A new wave of startups is developing next-generation large language models (LLMs) to address limitations in current transformer-based AI systems, according to MIT Technology Review. These companies aim to build what the publication calls LLMs+, a future generation of models that improve on existing architectures. The shift comes nearly a decade after the transformer neural network was introduced in 2017 and became the foundation for all major LLMs today.
The transformer architecture, introduced by Google researchers in their 2017 paper “Attention Is All You Need,” revolutionized AI by enabling efficient processing of long data sequences, especially text. However, recent advances in LLMs have relied on workarounds to overcome fundamental flaws in transformers. Startups like Subquadratic, led by CEO Justin Dangel, are exploring new approaches to push beyond these constraints and create more capable AI models, as detailed in MIT Technology Review’s What’s Next series.
This development matters as the AI industry, heavily dependent on transformers, faces challenges in scaling and reasoning capabilities. The emergence of LLMs+ signals a potential shift in how AI models are built, with implications for applications across industries. The ongoing innovation could redefine the competitive landscape, as companies seek to improve performance and handle larger inputs more effectively than current transformer-based models.
MIT Technology Review’s coverage highlights that while LLMs will remain central to AI, their underlying technology is evolving. The publication’s 2026 list of 10 things that matter in AI includes LLMs+ as a key trend shaping the future of artificial intelligence research and development.