The MAD Podcast with Matt Turck

Episode: "AI Eats the World: Benedict Evans on What Really Matters Now"

Date: May 22, 2025
Guests:

• Host: Matt Turck
• Guest: Benedict Evans, Analyst, Advisor on Platform Shifts

Episode Overview

Matt Turck hosts returning guest Benedict Evans for a penetrating, candid discussion on the current reality and future prospects of generative AI. The conversation ranges from model commoditization, the hype vs. actual enterprise deployment, the fate of AI "doomerism," real-world utility, the evolution of interface and distribution, and the practical limits of today’s technology. Evans brings perspective from two decades following tech platform shifts, and he doesn’t shy away from challenging industry dogma or media-driven narratives.

Key Discussion Points & Insights

1. Is AI a Platform Shift or a Paradigm Shift?

• No Clear Consensus Yet (02:13)

  • Despite model improvements and a shift to post-training, it's not obvious AI is driving radical change akin to previous tech transformations. Many questions from early 2023 remain unanswered.
  • Benedict Evans: "A lot of the questions you could have asked in the beginning of 2023 don't really have answers yet." (03:12)

• AI Models as Commodities (03:30)

  • There are now several state-of-the-art models (OpenAI, Anthropic, Gemini, etc.), and the gap between them is closing. The bigger differentiator is brand and distribution, not core technology.

2. Three-Layer Model of AI Industry

• Benedict’s Framework:
  1. Model Wars: Technical arms race (comparable to Moore’s Law era, but with many players)
  2. Enterprise SaaS Adopters: Companies leveraging APIs to solve narrow, industry-specific problems (e.g., automating HR or accounts payable) (04:50)
  3. The "Fuzzy Middle": Grand narratives about AI that don't match operational reality—reminiscent of the Metaverse hype (05:39)

"On one hand, oh my God, have you seen the new model? ... but it still can't actually replace any of the software you use. It can't replace Excel." (06:30, Evans)

3. Error Rates: Probabilistic vs. Deterministic Systems

• Critical Distinction (07:48)
  • Many use cases are tolerant to error; others require determinism. LLMs are not reliable for "right answers" every time (i.e., not Oracle), limiting some product opportunities.

"There's an enormous difference between saying that was correct 89% of the time and now it's correct 91% of the time, and on the other hand saying that was wrong and now it's right. Those are completely different things." (07:48, Evans)

• "Deep Research" as Case Study (11:55)
  • Using AI for generating reports in known areas is valuable, but for unfamiliar topics, trustworthiness collapses due to hallucinations and subtle error propagation.

"If I actually want that table, I'm going to need to check every single cell in the table myself. In which point, why would I use deep research in the first place if I'm going to have to check every single thing it gives me?" (13:17, Evans)

4. Industry Overpromising & Adaptation

• Forcing Old Use Cases (16:15)
  • Early adoption forces AI to mimic existing deterministic tools instead of creating new workflows tailored to probabilistic capabilities.
• Religious Fervor & Critique (17:31)
  • Belief often substitutes for practical understanding, leading to bubbles and flawed criticism.

5. AI Hype vs. Real Adoption

• Real-World Use Cases Proliferate (17:44)
  • Unlike crypto’s search for application, generative AI is already deployed in thousands of companies, especially in marketing, customer service, and code generation.

"There are hundreds and hundreds of companies who've already got this in production... but at the same time it's not good at everything." (17:44, Evans repeats his podcast opener for emphasis)

6. Generational Change in AI Acceptance

• AI as Intuitive Tool for Next Generation (19:18)
  • Usage patterns (e.g., ChatGPT dips aligning with school holidays) hint at homework and educational use.
  • Younger users may naturally accept non-deterministic results, unlike previous generations.

"As this generation that grows up with these tools enters the workplace... you'll have people that say, of course it's AI, it's non-deterministic, you have to use it for what it's good at." (20:03, Turck and Evans)

7. Model Commoditization, Brand & Distribution

• Biggest Moat is Distribution (22:10)
  • OpenAI dominates consumer mindshare, despite nearly equivalent models from competitors.
  • Sam Altman’s primary role now is brand and distribution, not just tech leadership.

"Why is ChatGPT at the top of the App Store chart and has been for a year?... It's kind of a distribution and brand and reach story." (22:26–24:25, Evans, echoed throughout)

• Hiring Signals Strategic Shift (24:38–25:19)
  • OpenAI and Anthropic hire consumer-oriented executives (e.g., Instacart & Facebook alumni) to build application layers, signaling focus beyond pure API access.

8. Thin Wrappers vs. Real Applications

• Enterprise SaaS is not Just a Wrapper (25:29–26:57)
  • Vertical applications require deep domain knowledge and customization; “thin GPT wrapper” is a misnomer outside simple chatbots.

9. Corporate Strategy & the Infrastructure Arms Race

• Big Tech Motives (30:05–33:04)

  • Capital expenditure and infrastructure development (data centers, GPUs) lead the charge; companies hedge on where value will concentrate—APIs, platforms, or consumer apps.

• No Master Plan Myth (32:15)

  • Much as with past tech giants, AI labs are improvising, not enacting a secret, perfect strategy.

10. Big Tech: Apple, Google, Meta, and AWS

• Apple: Struggles With Ambition & Reality (34:14–37:42)

  • Apple demoed multi-modal, agentic AI features (Siri 2.0) at WWDC—but much of it wasn’t built, exposing internal confusion and the difficulty of moving from demo to reality.

• Google: Incumbent’s Dilemma (37:48–39:11)

  • Google’s search franchise is directly threatened, but adapting is hard for an incumbent with decades of assumptions.

• Meta: UI Experiments & Distribution (43:01–46:53)

  • Meta tries to bundle AI into existing platforms (Instagram, WhatsApp), and experiments with standalone apps and feeds to suggest use cases, aiming for the elusive “viral loop” in AI.

• AW$ and Infrastructure Providers (41:57–42:41)

  • AWS and Meta benefit as AI drives insatiable demand for compute; commoditization at the infra layer is their core business logic.

11. Consumer AI & the Search for a Breakout App

• Current Consumer Apps—Few True Successes (46:53–48:49)

  • Image generators (Midjourney) and "AI companion" tools are notable, but no verticalized consumer killer app has emerged yet, unlike in enterprise.

• Generative Content: New Possibilities & Uncertainties (49:00–51:02)

  • Generative AI challenges the meaning of content (e.g., on Instagram) and can disrupt markets like advertising or e-commerce, but broad consumer impact is still forming.

12. Business Models, Memory, and Data Moats

• Ads and Personalization (51:21–53:31)
  • Is ChatGPT evolving toward an ad-supported model like Google? Data privacy regimes (EU) will shape monetization approaches.
  • Companies are building “interest graphs” of users; whether this yields a strong moat is up for debate.

13. Enterprise AI: Real Use, Real Limitations

• Pilots Everywhere, Real Deployment Growing (55:25–58:20)
  • Most large enterprises have piloted or deployed LLM-based features (automation, search, tagging, recommendations). Consultants (Accenture, BCG) thrive in this environment.

"Everyone's done a bunch of pilots. Did some of it not work? Well, yeah. That's what you do pilots for." (56:23, Evans)

• Stages of Adoption (58:34–60:07)
  • Step 1: Automate known workflows
  • Step 2: Innovate on products
  • Step 3: New business models, e.g., Airbnb/Uber-level disruption (not yet visible)

14. AI Agents: Hype and Hope

• Skepticism Re: Agents (64:07–67:10)
  • AI agents purported to automate multi-step tasks, but real-world demos are limited or fail under exception handling and reliability requirements.

"These agent demos where they do all these multi-stage things...it's not a real demo, it's not working." (67:10, Evans)

15. AI Doomerism: Why it Faded

• Circular Debates & Social Realities (69:42–73:44)
  • AI existential-risk ("doomer") arguments lost traction; consensus in elite circles now is that practical and social concerns (abuse, deepfakes, biases) are real, but apocalyptic logic is circular and unpersuasive.

"They invited all the doomers to Davos in 2024 and they listened to them and thought these people are idiots and didn't invite them back ... a logically flawless circular argument." (70:09, Evans)

Notable Quotes & Memorable Moments

• "You can't just hand wave away the fact that these things are wrong sometimes and you have to think about what you do with that." (10:59, Evans)
• "The emperor isn't naked ... it's already being used and it's really useful, but at the same time it's not good at everything." (17:44, Evans)
• "The challenge ... is there's a sort of Emperor's new Clothes problem in that. But it's not ... the emperor isn't naked." (17:44, Evans)
• "All AI questions have one of two answers: either it will be exactly like every other platform shift, or no one knows." (20:12, Evans)
• "Distribution and brand and reach ... that's the story." (24:25, Evans)
• "Why did they show something that wasn't built (Apple)? That's a bigger problem. And why hasn't it been built yet? Because like nobody's got that built working." (37:48, Evans)
• "They were all really clever people who all lived in group houses in Berkeley and all talked to each other and told each other how clever they were and constructed these logically flawless circular arguments." (70:09, Evans)

Important Timestamps

• Commoditization & Distribution of AI Models: 02:13–04:50
• Error Rates & Probabilistic Systems: 07:48–15:07
• Forcing Old Use Cases on New Tech: 16:15–19:18
• Generational Shift in AI Acceptance: 19:18–20:12
• OpenAI’s Brand & Consumer Dominance: 22:10–24:25
• Enterprise AI Reality: 55:25–58:20
• AI Agents & Hype: 64:07–67:10
• Death of Doomerism: 69:42–73:44

Takeaways for Listeners

• AI progress is real and significant, but not evenly distributed; utility is booming in enterprise and coding, less so in revolutionizing consumer apps or creative autonomy.
• Error rates, trust, and workflow adaptation are pivotal in determining what AI can do today—and what remains out of reach.
• Distribution, user mindshare, and brand are currently bigger differentiators than “model quality.”
• The existential-dread narrative has lost influence, replaced by pragmatic risk assessment as AI weaves into daily life and enterprise software.
• While enterprise AI automates and adds value, entirely new paradigms—a "Step 3" like Uber for AI—remain undefined.

For more, listen to the full episode or visit Matt Turck’s MAD Podcast page.