Skip to content
The Chat Archive
sonar-pro-high LMArena Search Arena 24K

explain why yann le cunn says llms are on their way out. give specific quotes

  1. User

    explain why yann le cunn says llms are on their way out. give specific quotes

  2. Assistant

    Yann LeCun, a prominent AI researcher and Chief AI Scientist at Meta, has expressed skepticism about the long-term viability of large language models (LLMs) like GPT, Llama, and Claude. He believes they are nearing obsolescence due to fundamental limitations in their design and capabilities. Below are the key reasons and specific quotes from LeCun explaining his perspective:

    1. Lack of Reasoning and Planning

    LeCun criticizes LLMs for their inability to perform reasoning and planning, which he considers essential for advanced AI systems. He describes LLMs as "reactive" systems that generate text token by token without engaging in deeper reasoning. He stated, “An LLM generates one token at a time, processing a fixed amount of computation to produce each token. This is clearly a reactive process—there's no reasoning involved” [1][3].

    He contrasts this with the need for AI systems that can model the physical world, reason across different timescales, and plan actions hierarchically. “We need machines that understand the physical world. We need machines that are capable of reasoning and planning,” he explained [7].

    2. Overemphasis on Language

    LeCun argues that LLMs are limited because they focus solely on language, which he sees as a narrow and simplified representation of reality. He remarked, “Language is relatively straightforward due to its strong statistical properties. It exists in a low-dimensional, discrete space, essentially a serialized format of our thoughts” [1]. He also noted, “The goal is for AI systems to focus on the world being talked about, not the words themselves — but LLMs don’t grasp the distinction” [6].

    3. Inability to Handle Novel Problems

    LeCun highlights that LLMs struggle with tasks they have not been explicitly trained on. He stated, “If they see a new problem from scratch, current systems just cannot find a solution” [7]. This inability to generalize to new situations underscores their limitations compared to human intelligence.

    4. Inefficiency and Resource Intensity

    LeCun criticizes the computational inefficiency of LLMs, particularly their reliance on statistical tricks to simulate reasoning. He described this as a “horrible trick” and added, “It’s not the way humans reason” [7]. He also pointed out the high costs associated with fine-tuning LLMs for specific tasks, which he believes is unsustainable [1].

    5. Limited Understanding of Reality

    LeCun emphasizes the importance of AI systems having a "world model" to understand and predict physical reality. He explained, “Current AI systems have a very hard time solving new problems that they’ve never faced before. They don’t have this mental model, this world model… They don’t reason in that way” [7].

    6. Vision for the Future

    LeCun envisions a new paradigm for AI that moves beyond LLMs. He and his team are working on systems that integrate abstract representations of reality, enabling reasoning and planning. He stated, “If we can achieve our goals within three to five years, we will establish a vastly improved framework for systems capable of reasoning and planning” [1]. He believes these systems will surpass LLMs and play a more transformative role in AI development.

    Conclusion

    LeCun’s critique of LLMs centers on their lack of reasoning, inefficiency, and narrow focus on language. He predicts that LLMs will be replaced by more advanced systems within five years, stating, “LLMs may become obsolete within five years, supplanted by new models… which possess an understanding of the physical world and exhibit System 2 capabilities” [3]. Despite his criticisms, he acknowledges the utility of LLMs in specific applications but sees them as a stepping stone rather than the future of AI.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.