Did the LLM Race Send AI Down the Wrong Path?

Large language models have become almost synonymous with artificial intelligence.

ChatGPT changed how millions of people interact with software. Models learned to write code, analyse documents, generate images, operate tools and increasingly perform tasks that once seemed comfortably beyond machines.

But what if the industry's biggest success also became a distraction?

That is the uncomfortable argument made by AI researcher François Chollet, creator of Keras and co-founder of ARC Prize.

In a 2024 interview with Dwarkesh Patel that has recently resurfaced in AI coverage, Chollet argued that OpenAI had set progress toward artificial general intelligence back by "probably like 5–10 years."

It is a provocative claim. It is also important to understand what Chollet was actually arguing.

The five-to-ten-year figure is not a measured delay, nor is it a scientific consensus. It is Chollet's assessment of how the direction of AI research changed as the industry concentrated around large language models.

The problem isn't that LLMs are useless

Chollet's argument shouldn't be interpreted as saying language models don't work.

Clearly, they do.

Modern LLMs can write software, summarize enormous documents, translate languages, generate creative material and act as the reasoning layer behind increasingly capable AI agents.

His criticism is more fundamental.

Chollet distinguishes between skill and intelligence.

A model can become extraordinarily capable at tasks represented within its enormous training distribution. But genuine intelligence, in his view, requires something more difficult: the ability to encounter a genuinely new problem and efficiently figure out how to solve it.

That distinction sits at the centre of ARC-AGI, the benchmark Chollet originally developed to test generalisation and adaptation rather than simply accumulated knowledge.

ARC problems deliberately present unfamiliar visual reasoning tasks. A system has to infer the underlying transformation from a small number of examples and apply it to a new case.

For humans, many of these puzzles appear remarkably simple.

For AI systems, they have historically exposed weaknesses that conventional benchmarks can hide.

When everyone started building the same thing

Chollet's criticism extends beyond model architecture.

He argues that the enormous commercial success of LLMs changed where money, talent and attention flowed.

Once scaling language models began producing increasingly impressive results, the incentives became obvious.

Build larger models.

Acquire more compute.

Collect more data.

Improve post-training.

Scale inference.

Build products around the resulting models.

The approach has produced extraordinary technology.

But Chollet's concern is what wasn't explored while the industry raced in the same direction.

Before the current LLM boom, AI research involved a wider range of competing ideas. Today, an enormous percentage of the industry's resources are concentrated around transformer-based foundation models and techniques built around them.

If AGI ultimately requires a substantially different architecture, concentrating so much research effort on one family of approaches could become a very expensive detour.

The second problem: AI research became less open

There is another part of Chollet's argument that is easy to overlook.

He believes frontier AI research becoming increasingly closed has also slowed progress.

Many of the foundations behind today's AI boom came from openly published research. Researchers could inspect an idea, reproduce it, challenge it and build something better.

As competition between major AI laboratories intensified, that culture changed.

Frontier systems increasingly arrived with limited information about their architecture, datasets and training methods.

The commercial logic is understandable. Training state-of-the-art models can cost enormous amounts of money, and companies have strong incentives to protect the techniques that give them an advantage.

But there is a trade-off.

When fewer researchers can examine the frontier, fewer people can independently build upon it.

Chollet's argument is essentially that AI may have gained enormous investment while simultaneously narrowing the space in which fundamental experimentation happens.

But the LLM story isn't finished

There is a strong counterargument.

Modern AI systems are no longer simply larger versions of early language models.

The industry has added reinforcement learning, tool use, multimodality, external memory, retrieval, agents and increasingly sophisticated reasoning techniques.

Test-time compute allows systems to spend additional resources solving difficult problems rather than immediately predicting an answer.

Researchers are also combining neural models with search, verification and other techniques.

The eventual path to more general intelligence therefore may not be LLMs versus something else.

LLMs could become one component of a larger architecture.

And the representations learned by enormous foundation models may prove extremely useful even if scaling alone doesn't produce AGI.

The bigger question

This debate matters because AI is reaching a point where simply asking which model has the highest benchmark score isn't enough.

The deeper question is:

What are these systems actually learning to do?

A model that has absorbed an enormous portion of humanity's digital knowledge can appear extraordinarily intelligent.

But knowledge and adaptability aren't necessarily the same thing.

If a machine encounters a problem that is genuinely outside its experience, can it construct a new solution efficiently?

Humans do this constantly.

For AI, it remains one of the most important challenges.

TCF Journal Take

The most interesting part of Chollet's argument isn't whether the industry literally lost five or ten years.

That number is impossible to verify.

The important part is the warning behind it.

Technological progress can become strangely conservative.

Once one approach starts working, investment pours into making that approach bigger, faster and cheaper. Success attracts resources, and resources create even more success.

Eventually, the dominant idea can become difficult to question.

LLMs have already changed computing.

They may also become a crucial part of whatever eventually deserves to be called AGI.

But if the next major breakthrough requires a fundamentally different idea, the AI industry needs enough researchers willing — and funded — to look somewhere else.

Sometimes the biggest risk in innovation isn't pursuing the wrong idea.

It's becoming so successful with one idea that we stop searching for another.