Luke Thinks Learning Lab

Luke Thinks Podcast

A learning lab in audio. Analysis, experiments and half-formed ideas from a finance professional in the Netherlands — the thinking, not just the conclusions.

Latest episode

Episode 5 Debate 30 min

The Ironies of Automation & The Art of Oversight

What happens when human operators supervise automated systems: Bainbridge's ironies of automation applied to Air France 447, medical deskilling trials, and code defects.

Show notes →

In this episode

  • The Ironies of Automation: Automating routine tasks leaves humans with only rare, catastrophic edge cases they are deskilled from solving.
  • Automation Bias in High-Stakes Domains: Board-certified radiologists saw diagnostic accuracy drop from 82.3% to 45.5% when primed with faulty AI advice.
  • The Illusion of the Loop: Keeping a human in the loop is useless if deskilling destroys the human’s independent mental model.

Connected to

Series

Multi-part episodes, best heard in order

All episodes

5 episodes, newest first

  1. Episode 5 Latest Debate 30 min

    The Ironies of Automation & The Art of Oversight

    What happens when human operators supervise automated systems: Bainbridge's ironies of automation applied to Air France 447, medical deskilling trials, and code defects.

    Connected to

    Key takeaways
    • The Ironies of Automation: Automating routine tasks leaves humans with only rare, catastrophic edge cases they are deskilled from solving.
    • Automation Bias in High-Stakes Domains: Board-certified radiologists saw diagnostic accuracy drop from 82.3% to 45.5% when primed with faulty AI advice.
    • The Illusion of the Loop: Keeping a human in the loop is useless if deskilling destroys the human’s independent mental model.

  2. Episode 4 Debate 30 min

    Why Your Brain Needs Cognitive Friction

    Examining whether general thinking skills exist independently of subject matter, and measuring what happens to the human brain when the friction of thinking is offloaded to AI.

    Connected to

    Key takeaways
    • The Myth of Far Transfer: Second-order meta-analyses show zero far transfer from brain training; expertise is domain knowledge in disguise.
    • Cognitive Debt & Persistence: AI offloading produces immediate test gains but causes an 11% drop in 45-day retention and doubles skip rates.
    • Evaluating Requires Producing: Reviewing AI outputs requires the exact mental templates built through effortful manual production.

  3. Episode 3 Debate 49 min

    Why AI Verification Replaced Prompt Engineering

    From vibe coding to context engineering: how the $300k prompt engineering hype gave way to the disciplined era of agent harnesses, context window curation, and verification.

    Connected to

    Key takeaways
    • Context Window Dynamics: Models do not reason uniformly; keeping context below 40% capacity preserves peak model intelligence.
    • The Verification Bottleneck: The scarce resource in the AI economy is specifying requirements, managing context, and verifying output.
    • Harness Governance: Autonomous agents without strict boundary harnesses and human verification loops compound errors exponentially.

  4. Episode 2 Debate 15 min

    Big Tech AI Investments: The Great Accounting Debate

    An in-depth 15-minute debate dissecting both sides of the AI investment thesis. Is Microsoft's $13B OpenAI commitment an ingenious infrastructure land-grab, or a precarious house of cards vulnerable to sudden regulatory intervention and efficiency gains?

    Connected to

    Key takeaways
    • Capex vs. Opex Depreciation: Hyperscalers extending server useful lives from 3 to 6 years to temporarily lower annual depreciation charges.
    • Circular Velocity: Compute credits deployed as equity versus direct cash distributions obscuring genuine customer economics.
    • Open vs. Closed Models: Open-source model parity compressing inference margins across all hyperscalers.

  5. Episode 1 Briefing 4 min

    The $50B AI Partnership Bubble: Accounting Crisis in Big Tech

    Tech giants have deployed over $50 billion into generative AI startups through structured equity and compute-credit partnerships. But standard GAAP rules were never designed for investments where cash immediately loops back as cloud hosting revenue.

    Connected to

    Key takeaways
    • Governance Nuance: Equity stakes structured at 49% or non-voting rights circumvent mandatory consolidation under ASC 810.
    • Round-Tripping Concerns: Multi-billion capital deployed as equity loops back into Azure cloud hosting revenue within quarters.
    • Impairment Horizon: Rapid advances in model efficiency risk sudden multi-billion impairment write-downs on proprietary infrastructure.