J Studios/Getty Images
- Top AI companies are in a race to develop artificial general intelligence.
- The large language models powering popular chatbots, however, are showing their limits.
- Some researchers say world models or other strategies might be the clearer path to AGI.
The race to build artificial general intelligence is colliding with a harsh reality: Large language models might be maxed out.
For years, the world’s top AI tech talent has spent billions of dollars developing LLMs, which underpin the most widely used chatbots.
The ultimate goal of many of the companies behind these AI models, however, is to develop AGI, a still theoretical version of AI that reasons like humans. And there’s growing concern that LLMs may be nearing their plateau, far from a technology capable of evolving into AGI.
AI thinkers who have long held this belief were once written off as cynical. But since the release of OpenAI’s GPT-5, which, despite improvements, didn’t live up to OpenAI’s own hype, the doomers are lining up to say, “I told you so.”
Principal among them is perhaps Gary Marcus, an AI leader and best-selling author. Since GPT-5’s release, he’s taken his criticism to new heights.
“Nobody with intellectual integrity should still believe that pure scaling will get us to AGI,” he wrote in a blog post earlier this month, referring to the costly strategy of amassing data and data centers to reach general intelligence. “Even some of the tech bros are waking up to the reality that ‘AGI in 2027’ was marketing, not reality.”
Here’s why some think LLMs are not all they are cracked up to be, and the alternatives some AI researchers believe are the better path to AGI.
The AI bubble
OpenAI is now the most valuable startup on the planet. It has raised about $60 billion, and a discussed secondary share sale could push the company’s valuation over $500 billion. That would make OpenAI the most valuable private company in the world.
There are good reasons for the excitement. According to the company, ChatGPT has 700 million weekly users, and OpenAI’s products have largely set the pace of the AI race.
There are a couple of problems, however. First, and perhaps foremost for its investors, OpenAI is not profitable and shows few signs of becoming profitable soon. Second, the company’s founding mission is to develop AGI in a way that benefits all of humanity, yet there’s a growing feeling that this world-changing technology, which props up much of the hype around AI, is much further away than many engineers and investors originally thought.
Other companies, too, have been riding this hype wave. Google, Meta, xAI, and Anthropic are all attracting and pouring billions of dollars into scaling their LLMs, which means snapping up talent, buying data, and building vast arrays of data centers.
The mismatch between spending and revenue, and hype and reality, is provoking alarm that the AI industry is a bubble on the verge of bursting. OpenAI CEO Sam Altman himself thinks so.
“Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes. Is AI the most important thing to happen in a very long time? My opinion is also yes,” he told journalists earlier this month.
While other tech leaders, like former Google CEO Eric Schmidt, are less certain, a $1 trillion stock market tech sell-off last week showed the concerns are widespread. The market recovered on Friday after Federal Reserve Chair Jerome Powell said he is considering a rate cut in September.
Now, everyone is eagerly anticipating Wednesday’s earnings report from Nvidia, which makes the chips powering LLMs and is the pick-and-shovel company of the AI rush. If the company’s earnings show signs of slowing and its outlook is more cautious, there will be a whole new round of worry, and the AI doomers will again remind everyone of what they’ve been saying for years: LLMs are not the way.
The problem with LLMs
In June, Apple researchers released a paper called “The Illusion of Thinking.” What they found sounded positively human: Advanced reasoning models give up when faced with more complex tasks.
Their conclusion, however, was that these models rely on pattern recognition rather than logical thinking, and the researchers cautioned against the belief that they could result in AGI. “Claims that scaling current architectures will naturally yield general intelligence appear premature,” the researchers wrote.
The paper was widely mocked online, largely because Apple, despite its size and vast resources, is perceived as far behind in the AI race. For skeptics, however, it was validating.
Andrew Gelman, a professor of statistics and political science at Columbia University, has argued that the level of textual comprehension shown by LLMs falls short of expectations. What LLMs do compared to what humans do is the difference between “jogging and running,” Gelman wrote in a 2023 blog post.
“I can jog and jog and jog, thinking about all sorts of things and not feeling like I’m expending much effort, my legs pretty much move up and down of their own accord … but then if I need to run, that takes concentration,” he wrote.
Geoffrey Hinton, the Nobel Prize winner known to some as the Godfather of AI, disagrees. “By training something to be really good at predicting the next word, you’re actually forcing it to understand,” he told The New Yorker back in 2013.
Another potential problem with LLMs is their tendency to misinterpret the meanings of words, hallucinate, and spread misinformation. This reality is why, for now, most companies adopting AI require a human in the mix.
In a report published earlier this year, a group of academic researchers in Germany specializing in computational linguistics surveyed “in-the-wild” hallucination rates for 11 LLMs across 30 languages. They found that the average hallucination rate across all languages varied between 7% to 12%.
Leading AI companies like OpenAI have, in recent years, operated under the belief that these problems can be mitigated by feeding LLMs more information. The so-called scaling laws, which OpenAI researchers outlined in a 2020 paper, state that “model performance depends most strongly on scale.”
However, recently, researchers have begun to question whether LLMs have hit a wall and are facing diminishing returns as they scale. Yann LeCun, Meta’s chief AI scientist who heads a lab under the company’s superintelligence unit, is largely focused on next-generation AI approaches instead of LLMs.
“Most interesting problems scale extremely badly,” he said at the National University of Singapore in April. “You cannot just assume that more data and more compute means smarter AI.” Apple’s analysis also found that current LLM-based reasoning models are inconsistent due to “fundamental limitations in how models maintain algorithmic consistency across problem scales.”
Alexandr Wang, the head of Meta’s superintelligence division, appears equally uncertain. He said scaling is “the biggest question in the industry” at the Cerebral Valley conference last year.
Even if scaling worked, access to high-quality data is limited.
The hunt for unique data has been so fierce that leading AI companies are pushing boundaries — sometimes at the risk of copyright violations. Meta once considered acquiring publisher Simon & Schuster as a solution. Anthropic collected and scanned millions of pirated books while training Claude, which a district judge ruled in June did not constitute fair use.
Ultimately, some leading AI researchers say language itself is the limiting factor, and that’s why LLMs are not the path to AGI.
“Language doesn’t exist in nature,” Fei Fei Li, the Stanford professor famous for inventing ImageNet, said on an episode of Andreessen Horowitz’s podcast in June. “Humans,” she said, “not only do we survive, live, and work, but we build civilization beyond language.”
LeCun’s gripe is similar.
“We need AI systems that can learn new tasks really quickly. They need to understand the physical world, not just text and language but the real world, have some level of common sense, and abilities to reason and plan, have persistent memory — all the stuff that we expect from intelligent entities,” he said during his talk in April.
New ways to AGI
Researchers like Li and LeCun are pursuing an alternative to LLMs, called world models, that they believe is a better path to AGI.
Unlike large language models, which determine outputs based on statistical relationships between words and phrases, world models make predictions by simulating and learning from the world around them. These kinds of models feel more akin to how humans learn, while LLMs rely on vast troves of data that humans have no access to.
Computer scientist and MIT professor Jay Wright Forrester outlined the value of this kind of model all the way back in a 1971 paper.
“Each of us uses models constantly. Every person in private life and in business instinctively uses models for decision-making. The mental images in one’s head about one’s surroundings are models,” he wrote. “All decisions are taken on the basis of models. All laws are passed on the basis of models. All executive actions are taken on the basis of models.”
Recent research has found that world models not only capture reality as it is, but can also simulate new environments and scenarios.
In a 2018 paper, researchers David Ha and Jurgen Schmidhuber built a simple world model inspired by humans’ cognitive systems. This was used to not only model hypothetical scenarios, but also to train agents.
“Training agents in the real world is even more expensive,” the authors wrote. “So world models that are trained incrementally to simulate reality may prove to be useful for transferring policies back to the real world.”
In August, Google’s DeepMind released Genie 3, a world model that it says “pushes the boundaries of what world models can accomplish.” It can model physical properties of the real world, like volcanic terrain or a dimly lit ocean. This could allow AI to make predictions based on what it learns from these real-world simulations.
There are other ideas in the works, too. Neuroscience models try to mimic the processes of the brain. Multi-agent models operate on the theory that multiple AIs interacting with each other is a better analogy to how humans function in real life. Researchers pursuing multi-agent models believe AGI is more likely to emerge through this kind of social exchange.
Then, there is embodied AI, which adapts world models into physical forms, allowing robots to interpret and train on the world around them. “Robots take in all kinds of forms and shapes,” Li said on the No Priors podcast in June.
The potential of these alternatives, and in particular world models, gives hope to even Marcus, the premier LLM doomer. He refers to world models as cognitive models and urges AI companies to pivot from LLMs and focus on these alternatives.
“In some ways, LLMs far exceed humans, but in other ways, they are still no match for an ant,” Marcus said in a June blog post. “Without robust cognitive models of the world, they should never be fully trusted.”