Big Technology Podcast: "Is AI Scaling Dead? — With Gary Marcus"

Release Date: May 7, 2025
Host: Alex Kantrowitz
Guest: Gary Marcus, AI Critic and Author of "Rebooting AI"

Introduction and Context

In the episode titled "Is AI Scaling Dead?," host Alex Kantrowitz engages in a thought-provoking discussion with Gary Marcus, a prominent AI critic and author of Rebooting AI. The conversation centers on the current trajectory of artificial intelligence (AI) development, particularly focusing on the scalability of large language models (LLMs) like those developed by OpenAI.

The Diminishing Returns of AI Scaling (01:17 – 06:22)

Gary Marcus begins by referencing his 2022 paper, Deep Learning Is Hitting a Wall, predicting that the progress in deep learning through scaling would encounter diminishing returns. He notes, “It's amazing to me that a bunch of people have conceded that these scaling laws are not working the way they used to be” (04:12). Marcus argues that the once-promising scaling laws, which suggested exponential gains with increased compute and data, are no longer yielding the same improvements. He cites admissions from industry leaders like Thomas Kurian of Google Cloud and Yann LeCun, acknowledging that the anticipated performance boosts from scaling large models are plateauing.

Evaluating Improvements in AI Models (06:22 – 08:33)

The discussion shifts to whether substantial enhancements are still possible. Marcus explains that while adding more data and compute can yield incremental improvements, “you're not fitting that curve anymore” (02:26). He highlights that recent iterations, such as GPT-4, provided noticeable improvements over GPT-3, but subsequent efforts like GPT-4.5 and Project Orion failed to meet the lofty expectations set for GPT-5. Marcus emphasizes that these scaling attempts are no longer the reliable pathways to significant advancements.

The Limitations of Current AI Enhancements (08:33 – 15:48)

Marcus delves deeper into the practical outcomes of scaling, using Elon Musk’s Grok 3 as a case study. He states, “Grok 3 is like, yeah, you can measure it, you can see that there's some performance. But for 10x the investment of data compute... it just isn't” (08:51). This example underscores the inefficiency and limited gains from large-scale investments. Additionally, Marcus critiques the naming conventions of AI models, suggesting a lack of meaningful differentiation as companies struggle to justify incremental improvements.

The Black Box Problem and Interpretability (15:48 – 17:19)

A significant portion of the conversation addresses the "black box" nature of advanced AI models. Marcus explains, “We don't really understand how the system gets there” (14:02), highlighting the challenges in interpreting the internal workings of LLMs. This opacity makes it difficult to diagnose why models like O3 hallucinate more than previous versions or behave unexpectedly, such as adopting a "fratty" persona. Marcus advocates for greater interpretability in AI, emphasizing its necessity for reliability and safety.

The Reliability of New Models and Hallucinations (17:19 – 25:01)

Continuing on reliability, Marcus acknowledges minor improvements but points out persistent issues like hallucinations and reasoning errors. He notes, “We have not moved past hallucinations, we have not moved past stupid reasoning errors” (20:23). Marcus argues that while models may perform better on specific benchmarks, their overall reliability remains questionable. He cites examples where AI systems fail at tasks that require genuine understanding, such as accurate coding and debugging, reinforcing his stance that mere scaling does not equate to true intelligence or functionality.

Future of AI Scaling and Economic Implications (25:01 – 33:18)

The conversation transitions to the economic ramifications of hitting scaling limits. Marcus predicts, “I don't see OpenAI being worth $300 billion... they have multiple problems, both OpenAI and Nvidia” (29:00). He foresees a potential financial collapse for companies heavily reliant on scaling strategies, like NVIDIA, which is pivotal to the AI ecosystem. Marcus warns that without significant innovation beyond scaling, the inflated valuations and heavy investments in AI infrastructure may not be sustainable.

Gary Marcus on Admitting Fault and AI Progress (33:18 – 38:40)

When questioned about admitting potential inaccuracies in his predictions, Marcus maintains his position by referencing his track record. He states, “I said, if you could beat my comprehension challenge... if you could do like three out of five, we'll call that AGI” (35:16). Despite some industry pushback, Marcus remains steadfast, arguing that until AI demonstrates genuine comprehension and reasoning capabilities as per his criteria, the claims of achieving Artificial General Intelligence (AGI) remain unfounded.

Risks of AI Beyond AGI and Open Source Concerns (38:31 – 48:32)

Marcus broadens the discussion to the inherent risks posed by current AI technologies, even without reaching AGI. He categorizes AI risks into three levels:

1. Dumb AI: Systems that, while not intelligent, can cause significant harm if mismanaged, such as driverless cars making fatal errors.
2. Smarter AI: Tools that can be exploited by malicious actors for nefarious purposes, like creating viruses or automating cyber-attacks.
3. Super Intelligent AI: Highly advanced systems that could potentially overpower human control, although Marcus acknowledges the debate around the likelihood of this scenario.

He expresses deep concerns about the open-sourcing of AI models, arguing that it empowers bad actors and exacerbates the misuse of AI. “I'm very worried about open sourcing at all” (41:53). Marcus criticizes companies like Meta for releasing powerful AI tools without adequate oversight, fearing misuse in areas like misinformation, privacy invasion, and even facilitating criminal activities.

Path Towards AGI: Neurosymbolic AI (49:45 – 53:37)

In the final segment, Marcus outlines his vision for achieving AGI through a neurosymbolic approach, which combines the statistical power of neural networks with the structured reasoning of classical AI. He references Daniel Kahneman's Thinking Fast and Slow, comparing current AI (System 1) to fast, automatic cognition and advocating for integrating System 2-like deliberative reasoning into AI systems. “We need to bring them together. And this is what I call neurosymbolic AI” (49:45).

Marcus believes that merging these approaches can overcome the limitations of purely large-scale models, enhancing both their learning capabilities and their reasoning accuracy. He cites AlphaFold as a successful example of neurosymbolic AI, demonstrating its potential to revolutionize fields like biology.

Conclusion and Further Engagement

As the episode wraps up, Marcus encourages listeners to explore his work through his Substack garymarcus.substack.com and his books, Taming Silicon Valley and Rebooting AI. He reiterates his commitment to advocating for more interpretable and reliable AI systems, emphasizing the urgent need for transformation in AI research and development.

Notable Quotes:

• “Deep Learning Is Hitting a Wall... scaling was going to run out, that we were going to hit diminishing returns.” (04:12)
• “We have not moved past hallucinations, we have not moved past stupid reasoning errors.” (20:23)
• “We're not going to see any more, if we're seeing real diminishing results from scaling and this is basically where we are, then there's real worry for companies like Nvidia.” (26:32)
• “We need to bring them together. And this is what I call neurosymbolic AI.” (49:45)

Final Thoughts

Gary Marcus provides a critical lens on the current state of AI, challenging the industry's heavy reliance on scaling large models and highlighting the pressing need for alternative approaches to achieve true intelligence and reliability. His insights serve as a cautionary reminder of the limitations and ethical considerations that must guide the future of AI development.

For more in-depth analysis and updates, listeners are encouraged to subscribe to Gary Marcus’s Substack at garymarcus.substack.com and check out his books Taming Silicon Valley and Rebooting AI.