AI critic Gary Marcus: Meta's LeCun is finally coming around to the things I said years ago

Abstraction: Gary Marcus argues deep learning alone cannot achieve general intelligence

Key points:

• Marcus and LeCun now agree on three key points: scaling alone is insufficient, reinforcement learning is inadequate on its own, and large language models fall short of general intelligence.
• Marcus accuses LeCun of dismissing his 2018 "Deep Learning: A Critical Appraisal" without reading it, and of ignoring Judea Pearl's causality work and Schmidhuber's contributions.
• Core disagreement: Marcus advocates for "innateness" (hard-wired priors), pointing out that LeCun's own greatest contribution — convolutional neural networks — embeds the innate prior of translation invariance.
• LeCun's "A Path Towards Autonomous Machine Intelligence" (2022 manifesto) proposes multi-module world models but lacks an implemented theory, in Marcus's view.
• GPT-3 failure example: models incorrectly say Trump is still president due to lack of temporal reasoning; also answer "brilliant table" > child for advice due to statistical pattern-matching only.

Connections: Gary Marcus · Yann LeCun · Meta · Judea Pearl · Deep Learning · Neuro Symbolic AI · Scaling Laws · Large Language Models

Source: https://www.zdnet.com/article/ai-critic-gary-marcus-metas-lecun-is-finally-coming-around-to-the-things-i-said-years-ago/