

Is AI Scaling Dead? — With Gary Marcus
210 snips May 7, 2025
Cognitive scientist and AI skeptic Gary Marcus joins the discussion to explore whether large language model scaling has hit a ceiling. He shares insights on diminishing returns in AI effectiveness and critiques the reliance on GPU scaling. The conversation touches on data privacy issues, ethical considerations in AI development, and the risks associated with both open-source and proprietary models. Marcus emphasizes the need for transparency and a more nuanced understanding of AI's future trajectory.
AI Snips
Chapters
Books
Transcript
Episode notes
Diminishing Returns from AI Scaling
- The scaling laws of AI models suggested predictable improvement with more data and compute but are now showing diminishing returns.
- The industry is quietly acknowledging that making models bigger no longer yields the past exponential performance gains.
Scaling Up Models Yields Small Gains
- Recent huge increases in model size do not correspond to proportionate improvements in performance.
- The era of large jumps in AI model quality from just scaling up appears to be over; gains now are incremental and costly.
Test Time Compute Offers Narrow Gains
- Adding more compute at test time (reasoning steps) helps improve performance only in narrow, closed domains.
- This method doesn't yield broad reasoning improvements and the models often just mimic human reasoning patterns, not deeply understand them.