GPT-5’s underwhelming August 2025 release became fresh evidence for Gary Marcus’s long-running argument that scaling LLMs alone will not produce AGI.
Who they are
Gary Marcus appears in coverage as a cognitive scientist and prominent AI critic whose commentary connects technical limits of large language models with arguments for stronger AI governance. Stories place him chiefly in debates around OpenAI, AGI, model hallucinations, and the adequacy of scaling-driven development.
The recent arc
Coverage was most concentrated in 2023Q4, then remained consistently active through 2024 and 2025 as generative-AI launches and governance disputes created recurring openings for Marcus’s critique. OpenAI’s leadership upheaval, the New York Times copyright suit, GPT-4o’s launch, and staff warnings about secrecy and recklessness formed the wider industry backdrop to reporting that repeatedly involved OpenAI, Sam Altman, Google, and frontier-model risks.
By 2025, the emphasis sharpened around whether new models substantiate AGI claims. Marcus characterized GPT-5’s August release as incremental and argued that it showed pure scaling is not the route to AGI; this extends his earlier case that deep learning needs a paradigm shift. He also remained active in disputes over OpenAI’s policy conduct and in rebuttals surrounding Apple’s “The Illusion of Thinking” paper.
The tension
The central tension is Marcus’s challenge to the prevailing frontier-AI premise that larger, scaled LLMs can steadily deliver general intelligence. His criticism targets the reliability, reasoning, hallucination, and moral limitations of LLMs, putting him at odds with the product and AGI narratives associated with OpenAI and its peers, while his exchanges with Casey Newton and responses to defenses of Apple’s paper show that the dispute extends to how those limitations are interpreted.
Why it matters
If model releases continue to yield contested or incremental gains rather than clear advances in robust reasoning, Marcus’s technical critique could increasingly reinforce the case for oversight alongside capability development. His proposals for a dedicated U.S. AI agency and an FDA-modeled federal department make the stakes institutional as well as technical, though the corpus does not establish whether policymakers or major labs will adopt that approach.
Related: OpenAI · LLM · AGI · Sam Altman · OpenAI unveils GPT-4o, a new flagship generative AI model that is fast
Gary Marcus has appeared in 81 articles since 2016-12.
Coverage peaked in 2023Q4 with 10 articles.
Frequently mentioned alongside OpenAI, @garymarcus, Mastodon, Google.