The godfather of Meta's AI thinks the AI boom is a dead end (www.businessinsider.com)

Yann LeCun, long-time head of Meta’s AI effort, publicly warned that the current industry frenzy around large language models (LLMs) — the text-trained architectures behind ChatGPT, Gemini and Llama — is a dead end for reaching human-level intelligence. Speaking in Brooklyn, he said LLMs are useful and worth investing in, but argued they “suck the air out of the room,” diverting resources from what he sees as the missing ingredients: grounded “world models” built from visual and embodied data rather than just internet text. His comments come as he’s reportedly preparing to leave Meta and possibly launch a startup, making the critique more than academic dissent inside a company that has poured billions into LLM talent and infrastructure. LeCun’s stance matters because it highlights a deep, unresolved scientific split about how to build general intelligence: scale up text-only transformers or pursue multimodal, perceptual models that learn from interaction with environments. Technically, the debate centers on data modality (text vs. vision/embodiment), architecture/learning objectives, and whether statistical pattern-matching in LLMs can yield true reasoning and world grounding. If influential researchers act on LeCun’s view, funding and talent could shift toward alternative paradigms, reshaping research priorities and industrial bets — and underscoring how unsettled AI’s roadmap remains.