marked the eighth anniversary of the Panmunjom Declaration, the 2018 inter-Korean agreement in which the two Koreas pledged to
I wrote this before US President Donald Trump embarked on his uneventful state visit to China on Thursday. So, I shall confine my observations to the joint US-Philippine military exercise of April 20 through May 8, known collectively as “Balikatan 2026.” This year’s Balikatan was notable for its “firsts.” First, it was conducted primarily with Taiwan in mind, not the Philippines or even the South China Sea. It also showed that in the Pacific, America’s alliance network is still robust. Allies are enthusiastic about America’s renewed leadership in the region. Nine decades ago, in 1936, America had neither military strength
RSS