The Download: the next big thing in LLMs and how AI academic research is shifting
| Source: MIT Technology Review AI
Tags: transformers, LLM architecture, Nvidia, AI infrastructure, BlackRock, Meta, Llama
Startups are racing to replace transformers — now a bottleneck as LLMs scale — with faster architectures, while Nvidia secured $500 billion from institutional investors including BlackRock and Goldman Sachs for AI infrastructure.
Details
MIT Technology Review's daily newsletter surfaces two threads worth tracking: post-transformer architecture research and the institutionalization of AI infrastructure as an asset class.\n\nOn the architecture front, MIT Technology Review is running a longer feature on startups pursuing alternatives to the transformer. The transformer's attention mechanism becomes exponentially expensive as context length grows and struggles to maintain large amounts of information simultaneously — a growing bottleneck as LLMs scale. Four alternative approaches are highlighted in the associated article, though newsletter format limits detail.\n\nThe second major data point is Nvidia securing $500 billion from Wall Street for AI infrastructure, with reported deals involving BlackRock, Goldman Sachs, and four other institutional investors. This reinforces AI compute shifting from a technology investment to a formal infrastructure asset class — something institutional capital can allocate to like real estate or toll roads.\n\nThe newsletter also references Zuckerberg's open-source AI manifesto (coinciding with Meta's latest Llama release) and academic AI researchers navigating tension between university norms and the commercial AI industry's pull, following an AI2050 Schmidt Sciences convening in Mountain View.