The transformer, introduced by Google nine years ago, powers every major large language model. Yet its dense attention mechanism is becoming a bottleneck as models grow, making them expensive and less able to manage large amounts of information. Four fresh concepts promise to end this limitation, potentially making LLMs faster, more efficient, and maybe even smarter.

Meanwhile, AI professors are rethinking academic research. At a recent Schmidt Sciences AI2050 convening in Mountain View, California, leading scholars gathered to discuss how university‑based AI work is evolving in a world where industry funding and policy pressures shape the field.

These developments highlight a shift from transformer dominance to innovative architectures, while academia grapples with funding, ethics, and the practical realities of AI research.

The conversation continues as Nvidia lands $500 billion in AI infrastructure investment, and figures like Bernie Sanders call for an AI pause—underscoring the broader debate about AI’s future.

The next wave may redefine how we build and regulate intelligent systems.