Why AI Leaders Urge Students to Move Beyond Large Language Models

Stay Away from LLMs

During the VivaTech conference in Paris, Meta AI CEO Yann LeCun—often called the "Lu Xun of AI"—advised students interested in building the next generation of AI systems to avoid work on large language models (LLMs), saying that such research belongs to big companies and offers little room for student contribution.

LeCun emphasized the need to develop AI systems that overcome the limitations of current LLMs.

Exploring Alternatives

Devika founder Mufeed VH echoed this sentiment, suggesting that researchers should move away from Transformer‑based models and explore new architectures such as RMKV, an RNN‑style design that promises an unlimited context window and advanced reasoning capabilities, potentially rivaling GPT‑4.

But LLMs Keep Evolving

Despite LeCun’s caution, LLM research continues to advance. AI/ML consultant Dan Hou discussed GPT‑4o, a model designed to natively understand video and audio, expanding the amount of data future models can train on. Hou speculated that multimodal architectures could make AI dramatically smarter.

Sam Altman recently noted that data scarcity will no longer be a bottleneck, implying that the scaling laws governing LLM growth will persist if data issues are resolved.