In a striking act of self-criticism, one of the architects of the transformer technology that powers ChatGPT, Claudiusand virtually every major AI system told an audience of industry leaders this week that AI research has become dangerously narrow and that it is moving away from its own work.
Lion Joneswho co-authored the groundbreaking 2017 paper “Attention is all you need” and even coined the name “transformer”, he gave an extremely honest assessment of the topic TED AI conference on Tuesday in San Francisco: Despite an unprecedented investment and talent flowing into AI, the field has coalesced around a single architectural approach, potentially blinding researchers to the next major breakthrough.
“Despite the fact that there has never been so much interest, resources, money and talent, it has somehow narrowed the research we do,” Jones told the audience. He argued that the culprit is “tremendous pressure” from investors demanding profits and researchers trying to stand out in an overcrowded market.
The warning carries particular weight given Jones’ role in the history of artificial intelligence. The transformer architecture helped grow at Google, it became the basis for the generative artificial intelligence boom, enabling systems that can write essays, generate images and engage in human conversation. His article was cited over 100,000 timesmaking it one of the most influential computer science publications of the century.
Currently as CTO and co-founder of a company based in Tokyo SamanaJones is clearly abandoning his own work. “I personally made a decision earlier this year that I would drastically reduce the time I spend on transformers,” he said. “Now I’m researching hard and looking for the next big thing.”
More funding for artificial intelligence has led to less creative research, according to a transformer pioneer
Jones painted a picture of the artificial intelligence research community suffering from what he called a paradox: More resources led to less creativity. He described researchers constantly checking to make sure they weren’t “caught” by competitors working on identical ideas, and scientists choosing safe, publishable projects over risky and potentially transformative ones.
“If you’re doing standard AI research today, you kind of have to assume that there are maybe three or four other groups doing something very similar, or maybe the exact same thing,” Jones said, describing an environment in which “unfortunately, this pressure is hurting the science because people are rushing to publish, which reduces the level of creativity.”
He drew an analogy from artificial intelligence itself – the “exploration versus exploitation” trade-off that governs how algorithms search for solutions. When a system uses too much and explores too little, it finds mediocre local solutions while ignoring better alternatives. “We’re almost certainly in that situation right now in the AI industry,” Jones argued.
The consequences are sobering. Jones recalled the period just before the advent of transformers, when researchers were constantly tweaking recurrent neural networks – the previously dominant architecture – to achieve incremental benefits. When transformers came along, all this work suddenly seemed irrelevant. “How much time do you think these researchers would have spent trying to improve the recurrent neural network if they had known that something like transformers was just around the corner?” he asked.
He worries that the industry is repeating this pattern. “I’m afraid we’re in a situation now where we focus on one architecture and we just change it and try different things, and the breakthrough may be just around the corner.”
How the newspaper “Everything You Need” was born out of freedom, not pressure
To emphasize his point, Jones described the conditions that allowed the emergence of transformers – a stark contrast to today’s environment. The project, he said, was “very organic, grassroots” and born out of “lunch conversations or random writing on the whiteboard in the office.”
Most importantly, “we didn’t actually have a good idea, we had the freedom to spend our time, work on it, and, more importantly, we didn’t feel any pressure from management,” Jones recalls. “There is no pressure to work on a specific project or publish a few articles to raise a certain indicator.”
This freedom, Jones suggested, is largely absent today. Even researchers recruited for astronomical salaries — in some cases literally a million dollars a year — may not feel comfortable taking risks. “When they start in a new role, do you think they will feel empowered to try out their wilder and more speculative ideas, or do they feel a lot of pressure to prove their worth and go after the low-hanging fruit again?” he asked.
Why one artificial intelligence lab is betting that research freedom outweighs multimillion-dollar salaries
Jones’s solution is deliberately provocative: turn up the “exploration dial” and openly share your discoveries, even at a competitive price. He recognized the irony of his position. “It may sound a little controversial for one of the Transformers writers to stand on stage and say he’s absolutely unwell of them, but it’s kind of fair, isn’t it? I’ve been working on them longer than anyone else, except maybe seven people.”
On SamanaJones said he is trying to recreate the pre-transition environment, using nature-inspired research and minimal pressure to chase publications or compete directly with rivals. He offered researchers the mantra of engineer Brian Cheung: “You should only do research that wouldn’t happen if you didn’t do it.”
One example is “Sakana”a constant thinking machine“, which involves brain synchronization with neural networks. The employee who pitched the idea told Jones that he would be met with skepticism and pressure not to waste time at previous employers or in academic positions. At Sakan, Jones gave him a week to investigate the situation. The project was successful enough to attract attention in NeuroIPSimmense artificial intelligence conference.
Jones even suggested that freedom is more essential than pay in recruiting. “It’s a really good way to acquire talent,” he said of the discovery environment. “Think about it, talented, intelligent and ambitious people will naturally seek this type of environment.”
Transformer’s success could block the next AI breakthrough
Perhaps most provocatively, Jones suggested that transformers may become victims of their own success. “The fact that current technology is so efficient and flexible … has stopped us from looking for better ones,” he said. “It makes sense that if current technology were inferior, more people would seek better ones.”
He tried to make clear that he was not disregarding ongoing transformer research. “There is still a lot of very important work to be done on current technology that will add a lot of value in the coming years,” he said. “What I’m saying is that given the amount of talent and resources we have now, we can afford to do a lot more.”
His ultimate message was one of cooperation rather than competition. “Really, from my point of view, it’s not a competition,” Jones concluded. “We all have the same goal. We all want to see technology advance so that we can all benefit from it. So if we all turn up the exploration dial together and then openly share what we discover, we can achieve our goal much faster.”
The high stakes of the AI mining problem
The remarks come at a crucial time for artificial intelligence. The industry is grappling with growing evidence that larger transformer models are simply being built they may be approaching diminishing returns. Leading researchers have begun to openly discuss whether the current paradigm has fundamental limitations, with some suggesting that architectural innovations – not just scale – will be needed to continue progress toward more capable AI systems.
Jones’ warning suggests that finding these innovations may require dismantling the incentive structures that have fueled the recent artificial intelligence boom. WITH tens of billions of dollars annually affect the development of artificial intelligence and fierce competition between laboratories ensuring secrecy and rapid publication cycles, the exploratory research environment he describes seems increasingly distant.
However, his internal perspective is of extraordinary importance. As someone who helped create the technology that dominates the field today, Jones understands both what it takes to achieve breakthrough innovation and the risks the industry risks by abandoning this approach. His decision to move away from transformers – the architecture that gave him his reputation – adds credibility to a message that might otherwise sound like adversarial positioning.
It remains to be seen whether top AI players will heed the call. Jones, however, provided a stark reminder of what’s at stake: The next transformer-scale breakthrough could be just around the corner, and will be pursued by researchers with the freedom to explore. Or it could remain unexplored while thousands of researchers race to publish more architectural improvements that, in Jones’ words, one of its creators is “absolutely fed up with.”
After all, he’s been working on transformers longer than almost anyone else. He would know when it was time to move on.