What you’ll learn:
AI systems built or run with limited resources may soon be able to match the performance of leading large models at significantly lower costs, according to new research. This could revolutionize how companies build and use AI by improving access and lowering barriers to innovation, but it also makes oversight and safety more difficult.
Over the past few years, major technology companies have poured billions of dollars into a simple concept. As artificial intelligence systems have access to more data and computing power, their performance tends to improve.
But new research from MIT FutureTech suggests that the “bigger is better” approach to AI development may have reached a stage of diminishing returns. Previous research has shown that as an AI model grows, the performance gains from additional computing power begin to fade, based on the laws of neural scaling. MIT researchers Hans Gundlach, Jason Lynch, and We find that the reduction in performance gains is so large that companies end up with little comparative advantage even if they scale their models much faster than other organizations.
Their findings suggest a future where AI systems built or run with limited resources (what they call “soft models”) can perform as well as today’s leading models, while significantly reducing costs. And access to advanced systems could transform the way companies compete to build and use AI, as it becomes less about who can afford it and more about who can make the best strategic decisions.
limits of scale
The impact of this new research was already visible earlier this year, when Chinese startup DeepSeek’s R1 model, trained at a cost of around $6 million, rivaled the performance of much larger systems and attracted global attention. Competitor OpenAI reportedly spent hundreds of millions of dollars training its GPT-4 model.
In the researchers’ simulations, a model that increased its computing budget by about 3.6 times each year initially outperformed a smaller model with a $1,000 budget, but the difference peaked after about five years and then steadily narrowed.
“If you look at the laws that predict how these models will perform, the returns are significantly diminishing,” said Lynch, a researcher at the MIT FutureTech Lab. “You have to put in more compute to get an advantage.”
What this means for your business strategy
Lynch said the competitive advantage companies gain from building ever-larger models won’t last long. “Scale can be thought of as a short-term strategy,” they said. “The question is what can we build on top of that advantage before it fades.”
Lynch noted that companies that fine-tune models for specific applications and apply their own high-quality data are in a better position than those that are building ever-larger systems.
They added that as powerful AI becomes more available, competition will also increase. “A lot of people are going to have access to a really powerful model,” Lynch said. “This lowers the barrier to innovation, but also makes monitoring and safety much more difficult.”
The researchers compared this trend to the proliferation of personal computers, suggesting that access to advanced AI could one day become as common as owning a computer itself.
“Given that computers are also becoming cheaper and more widely used, it seems quite likely that a large portion of the world’s population will have access to powerful deep learning models,” the researchers wrote.
Broader access to advanced AI can lead to productivity gains as more people and organizations can share in its benefits.
Top technical experts and team
Face-to-face at MIT Sloan
get information
Monitoring lags as access grows
However, expanding access to AI tools also creates new challenges for surveillance. Current policies, such as U.S. export controls on Nvidia’s GPUs, are often aimed at limiting who can build large-scale “frontier AI” systems by restricting access to computing power. But researchers have found that such measures may no longer be enough if more efficient models start to behave nearly as well as today’s largest models.
“If you’re worried that the models are unsafe or harmful, this suggests that in the not-too-distant future, many people will have access to really powerful models,” Lynch said. “You can’t expect the largest models to be significantly more powerful over long periods of time.”
The researchers describe a brief “governance window” – a period during which large organizations still have an advantage – as an opportunity for regulators to develop stronger safety standards before advanced AI becomes widespread.
They argue that future surveillance must move beyond the limits of computing and focus on the data, algorithms, and safeguards that shape how AI systems are built and used.
Read the paper: “The meek model will inherit the earth”
The MIT FutureTech Lab is an interdisciplinary group that studies the economic and technological foundations of advances in computing.
Hans Gundlach I am a research assistant at MIT FutureTech. He has a background in research in AI, AI safety, and quantum computing, and wants to use these technologies to improve the future.
jason lynch He is a research scientist at MIT FutureTech, working to predict future advances in algorithm development and understand the fundamental limits to improvements in computing performance.
neil thompson He is a principal investigator at the MIT Computer Science and Artificial Intelligence Laboratory and the MIT Sloan Initiative on the Digital Economy, and director of MIT FutureTech. Researches technological innovation and corporate strategy.
