Keynotes

Dr Gary Marcus: The Grand AGI Delusion

Marcus urges the field to move toward more structured, hybrid, and modular systems for genuinely capable and trustworthy AI. 

Dr. Gary Marcus delivered the keynote talk titled “The Grand AGI Delusion” at The Royal Society in London on October 2, 2025.

It was hosted by the Web Science Institute to mark the 75th anniversary of Alan Turing’s seminal 1950 paper on machine intelligence (often linked to the Turing Test).

Marcus, Professor Emeritus at New York University and a longtime AI skeptic, critiques the hype around large language models (LLMs) and the pursuit of artificial general intelligence (AGI) via pure scaling.

Core Arguments

Marcus opens by referencing Turing’s question—”Can machines think?”—and distinguishes it from the Turing Test itself. He argues that passing conversational tests (e.g., GPT-4 achieving ~73% success in 5-minute interactions) does not prove real intelligence. He cites the classic “Eliza effect” (from the 1966 chatbot) as an example of how easily people anthropomorphize simple systems and over-attribute understanding to them.

Critique of Current LLMs and Scaling

Marcus highlights fundamental limitations of LLMs:

  • They excel at pattern matching and imitation but struggle with genuine reasoning.
  • Examples of failures include generating images with anatomically correct hands/fingers, solving novel variants of classic puzzles (like river-crossing problems), or performing safe, practical tasks like wiring a plug.
  • These stem from distribution shift—models break down on inputs meaningfully different from their training data.

He attacks the “scaling hypothesis” (pushed by figures like OpenAI’s Sam Altman), which claims that simply adding more data and compute will inevitably lead to AGI. Marcus points to diminishing returns: models like Meta’s Llama series and anticipated GPT-5 have underperformed relative to the hype. While Richard Sutton’s “bitter lesson” (favoring computation over hand-engineered knowledge) holds for some pattern-recognition tasks, it falls short for robust reasoning, generalization, and handling novelty.

Societal and Practical Concerns

Marcus warns about the broader downsides of the current AI frenzy:

  • Massive overinvestment (potentially trillions of dollars) with limited real returns—studies show ~95% of businesses report no meaningful ROI from generative AI.
  • Rising issues like “work slop” (low-value AI-generated content), environmental costs, bias, misinformation, and surveillance risks.
  • He compares unchecked AI development to the regrets around social media, arguing that governments are handing too much power to tech companies with insufficient oversight.

Proposed Path Forward

Instead of chasing a monolithic AGI through scaling alone, Marcus advocates:

  • Modularity: Build specialized, high-performing systems (e.g., AlphaFold for protein folding) rather than expecting one model to do everything.
  • Hybrid architectures: Combine fast, intuitive, data-driven LLMs (akin to Daniel Kahneman’s “System 1” thinking) with slower, symbolic, rule-based reasoning (“System 2”) for better interpretability, verifiability, and reliability.
  • Innate structure and world models: Incorporate prior knowledge and internal models of the world (like rules in chess or physics) instead of forcing everything to be learned statistically from data. He draws an analogy to human cognition, which blends “nature” (innate structures, like a genome) and “nurture” (learning).
  • Overall, prioritize reliable, safe, and understandable AI over hype-driven generality.

Closing Tone

Marcus criticizes hubris among some AI leaders (e.g., Sam Altman and Dario Amodei), who he says risk catastrophic outcomes in pursuit of profit or glory. He calls for humility, echoing Turing’s openness to multiple approaches, and stresses that true progress requires addressing safety and real-world utility rather than chasing the “grand delusion” of imminent superintelligent AGI.

In summary, the talk is a detailed, evidence-based rebuttal to the dominant narrative that current LLMs are on the cusp of AGI. Marcus urges the field to move toward more structured, hybrid, and modular systems for genuinely capable and trustworthy AI.

Related Articles

Back to top button