Where AI Is Headed: Platform Shifts, Employment Anxiety & the Real Value of Model Companies

#564. Where Will AI Go: Platform Migration, Employment Anxiety, and the True Value of Model Companies

📝 Podcast Summary

This episode is a clone of: Silicon Valley’s top venture podcast, “Lenny’s Podcast” **A rational conversation on where AI is actually going | Benedict Evans**

Guest this episode: Benedict Evans, an independent tech analyst who has long tracked platform migration in technology. He was previously a partner at a16z and has years of experience in equity research. His latest presentation, titled “AI Is Eating the World,” seeks to answer a question on everyone’s mind: How will AI truly transform our work, business, and lives?

In this episode, Benedict offers a冷静 yet controversial perspective: AI’s significance will be as profound as the internet or mobile internet — but also “only” as significant as those. He believes we may currently be at a stage similar to 1997 in the internet era — direction matters immensely, but most product forms, business models, value flows, and organizational changes remain undefined.

This conversation covers AI’s impact on employment, why AI labs are hiring more people, whether model companies will become low-margin infrastructure, why application layers and distribution might become more important, where anti-AI sentiment comes from, and what ordinary people should do in this uncertain future. Benedict’s core advice is direct: don’t bury your head in the sand, nor just vent anger on social media. The real value lies in personally diving into AI, understanding what it can do for you, and how it will reshape your industry.

👨‍⚕️ Guest Profile

Benedict Evans is an independent tech analyst specializing in the evolution of the internet, mobile internet, platform migration, AI, and structural shifts in the tech industry. He served as a partner at a16z and previously worked in equity research. In recent years, he has consistently tracked the evolution of AI, software, consumer internet, and tech business models through newsletters, presentations, and public talks.

⏱️ Timestamps

00:00 Introduction & Podcast Overview AI is a Big Deal — Internet-Level Significance

04:02 AI Is Eating the World: What We Still Don’t Realize 06:25 Now Like 1997: Most Things Haven’t Been Invented Yet 06:46 Software Development Has Already Changed: Before and After Claude Code 08:10 Jagged Frontier: Where AI Works, and Where It Doesn’t

Why AI Companies Need More “People”

09:18 Forward Deployed Engineer: Why AI Labs Are Starting to Look Like Consulting Firms 11:16 AI Didn’t Kill Consulting — It Made Professional Services More Important 11:32 Task ≠ Job: Why PowerPoint Isn’t What McKinsey Actually Sells 13:40 Jevons Paradox: Why Demand Might Increase After Automation 15:10 From Excel to Accounting: Why Automation Didn’t Eliminate Professional Roles

The End of Jobs? Or Another Platform Migration?

16:08 Even the Most Advanced AI Companies Are Hiring More — What Does This Mean? 16:32 Don’t Trust AI Lab CEOs’ Judgments on Labor Markets 17:20 Every Tech Revolution Destroys Jobs — And Creates New Ones 18:42 Why “All Companies Fire Everyone in Two Weeks” Is a Naive Fantasy 20:10 Corporate Change Is Slow: Sales Cycles, Organizational Systems, Industry Inertia 21:06 “This Time Is Different” — But So Was Every Previous Revolution 22:08 Barcodes, Internet, Google: How We Forgot the Last Great Shift

AGI, Superintelligence, and “We Don’t Know”

23:24 What Makes This Time Truly Different: AGI and Superintelligence 23:46 We Lack a Theory of Intelligence — and Also No Theory of Model Progress 24:50 AI, AGI, Superintelligence: Terms Are Being Redefined 26:02 Even If Models Stop Improving Tomorrow, AI Will Still Reshape the Next Decade

Where Will Value Flow?

26:58 Will Companies Become Unprecedentedly Large? 27:33 Software Continues to Eat the World: How TAM Expands Outward 28:50 Electricity, Utilities, and AI Intelligence: A Useful Analogy 30:10 Will Foundation Models Capture All Value? 31:30 If Models Become Commodities, Value May Shift Up to the Application Layer 32:54 Will Margins Be Squeezed for Model Companies? 33:11 Why Model Companies Are More Like Cloud Providers Than Windows

Investment, Giants, and Distribution Moats

34:58 If Investing, Which AI Companies or Categories Should You Target? 35:50 Platform Migration Doesn’t Necessarily Disrupt All Giants — Lessons from Mobile Internet 36:57 After Software Gets Easier to Build, Why Distribution Becomes More Critical 37:29 GPT Wrapper Isn’t Enough — The Real Value Lies in Harnessing 38:40 Browser Analogy: Thin Product Layers, But Distribution and Defaults Matter 39:28 How Google, Meta, Apple Use Distribution to Drive AI Adoption 40:25 Apple Intelligence Vision: Why Personal AI Assistants Are Hard to Build

Anti-AI Sentiment and Social Backlash

41:39 Is Anti-AI Sentiment Growing? 42:01 Data Centers, Electricity, Water — Are Concerns Overstated? 43:22 Employment Data Remains Unclear — We Lack Real-World AI Usage Metrics 44:32 AI Slop, Creator Anxiety, and Cultural Wars 45:20 Similar to Social Media Backlash: Some Worries Are Real, Others Half-True

Children, Careers, and Technological Risks

46:00 How Should We Educate Children in the Age of AI? 46:22 If Your Child Is Entering the Job Market Soon, It’s Especially Worrying 47:32 “It’ll Probably Be Fine” — But Not Risk-Free 48:26 Deepfake Nudes, Social Networks, and the Cost of Connecting Bad Actors 49:05 UK Post Office Scandal: How Technology Can Unintentionally Destroy Lives 50:16 Which Jobs to Avoid, Which to Pursue 50:42 Skills, Interests, and What People Are Willing to Pay For

The Real Questions We Should Be Asking

51:02 What We’re Still Not Asking Enough About AI 51:12 Do Model Labs Have Pricing Power? 51:42 What Is a Task? What Is a Job? 52:15 From CD to Spotify: Not Doing More of the Old — But Redefining the Problem

53:17 Why was coding the first thing AI changed

53:45 Don’t mechanically calculate “what percentage of a profession can be automated”

54:42 Uber’s test: It’s hard to predict in advance which industries will be affected

55:30 Airbnb vs. hotels: Every industry gets more complex the deeper you go

Actionable advice for individuals

56:42 How should ordinary people respond to fundamental uncertainty?

57:08 Don’t bury your head in the sand, nor just chase moral superiority

57:45 Dive into using AI: Understand what it can do for you

58:10 Make yourself more employable

AI Corner: How Benedict Uses AI

58:25 What Benedict uses AI for most often

58:50 Precise information retrieval remains AI’s weakness

59:25 Using AI for proofreading, image generation, and interior design

01:00:12 Chatbots are blank screens — real value lies in specific use cases

01:00:40 AI will disappear into products: Is voice-to-text still AI?

01:00:56 Why Apple Notes’ voice-to-text is already good enough

Quick Q&A

01:01:36 Benedict’s newsletter, presentations, and “non-actionable” wisdom

01:01:44 Recommended books: *Three Men in a Boat* and Chicago economic history

01:03:00 Recommended movies: Watch those classics you’ve always felt you “should have seen”

01:03:30 Recently liked product: A pair of shoes endorsed by a CEO

01:04:36 Life motto: It depends; things will probably be fine

01:05:01 Old phone collection: Hardware innovations before iPhone

01:06:58 How to find Benedict: benevans.com and his newsletter

🌟 Highlights

💡 AI is big, but don’t mythologize it

Benedict’s core argument is that AI’s importance can be compared to the internet or mobile internet — but it shouldn’t be imagined as a magic wand that will end all old worlds tomorrow. He believes we’re at a stage similar to 1997’s internet era: the technology is profoundly significant, but most products, business models, and value distribution mechanisms haven’t yet emerged.

“My most controversial view is that I believe AI’s significance is as large as the internet or mobile internet — and no larger.”

🧩 Task ≠ Job

One of the key analytical frameworks discussed in the episode is distinguishing between “task” and “job.” AI may automate certain tasks — like writing code, creating slides, or generating summaries — but that doesn’t mean it automates entire jobs. McKinsey doesn’t sell 75-page PPTs; they sell understanding organizational politics, client needs, implementation barriers, and business judgment.

“You hire them not to get a 75-page slide deck — you pay Bain to walk through your entire company and figure out: why didn’t you do this before?”

⚙️ Automation doesn’t necessarily reduce jobs — it might expand demand

Using historical examples like accounting, Excel, and software development, Benedict illustrates that when something becomes cheaper, companies don’t necessarily do less with less money — they might do more with the same money, or even spend more because ROI has shifted. This explains why the number of accountants has continued to grow despite spreadsheets, ERP systems, and cloud computing.

“If you make something cheaper, what happens? Do you do the same thing with less money, or do you do more with the same money?”

🏭 Model companies may resemble clouds rather than Windows

Regarding how value is captured in the AI industry, Benedict raises a critical question: Do foundational model companies actually have pricing power? If multiple models become functionally equivalent and compete fiercely, while real user experience and business logic reside at the application layer, then models may become infrastructure — like cloud services, telecom networks, or electricity — with value flowing upward to higher-level products and distribution channels.

“If chatbots aren’t the final UX, if apps are still needed, and model companies won’t build those apps anyway, and models themselves are essentially commodities — why would model companies have pricing power?”

📣 Distribution will become more important

As foundational capabilities become increasingly commoditized, distribution, branding, and default entry points will become key competitive advantages. Google can embed Gemini into search and Android; Meta can integrate AI across all social products; Apple owns a billion-device ecosystem. For average users, as long as a product is “good enough,” they won’t actively switch.

“When the field becomes largely commoditized, a product that’s merely adequate, combined with distribution and branding, becomes extremely valuable.”

😰 Anti-AI sentiment is a complex mix of issues

Benedict argues that anti-AI sentiment isn’t caused by one single factor, but rather stems from employment anxiety, data center energy consumption, electricity costs, creator rights, AI slop, societal panic, and technical misunderstandings. Like the backlash against social media, some concerns are valid, some are half-true, and others simply don’t hold up.

“It’s a messy, amorphous bundle. Yes, AI will change many things — and we should worry about these changes. But this is actually normal. We’ve always navigated such transitions.”

🧠 The biggest problem in AGI discussions: We don’t know

On AGI and superintelligence, Benedict takes a cautious stance. He points out that we lack a complete theory of human intelligence, and also lack a theoretical understanding of why large models work so well or how much they can improve in the future. Thus, many predictions are essentially intuitive guesses. But this doesn’t diminish AI’s status as an incredibly important technology.

“Even if models stop getting stronger tomorrow — even if this is the endpoint — it’s still an extraordinarily useful technology that will reshape the world over the next decade.”

🛠️ What ordinary people should do most: Dive in and use it

Facing the uncertainty brought by AI, Benedict’s advice is neither panic nor moral rejection — but rather deep, hands-on engagement. You need to understand what AI can and cannot do for you, how it will transform your industry, and how you can become more valuable in the new environment.

“Don’t bury your head in the sand and say you hate all of this. The real help comes when you fully dive in, immerse yourself, and emerge knowing exactly what you can do with it.”

🌐 Podcast Info Supplement

This podcast uses original human voice recordings for audio production, which may sound slightly off in places.

AI translation was used, so some parts may feel awkward or unnatural.

If you’d like to hear Chinese versions of other foreign podcasts in the future, feel free to contact us via WeChat: iEvenight