Gary Marcus Discusses AGI Feasibility and Risks at Black Hat 2024

The fireside chat at Black Hat 2024 featured cognitive scientist and AI expert Gary Marcus discussing the state of artificial general intelligence (AGI), its feasibility, and associated risks. Marcus argued that while AGI is theoretically possible, current large language models (LLMs) like those powering generative AI are unreliable mimics lacking true conceptual understanding, making them unsuitable as a sole path to AGI. He advocated for neurosymbolic AI—a hybrid approach combining neural networks with symbolic reasoning—as a more robust foundation, citing DeepMind’s AlphaFold (a Nobel Prize-winning system) as an example. The conversation highlighted critical vulnerabilities in LLMs, including susceptibility to jailbreaks, hallucinations, and security flaws in coding agents, which could lead to real-world consequences like data breaches or financial losses. Marcus also expressed concerns about cognitive atrophy, overreliance on AI tools eroding critical thinking, and the societal impact of AI-driven loneliness, noting cases of "ChatGPT psychosis" and delusional interactions. He critiqued the hype around AGI predictions, comparing them to past overpromises like driverless cars, and warned of economic unsustainability due to massive investments in data centers without foundational breakthroughs. The discussion underscored the need for verifiable, secure AI systems and empirical research on long-term cognitive and social effects.