The app for independent voices

**What Marcus gets wrong:**

The framing of LeCun doing a "180" and "changing teams" is historically inaccurate. LeCun has been advocating for energy-based models and non-autoregressive architectures for *over twenty years*. His JEPA work, his critiques of LLMs, his emphasis on world models—none of this is new. LeCun didn't "convert" to anything; he joined a company operationalizing his own long-held vision.

More substantively, Marcus is conflating "energy-based reasoning" with "neurosymbolic AI," and they're not the same thing. Classical neurosymbolic approaches combine neural networks with explicit symbolic logic systems—rules, knowledge graphs, formal inference engines. Energy-based models are still fundamentally continuous and neural; they just optimize globally rather than predicting sequentially. The theoretical lineage traces to statistical physics and LeCun's own decades of work, not to Marcus's 2020 paper.

The claim that Logical Intelligence's mission "reminds him intensely" of his blueprint is... generous self-assessment. It's like saying anyone who criticizes pure behaviorism is vindicating Freud.

**What Marcus gets partially right:**

There *is* a general movement away from "scale the transformer until AGI" thinking. Bodnia's ecosystem framing, LeCun's persistent critiques, the rise of diffusion models—something is shifting. And Marcus has been consistently critical of LLM maximalism when it was unfashionable to be so.

But "the field is moving away from pure LLMs" isn't the same as "the field is moving toward Gary Marcus's specific proposals." That's the sleight of hand.

**The deeper issue:**

Marcus's frame is still "which approach wins"—he's just claiming his team is now winning. He's not engaging with what's actually interesting here: the convergence on entropy-based computation as a principle, the ecosystem framing that moves beyond single architectures, the question of what intelligence *is* rather than which technique achieves it.

He's scoring points in a game that the more interesting thinkers are starting to transcend.

Breaking: Yann LeCun, longtime critic of neurosymbolic approaches, changes teams
Jan 24
at
10:20 PM
Relevant people

Log in or sign up

Join the most interesting and insightful discussions.