Gary Marcus argues that generative AI is failing to meet expectations. He points out that large language models still can't be trusted and rely too heavily on memorization instead of genuine reasoning. These systems aren't delivering the quantifiable value promised to businesses and the public.

The core issue involves overpromising by tech creators. The technology works for specific tasks, but its scope is substantially smaller than marketed. This gap between hype and reality matters because economies and policies are being shaped around unproven hopes for dramatic improvements that aren't materializing.

Recent findings support Marcus's skepticism. A Washington Post report on the Remote Labor Index showed AI could only handle about 2.5% of jobs. Attempts to scale these models are hitting walls, suggesting they won't solve fundamental limitations. Investing heavily in this 'shoddy technology' looks increasingly risky.

What happens next? Marcus has consistently warned against betting everything on current AI architectures. His critique suggests the industry needs a reality check before more capital is wasted. The focus should shift toward more reliable, less hyped approaches that actually deliver measurable business value.