Limits Of Large Language Models And Hallucinations

Edward Chancellor joins Kai Wu to discuss what financial history and capital cycle theory can teach investors about today’s AI boom. They explore why transformative technologies can still produce terrible investor returns, how overinvestment develops, where anti-bubbles may be forming, and what past episodes like the railway mania, the dot-com bubble, China’s investment boom and the post-2008 interest rate regime suggest about the risks and opportunities today.

Guest links:

Edward Chancellor
https://www.edwardchancellor.com/

Papers and articles discussed:

Valuing AI: Extreme Bubble, New Golden Era, or Both
https://www.gmo.com/americas/research-library/valuing-ai-extreme-bubble-new-golden-era-or-both_viewpoints/

Markets have poor scorecard for spotting AI losers
https://www.reuters.com/commentary/breakingviews/markets-have-poor-scorecard-spotting-ai-losers-2026-04-24/

There’s no such thing as a good bubble
https://www.reuters.com/commentary/breakingviews/theres-no-such-thing-good-bubble-2025-10-09/

Big Booze can sweat off its multi-year hangover
https://www.reuters.com/commentary/breakingviews/big-booze-can-sweat-off-its-multi-year-hangover-2025-07-10/

Topics covered:

How capital cycle theory applies to the AI data center boom

Why railway mania, autos, aircraft and the dot-com bubble offer lessons for today

Why markets often fund major technology transitions but fail to identify the winners

The prisoner’s dilemma driving hyperscaler AI spending

Whether AI demand can justify the supply being built

How GPU depreciation and AI capital spending may affect reported earnings

Why hallucinations and reliability may limit the total addressable market for large language models

The case for looking at AI anti-bubbles instead of shorting the bubble directly

Why China shows that strong GDP growth does not guarantee strong shareholder returns

How intangible capital, SaaS valuations and human capital fit into capital cycle analysis

Whether bubbles can be good for society while still being bad for investors

Why the long-term interest rate cycle may have changed

The role of gold in a world of expensive stocks, rising debt and vulnerable bonds

Timestamps:

00:00 Edward Chancellor on capital cycles, bubbles and AI

04:42 Why the railway mania became a classic overinvestment cycle

09:00 Why markets fund technology booms but often miss the winners

13:19 The prisoner’s dilemma behind AI spending

17:30 Will AI demand justify the supply being built

20:00 How capital spending can inflate profits before the bust

25:08 The AI Hindenburg moment and the limits of large language models

30:55 Why AI hype may exceed the proven technology

35:55 Why the anti-bubble may matter more than shorting AI

40:00 The energy transition bubble and the opportunity in overlooked assets

45:08 China’s lesson on GDP growth and shareholder returns

49:27 Big Booze, GLP-1s and the Lindy effect

54:23 Can intangible capital have its own capital cycle

59:54 SaaS valuations and the index creation warning signal

01:04:10 Why bubbles can help society but hurt investors

01:09:09 Why long-term rates may be in a new multi-decade cycle

01:14:07 Why Edward Chancellor still sees a role for gold