Gary Marcus and Developers Clash Over LLMs vs. Neurosymbolic AI

A broad argument broke out on X between Gary Marcus and developers including Matt Shumer, jachiam0, and danielmac8 over what should get credit for recent AI progress. The developers argued that Marcus has repeatedly predicted ceilings for deep learning and then moved the goalposts after ChatGPT and newer reasoning/tool-using systems succeeded. Marcus replied that he has never opposed AI progress itself; his target, he said, has always been the idea that pure LLMs alone are enough, while the real advances come from hybrid systems that combine models with tools and external structure.

Confirmed

• Matt Shumer said Marcus has spent years making pessimistic calls about deep learning and LLMs, including claims that the field was hitting a wall before ChatGPT, and argued that Marcus keeps shifting his standards when new capabilities appear.
• jachiam0 and danielmac8 criticized Marcus for long dismissing pure LLM approaches while, in their view, failing to show concrete neurosymbolic results that outperform them. They argued that when reasoning models and tool-using systems became strong, Marcus started counting those gains as validation of his own position.
• Marcus said he is not anti-AI and welcomes progress in areas like science and medicine. His objection, he said, is specifically to the claim that end-to-end, tool-free LLMs are sufficient. He defined systems that mix neural models with symbolic operations, tools, or outer frameworks as hybrid or neurosymbolic systems.
• Marcus further argued that neurosymbolic methods helped "rescue" deep learning, cited Anthropic-related progress as evidence that tools and symbolic structure improve performance, and claimed that leading AI companies are effectively building neurosymbolic systems whether or not they label them that way.
• Shumer pushed back that Marcus's earlier notion of "hybrid" referred to neurosymbolic architectures, not simply today's LLM-plus-tool workflows. He also argued that if relying on tools disqualifies a system from being AGI, then humans would fail that standard too.

Unconfirmed

• The strongest unresolved issue is definitional. Marcus treats tool use and outer frameworks as part of neurosymbolic AI, while critics argue that this broadens the term after the fact. Based on these posts alone, there is no shared operational definition that settles whether both sides are describing the same class of systems.

Why it matters

• The exchange captures a core split in AI discourse: are current gains mainly the result of scaling and training better LLMs, or of wrapping them in retrieval, tools, loops, and other system-level scaffolding?
• That distinction affects how people interpret progress toward reasoning and AGI. Posts from altryne, danielmac8, and ZeeshanZiaML also suggest that apparent progress on hard math may often depend on external tools, iterative frameworks, or specialized non-LLM systems rather than on a standalone model alone.

2026-08-02 ~ 2026-08-03 · 19 related posts

• [source] Matt Shumer Calls Out Gary Marcus for Constantly Moving the Goalposts on AI — mattshumer_ · 2026-08-02
• Gary Marcus Claps Back at Shumer Over Deep Learning 'Hitting a Wall' Claims — GaryMarcus · 2026-08-02
• [source] Mocking Gary Marcus: Are Reasoning Models Now 'Neurosymbolic'? — jachiam0 · 2026-08-02
• Gary Marcus Fires Back: I Don't Hate AI Progress, I Doubt LLMs — GaryMarcus · 2026-08-02
• Gary Marcus hits back: neurosymbolic AI rescued deep learning from the wall — GaryMarcus · 2026-08-02
• Matt Shumer refutes Gary Marcus: tools + LLMs is what worked, not neurosymbolic — mattshumer_ · 2026-08-02
• [source] Gary Marcus Defends Neurosymbolic AI: Pure Neural Nets Won't Reach AGI — GaryMarcus · 2026-08-02
• Dev Pushes Back on Gary Marcus: Humans Use Tools Too, Capability is What Matters — mattshumer_ · 2026-08-02
• Gary Marcus Claims Anthropic's Progress Vindicates Neurosymbolic AI — GaryMarcus · 2026-08-02
• Researcher Slams Neurosymbolic AI Advocates: Pure LLMs Haven't Hit a Wall — daniel_mac8 · 2026-08-02
• Scaling LLMs Isn't Enough for Math: The Need for Tools and Loops — altryne · 2026-08-02
• Discussion: Pure LLM Architectures Struggle with Reliable Math Reasoning — daniel_mac8 · 2026-08-02
• Gary Marcus: Leading AI Companies Are Secretly Building Neurosymbolic Systems — GaryMarcus · 2026-08-03
• LLMs Solving Math is Just RAG++: Debunking the Deep Reasoning Myth — ZeeshanZiaML · 2026-08-03
• Gary Marcus Endorses That Deep Learning is Hitting a Wall as Reasoning Loops Shift — GaryMarcus · 2026-08-03
• Gary Marcus: The Real LLM Breakthrough Is System-Level Integration — GaryMarcus · 2026-08-03
• antirez Slams Anti-LLM Camp: Moving the Goalposts is a Historic Scientific Error — solyarisoftware · 2026-08-03

2 near-duplicate retellings: GaryMarcus · GaryMarcus