AI skeptic Gary Marcus now says AI can’t do things it’s already capable of

The rapid advancement of AI capabilities has repeatedly challenged skeptics’ predictions about the technology’s limitations. In early 2020, AI critic Gary Marcus highlighted specific tasks that GPT-2 couldn’t perform, only to see subsequent AI models overcome these limitations, creating a pattern of premature criticism that continues today.

Key Timeline and Pattern: Gary Marcus has established a recurring pattern of identifying AI limitations that are quickly overcome by newer models.

• In 2020, Marcus published critiques of GPT-2’s limitations, suggesting a need for different approaches
• GPT-3 later solved most of these identified problems, prompting Marcus to create a new list of 15 challenges
• GPT-4 subsequently demonstrated capability in handling most of these new challenges

Recent Criticism and Reality Check: Marcus’s latest critique of ChatGPT demonstrates a disconnect between his assessment and current AI capabilities.

• Marcus based his criticism on the free version of ChatGPT (GPT-4o), an outdated model from May 2024
• Using GPT-o1, a more recent model, successfully completed tasks Marcus claimed were impossible
• The tested tasks included creating detailed state-by-state tables and performing accurate vowel counting exercises

Technical Context: The evolution of AI models shows significant progress in handling complex tasks through different approaches.

• Reasoning models, while still based on Large Language Models (LLMs), have enhanced capabilities through reinforcement learning
• These models can self-check their work, leading to improved accuracy
• While some errors may still occur, they’re comparable to human-level mistakes

Notable Quotation: Marcus’s Twitter bio stating he “Warned everyone in 2022 that scaling would run out” highlights his consistent skepticism about AI advancement.

Looking Forward: While Marcus correctly identifies current limitations in AI agents, his pattern of criticizing capabilities that have already been achieved raises questions about the validity of his broader skepticism regarding AI development. The continued progress in AI capabilities suggests that scaling limitations have not yet been reached, though they may eventually emerge as a constraint.