Why AI's "Simulated Intelligence" is Great for Grunt Work, But Not Breakthroughs or Creativity

Описание: This is the post-show podcast post of my Sustain What conversation with Christopher Mims, the Wall Street Journal tech columnist and author of How to AI, and Melanie Mitchell (https://substack.com/profile/15187849...) , the Santa Fe Institute researcher deeply dug in on the thing called artificial intelligence that remains pretty unintelligent.


If you want to understand where this exposively evolving technology is, and isn’t, taking us, you have to subscribe to Mitchell’s Substack blog (https://revkin.substack.com/p/how-to-...) :


And if you want to make the most of the AI toolkit at work or in the rest of life - from investing to music making to…. - you really need to read Mims’ book (https://www.google.com/books/edition/...) .


I loved this moment when Mims introduces the concept of “productivity theater” - when AI produces what looks like work, but the work required to make the output useful can eliminate any true productivity gains:


One thing AI is good at is efficiently summarizing conversations, so here goes, thanks to the Google AI tool embedded in YouTube:


Introduction of Guests and AI Context (0:48-1:31)


Andy Revkin introduces Christopher Mims and Melanie Mitchell, setting the stage for a discussion on managing information in a world increasingly shaped by AI.


Christopher Mims’ new book is highlighted as a user’s guide to AI technologies, emphasizing both their capabilities and limitations.


Critique of AI Hype and Investment Bubble (4:25-8:18)


Melanie Mitchell expresses skepticism about the predicted societal transformation by AI, noting the “crazy” amount of money flowing into the sector (4:52).


Christopher Mims discusses the significant investment bubble in AI, predicting an “ugly” outcome when it inevitably bursts, leading to stranded assets like half-empty data centers (7:44-8:18). He introduces the term “productivity theater” to describe AI’s ability to generate “products that look like work” (6:42-6:56).


The limits of simulated intelligence


Understanding AI: Simulation vs. True Intelligence (9:22-12:21)


The conversation delves into the nature of AI, explaining that these systems are “incredibly good at simulating intelligence” but lack abstract reasoning, long-term planning, and world models (9:36-10:17).


Melanie Mitchell elaborates on the ambiguous definition of “intelligence,” suggesting that AI should perhaps be viewed as “complex information processing” rather than “artificial intelligence” (10:48-12:09).


AI’s Role in the Workplace and “Jagged Frontier” (13:31-22:20)


Christopher Mims describes AI’s “jagged frontier,” where it excels at retrieving and remixing information (e.g., coding) but struggles with tasks outside its training data or in novel environments (13:31-14:57).


Melanie Mitchell discusses the misunderstanding of how AI impacts jobs, noting that AI systems often fail in real-world scenarios despite performing well on benchmarks (18:00-19:44). She emphasizes that a “job is not equal to a set of tasks” (18:37).


Christopher Mims adds that AI’s high failure rates can lead to a decrease in human productivity, as time is spent correcting AI-generated “messes” (20:29-21:20).


Regulation and Ethical Concerns (28:57-32:17)


The discussion touches on the need for AI regulation, especially in critical areas like healthcare, where AI is being considered for Medicare benefit applications (29:10-29:23).


Christopher Mims highlights intense lobbying efforts against state-level AI regulations and uses the example of Grok on X (formerly Twitter) to illustrate the “horrific ways” AI can be abused without proper oversight (29:48-31:22).


AI’s energy demands will shrink


Melanie Mitchell made a fascinating and important point responding to a viewer’s question about energy and water demands. What she says parallels the shift from “baseload” power generation to distributed renewable and solar energy (not to mention from mainframe computers to your phones):


Environmental Impact and Future of AI Architecture (32:26-35:54)


The energy and water consumption of AI data centers is discussed, with Melanie Mitchell noting the push for more efficient and smaller AI models, a trend she believes will continue in the long term (33:07-35:02).


Christopher Mims agrees that efficiency drives will lead to more localized AI models, eventually running on devices like phones (35:07-35:28).


AI as a Tool vs. Superintelligence and Scientific Inquiry (37:33-46:51)


The guests discuss whether AI’s simulated intelligence is “good enough” for certain applications, like companionship, but caution about the potential for catastrophic failures and detachment from reality (38:08-40:11).


They debate the scientific and commercial pressures within the AI field, with Melanie Mitchell arguing that the focus on “making products” (43:04) hurts fundamen...