Gary Marcus: Neurosymbolic AI Rescues Pure LLM Scaling That Hit Its Ceiling

Across September 2–3, Gary Marcus posted repeatedly to restate his long-held position: pure LLM scaling has hit a ceiling, and current progress comes from people patching LLM weaknesses with "tools" and "frameworks"—precisely the premise of Neurosymbolic AI, combining powerful-but-unreliable machine learning (neuro) with reliable-but-limited deterministic software (symbolic).

Confirmed

• Marcus quoted and fully endorsed Jason Pontin's assessment: LLMs trained on the full text of the web are genuinely useful for search and many generative tasks, but they lack theory of mind and world models, cannot do sustained computation, drift easily, and are costly to train and run; achieving AGI requires new methods.
• Asked whether he would claim victory for symbolic AI if neurosymbolic approaches succeed, Marcus replied this has been his core argument for the past 25 years—timing is secondary.
• He also amplified what he called a key point everyone in AI should understand: neurosymbolic methods by definition use deterministic software to compensate for machine learning's reliability shortcomings.

Why it matters

These posts continue Marcus's long-running debate with pure-scaling advocates. If the "tools plus frameworks" route indeed becomes mainstream, it would support his judgment that LLMs alone cannot reach AGI, with implications for the industry's choice of technical direction.

2026-09-02 ~ 2026-09-03 · 5 related posts

• [source] Gary Marcus: Neurosymbolic AI rescued pure LLM scaling — GaryMarcus · 2026-09-02
• [source] Gary Marcus reiterates neurosymbolic victory validates his 25-year claim — GaryMarcus · 2026-09-02
• [source] Gary Marcus Endorses Critique: LLMs Lack World Models, AGI Needs New Techniques — GaryMarcus · 2026-09-02
• Gary Marcus: neurosymbolic AI pairs reliable deterministic software with capable but unreliable ML — williamtp · 2026-09-03

1 near-duplicate retellings: GaryMarcus