The Real AI Future: Not Doom, Not Hype - Just Smart Reality | Ben Evans | The Knowledge Project | Podcast Notes | YouTube Summary

AI Isn't the End of the World (But It's Not Nothing Either)

• Why every platform shift feels confusing in real-time (and AI is no different)

• What past disruptions teach us about winners, losers, and unexpected outcomes

• Why the "data moat" advantage isn't what you think

• The real threats to Google, Apple, Microsoft, and Meta

• Why most people still don't "get" AI - and what that means

• The questions we should be asking (and the ones we're getting wrong)

The Centrist Position: Why AI Isn't Everything (But Still Matters)

• AI is the biggest thing since the iPhone - significant, but not civilization-ending

• It's not electricity, not the industrial revolution, not a path to superintelligence

• It's another platform shift that will dominate for 10-15 years, then something else will come along

• Every generation thinks "this time is different" - and they're both right and wrong

• The dotcom bubble was different from the 1980s financial bubble, but still a bubble

• AI will create new jobs, destroy others, and raise weird new questions - just like previous shifts did

History Repeats: What Past Platform Shifts Teach Us

The Internet (1995): Nobody Knew What Would Win

• Would it be centralized "information superhighways" controlled by cable companies?

• Would email be bigger than the web? (Mary Meeker thought so in 1995)

• Would browsers matter? (Microsoft dominated browsers but captured zero value)

• Search advertising and social media came 5-10 years later - nobody predicted that

• You can be certain something big is happening without knowing how it will unfold

• The obvious winners often aren't

• Value capture happens in unexpected places

Mobile Internet: The Small Mac Revolution

• Phones would become small computers, not just "phones with better UI"

• It would take 10 years to really take off

• Telecom companies would capture zero value

• Microsoft and Nokia would become irrelevant

• Mobile would replace the PC as the center of tech, not complement it

• People kept asking: "What would you do on mobile that you can't do on PC?"

• The answer turned out to be: everything

• We now don't even say "mobile internet" anymore - it's just the internet

The Elevator Story: How Innovation Becomes Invisible

• Otis created the "autotronic" elevator with "electronic politeness" (the infrared door sensor)

• It seemed weird and radical at the time

• Now we don't even think about it - it's just a lift

• Today's strange new technology becomes tomorrow's invisible infrastructure

• We forget how weird previous innovations seemed

• AI will eventually just be "software"

The Incumbent's Dilemma: Why Leaders Often Lose

The Pattern That Keeps Repeating

• Try to make the new thing just a "feature" of their existing product

• Use it to automate what they're already doing

• Protect their high-margin legacy business

• New companies unbundle the incumbent's offerings

• The new technology enables things that weren't possible before

• Sometimes incumbents adapt (Google), sometimes they don't (Kodak)

The Kodak Case Study: More Complex Than You Think

• 1975: Their "digital camera" was the size of a refrigerator - not viable

• Late 1990s: Kodak went all-in on digital, became the #1 digital camera seller in the US

• They thought people would print more photos - they invested heavily in photo printers

1. Smartphones - cameras became free with your phone

2. Social media - people stopped printing photos entirely

3. Commodity hell - digital cameras had low margins and no differentiation, unlike high-margin film

• Is Google's high-margin search business like Kodak's film?

• Is AI a low-margin commodity business?

• We don't actually know yet - the margins keep shifting

The Google Question: Reset or Reinforcement?

Why This Moment Is Dangerous for Google

• People reconsider their defaults during discontinuities

• It's no longer automatic that you "just Google it"

• Everyone's trying new things, forming new habits

• Google's advantages matter less when everyone's starting fresh

• Still the best traditional search engine by a wide margin (per the antitrust trial)

• Massive resources and talent

• Strong models (Gemini is competitive)

• But do they have the right org structure and incentives to win the new game?

The Data Advantage Myth

• LLMs need such enormous amounts of generalized text that no one company has enough

• The data everyone needs is roughly equally available to everyone

• Meta downloaded torrents of pirated books because they didn't have enough text

• Google's snippets of text aren't the right kind for training

• Anyone with a billion dollars can scrape the web

• Data is a level playing field for foundation models

• Differentiation will come from somewhere else

The Usage Gap: Why Most People Still Don't Get It

The Surprising Numbers

• ~10% use AI tools daily

• ~15-20% use them weekly

• ~20-30% tried it once or twice

• ~20-30% looked and didn't understand the point

• Many people who try it can't think of a reason to return

• Tech insiders live in a bubble where 90% of people they know use AI constantly

• The real world looks completely different

• This is early adoption, but with a twist - it's free and easy to access

The "Faster Than iPhone" Trap

• Smartphones cost $1,000 (or $5,000 for early PCs adjusted for inflation)

• AI is free - you just visit a website

• There are way more people online now than in 2007

• Of course absolute numbers are bigger and faster

• Why do people try it and not come back?

• Why do even regular users only think of something to do once a week?

• What does it mean that it's easy to access but hard to integrate into daily life?

The Product Problem: Why All AI Chatbots Feel the Same

The Commodity Challenge

• Give the same prompt to ChatGPT, Claude, Gemini, Grok, Mistral, DeepSeek

• Do a double-blind test

• Most people couldn't tell which is which

• Way more usage than competitors

• Top of app store rankings

• Gemini bounces between 50-100 in rankings

• Other AI chatbots don't crack the top 100

The Browser Parallel

• Different rendering engines underneath (like different LLMs)

• But the product is identical: input box, output box

• Only innovation in browsers in 25 years: tabs and merging search into the address bar

• Success came from distribution and branding, not product differentiation

• Is AI the same - where brand and defaults matter more than the actual product?

• Or is it more like social media, where Instagram succeeded over Flickr despite both doing "photo sharing"?

No Network Effects (Yet)

• Operating systems: more users → more apps → more users

• Google: more searches → better results → more searches

• Social media: your friends are there → you're there → more friends join

• More users doesn't make the model better

• No self-reinforcing cycle

• "Memory" features create switching costs, not network effects

• You could probably just ask one AI what it knows about you and tell another

• If AI develops true learning from usage

• But we're not there yet

• And we don't know what that would look like

What People Actually Use AI For (And Don't)

The Use Case Matrix

1. People with tasks AI is obviously good at (coding, brainstorming) → heavy users

2. People with tasks AI could help with, but not obviously → occasional users

3. People good at thinking about new tools → find creative uses

4. People not good at thinking about new tools → try once, abandon

• Doesn't write code (no use for code generation)

• Doesn't do brainstorming exercises

• Doesn't need summarization

• Thinks by writing, not by editing AI output

• Has to actively think: "What am I doing that AI could help with?" - that's mental overhead

The Spreadsheet Comparison

• Cost $15,000 (adjusted) for Apple II + VisiCalc + screen

• Showed accountants: change interest rate, all numbers update instantly

• Replaced literally a week of manual work

• Accountants would finish in a week, then play golf for three weeks (didn't want clients to know it was that fast)

• Spreadsheets had obvious, immediate value for specific jobs

• AI requires you to imagine new workflows

• The value isn't as clear or immediate

• Most people don't naturally think "how could a tool change how I work?"

The Salesforce Button Test

• You're in Salesforce, looking at a client

• There's a button: "Draft email reply"

• You click it, it works, you send

• Massive adoption

• Blank screen

• You have to think what to ask

• You have to form new habits

• You have to remember it exists

• Wrapped in existing workflows = adoption

• Standalone chatbot requiring behavior change = limited adoption

The "Thinking by Writing" Problem

Why AI Changes How We Create

• Writes to figure out what he thinks

• Writing is thinking, not just output

• Editing AI output is fundamentally different from writing

• It's like the difference between composing music and adjusting someone else's composition

• "Is this what ChatGPT would have said?"

• If yes, don't publish it

• Not because people can get it from ChatGPT

• But because anyone would have said that - it's not adding value

• AI raises the baseline of what counts as insight

• It makes obvious analysis worthless

• You have to go deeper to add value

The Slope Problem

• ChatGPT's "insight level" is increasing rapidly

• It's probably already at "intern level" in many domains

• Maybe it hits "master's level" next year

• In math, it's already superhuman in some areas

• Eventually, it might surpass most humans in most domains

• When the AI's slope of improvement is steeper than yours, what happens?

• Maybe it passes the average person in 5 years

• Maybe it passes Evans in 4 years

• Maybe it passed some people already

• Does originality require knowing you're being original?

• AlphaGo made "original" moves - but had a scoring system (win/lose)

• Music generation can make "new" music - but how do you know it's good?

• LLMs are trained to minimize variance - originality is penalized

• How would an AI know something is "different but good" versus just different?

The Regulation Trap: Why "AI Regulation" Misses the Point

The Wrong Level of Abstraction

• You don't regulate the technology, you regulate applications

• We regulate cars, but not "internal combustion engines"

• We regulate medical devices, not "software"

• The trade-offs depend entirely on what you're using it for

• All regulation has trade-offs

• To govern is to choose

• You can't pull one lever without something else moving

• If you make it hard to build AI, fewer people will build AI - that's a choice

The California Approach: Treating AI Like Nuclear Weapons

• Tight controls on who can build models

• Restrictions on what you can do with them

• Assumption: this could create bioweapons or kill everyone

• Yes, you reduce hypothetical risks

• But you also slow innovation dramatically

• You make it expensive to start companies

• You push development elsewhere

• The "AI will kill us all" narrative is "childish logical fallacies"

• It's like social media moral panic 2.0

• But even if you disagree, you must acknowledge the trade-off

The Housing Analogy

• If you make it really hard and expensive to build houses...

• Houses will be more expensive

• You can choose that, but you can't then complain about expensive houses

• If you make it really hard to build models and start companies...

• You'll have less innovation and fewer companies

• You can choose that, but you can't then complain about falling behind

• Neither free market (price signals work) nor government provision (like Singapore)

• Broke the free market without replacing it

• Same risk with AI regulation

Advice for Countries: How to Win at AI

What Evans Would Tell a President

• People used to ask: "How do we create another Silicon Valley?"

• The answer was usually: "You can't"

• Sometimes: create funding structures, make it easy to start companies

• Mostly: get out of the way

• Trying to pick winners rarely works

• This is an economist's question, not a tech question

• Where has industrial policy worked? Where hasn't it?

• From a tech perspective: focus on enabling ecosystems, not specific companies

1. Don't make it harder - avoid California's approach of treating it like nuclear weapons

2. Think of it as "more startups" - not picking a specific company to back

3. Remove barriers - funding availability, regulatory overhead, talent mobility

4. Accept trade-offs - tight control means less innovation, period

The EU vs. US Approach

• Treat AI as potentially dangerous

• Regulate heavily upfront

• Protect citizens from hypothetical harms

• Result: harder to build, slower innovation

• "This is social media 2.0"

• Social media was terrible and destructive

• Need to prevent those problems

• Result: still restrictive, but less than EU

• Both approaches slow development

• You're choosing safety/control over speed/innovation

• That's fine - but own the choice

What Students Should Learn (And Why)

The "Learn to Code" Debate

• No, you shouldn't assume you will or won't be a software engineer

• It's like asking "should you learn an instrument?" or "take theater classes?"

• Find out if you want to learn to code

• Don't presume you need to

• But don't presume you don't need to either

• Presume everything will change

• Presume you'll have many careers

• Presume you need to stay curious

• Focus on learning how to think

What "Learning to Think" Actually Means

• Didn't learn history facts

• Learned how to ask the next question

• Learned how to break things apart

• Learned how to read 50-100 books in a week and find what matters

• Learned how to synthesize information

• Learned to ask: "What does this actually mean vs. what it looks like?"

• Learned to evaluate credibility

• English literature

• Philosophy

• Engineering

• They weren't learning to be historians or philosophers - they were learning to think

The US vs. UK Education Philosophy

• If you want a good job: study math, business, engineering

• Practical degrees lead to employment

• Liberal arts seen as less useful

• Focus on immediate job market value

• Study what challenges you intellectually

• Philosophy, history, literature are equally valid

• You're learning how to think, not what to think

• The subject matter is almost secondary

• He's hesitant to think you can only succeed with certain degrees

• Goldman Sachs, McKinsey, law firms don't require specific majors

• 20 years to figure out what he was good at

• Students can't know yet what they'll be good at

• Try different things

• Find what you're good at

• Create options for yourself

• Learn what challenges and pushes you

• Develop the ability to learn in different ways

The "I Don't Know" Honesty

• "I don't fucking know"

• Sounds like a university commencement speech

• You don't know what you'll be good at

• Try to create options

• Find the skills that match how your brain works

• Most advice is overly prescriptive

• The honest answer is: it depends on you

• Different people need different paths

• The goal is discovering your strengths, not following a formula

Lessons from Venture Capital: What Evans Learned at a16z

The Maxims and Sayings

• How startups actually work

• How the "machine" of Silicon Valley creates companies

• What makes something a good or bad idea

• If it could work, what would it be?

• Could these people make it work?

• "That's a dumb idea"

• Could it work?

• If it did work, what would it become?

• Are these the people who could make it happen?

The Art Gallery Effect: Calibration

• Go to MoMA, the Met, the Louvre

• Masterpieces everywhere

• Hard to distinguish the truly great from the very good

• Wallace Collection in London

• Old aristocratic palace in Rome

• 10-15 rooms of paintings

• Then you see the Raphael - it glows across the room

• "Oh, THAT'S why he's Raphael"

• See hundreds of pitches

• "Oh, that's why he's Max Levchin"

• "Oh no, that's why this is bullshit"

• 10 minutes in, you know - but you have 45 more minutes to be polite

• Pattern recognition comes from contrast and texture

• What good looks like

• What worked vs. what didn't

• What people tend to say

• How things tend to work

• Pure pattern recognition

The High School Dynamics

• Everyone's working on the same thing

• Like a town with one subject

• Everyone around you is doing a PhD

• Of course you're going to do great work - everyone is

• World experts are down the street

• Surrounded by people who've done it before

• Want a CTO who's done it 5 times? They're here

• Want a head of growth who's scaled companies? They're here

• Powerful peer effects and expectations

• Resources and expertise everywhere

• You'll never meet anyone NOT working on exactly this

• No external context or perspective

• No one interested in anything else

• Want to see theater? Go to LA

• Want to see art? Go to LA or Chicago

• Intellectual monoculture

• Powerful for focus and execution

• Dangerous for perspective and judgment

• Easy to lose touch with how normal people think

The Current State of Play: Who's Winning?

The Great Capex Surge

• Google, Microsoft, AWS, Meta spent ~$220 billion in capex last year

• This year: probably over $300 billion

• Nearly tripled in just a couple of years

• Meta bought 49% of Scale.ai for $15 billion

• OpenAI spin-outs (Safe Superintelligence, xAI) valued at tens of billions

• Pre-product, pre-revenue labs

• Just because someone from OpenAI is involved

• Sam Altman complained: Mark offering people $100 million to join

• Going all-in on AI infrastructure

• Massive hiring spree

Model Quality: The Current Rankings

• Clearly firing on all cylinders

• Making great models

• Gemini is competitive

• Llama 4 was apparently an embarrassment

• Scrambling to catch up

• Open source strategy: make models commodity infrastructure

• Still sets the agenda, but less than 2 years ago

• Sam Altman is a "polarizing figure" (actually: opinions are unanimous and negative)

• Everyone who's worked with him has quit

• Weird, contentious relationship with Microsoft

• Own models aren't very good

• Hired Mustafa Suleyman, still struggling

• Dependent on OpenAI relationship

• But: will sell tons of Azure to run everyone else's stuff

• "Will China catch up?" - answer was always obviously yes

• DeepSeek demonstrated this

• Models are becoming commodities

The Apple Question: Different Game Entirely

• Don't need to be first

• Want to do it right

• Don't need every consumer internet thing

• No YouTube competitor, no ride-sharing, no grocery delivery

• Also: no chatbot (yet)

• Does integrating LLMs change the smartphone experience fundamentally?

• Could it shift competitive balance with Pixel?

• (Pixel only bought by Google employees and tech press - Google doesn't want to compete with Samsung)

• 2000s: Everyone needed internet

• To get internet, you needed a computer

• Everyone bought Windows PCs

• But they used them for web stuff, not Microsoft stuff

• Microsoft lost despite selling the hardware

• You'll still buy the new iPhone (best battery, chip, screen, camera)

• But everything you do will be someone else's cloud model

• Not an app from the App Store - a model running elsewhere

• Apple becomes a beautiful hardware shell for others' AI

• You already use iPhone for ChatGPT, DoorDash, Uber, Instagram, TikTok, games

• It's always been a platform for others' services

• As long as you're buying the iPhone, Apple wins

• Plus: if we move to AR glasses, Apple will make those too

The Google Revenue Question

• Search activity moves to LLMs

• Where does the revenue go?

• How do you map search behavior to LLM behavior?

• Do people just shift habits - using ChatGPT as "the new Google"?

• Google sends traffic to websites

• LLMs just answer questions

• Publishers lose traffic

• How does the ecosystem survive?

• They're not out of the game

• Could absorb this shift

• Instagram changed what advertising looks like - Google could too

• But it requires execution and adaptation

The Meta and Amazon Strategy: Make Models Commodity

• Make LLMs open source

• Drive models to commodity status

• Sold at cost

• They differentiate on top with Facebook/Instagram social stuff

• The model is just infrastructure

• Also wants models to be commodity infrastructure sold at cost

• That's what AWS is - commodity infrastructure done better than anyone

• Make money from doing it at scale

• AWS + ads = basically all of Amazon's profit ($50-60 billion in ad revenue alone)

• Neither company needs to own the model

• Both have other ways to capture value

• Commodity infrastructure benefits their core business

Microsoft: Grabbing God's Coattails

• First: VR/AR with HoloLens - "There it is!" (We don't talk about that anymore)

• Now: AI - grabbing on again

• Own models not ranking well

• Hired Mustafa Suleyman, still struggling

• Weird, contentious relationship with OpenAI

• Not really their models

• Will sell enormous amounts of Azure

• Everyone needs cloud infrastructure to run this stuff

• The tension: Do people use ChatGPT directly, or do great products get built on Azure?

• Someone builds amazing accounting software

• Connects to your bank, does the cool stuff

• Runs on Azure, uses some LLM (who cares which one)

• It's just better

• Microsoft wins without owning the model

• "Do my fucking invoicing for me"

• Better yet: "Figure out why that client's ERP doesn't like my bank account"

• "Stop me bouncing emails with someone in India for 3 months"

• LLMs can't do that yet

• When they can, that's the killer app

The Incumbent Analysis: Who's Positioned Best?

Google and Microsoft: The Disrupted Disruptors

• Incumbent business potentially disrupted by AI

• But also: cloud business that sells all the new AI stuff

• Search revenue at risk

• But Google Cloud sells AI infrastructure

• Own models are competitive

• Question: Can they navigate the transition?

• Office/Windows less directly threatened

• Azure is perfectly positioned

• Models are weaker

• Question: Can they capture value without owning the model?

Amazon: The Safe Play

• E-commerce not obviously disrupted by AI

• Might even be enhanced (better recommendations, easier shopping)

• AWS perfectly positioned to sell infrastructure

• Ad business ($50-60 billion) continues growing

• How does AI change how people shop on Amazon?

• LLM recommendations instead of search?

• But Amazon controls that experience

Meta: The Wild Card

• No cloud business to sell AI infrastructure

• Llama 4 was disappointing

• Playing catch-up on model quality

• New ways to monetize through AI

• Instagram well-positioned for AI-enhanced advertising

• Open source strategy could pay off

• Massive resources to invest

• Need better models

• Can't just rely on making models commodity if yours aren't competitive

Apple: The Hardware King

• Still going to sell the nicest glowing rectangle

• Best chip team in the world

• Best camera, screen, battery

• Even if all the AI is someone else's, you need great hardware to run it

• If we move to glasses/AR, Apple will make those too

• Becomes like Microsoft in the 2000s

• Everyone buys the hardware

• But uses it entirely for others' services

• Captures less value from the ecosystem

• This has always been somewhat true (Instagram, Uber, etc. aren't Apple services)

• As long as people keep buying iPhones, Apple wins

• The 30% App Store cut helps

• Hardware margins are healthy

• Does AI change the fundamental value proposition of the smartphone?

• Or is it just another set of apps/services on the platform?

• Too early to tell

Tesla: The Perpetual Question Mark

• It's a software/AI company

• Autonomous driving will create winner-take-all effects

• Camera-only approach will work eventually

• All that driving data creates a moat

• It's a car company

• Competing with entire Chinese industrial policy

• Flood of equally-good EVs coming

• Protected by tariffs in US, vulnerable everywhere else

• Will Tesla get cameras working before Waymo can remove the LIDAR?

• Waymo works now (with $50k of LIDAR per car)

• Tesla doesn't work yet (with cameras only)

• We've been asking this for a decade

• People said Tesla was "the iPhone of cars"

• Actually: cars are becoming Android with no iPhone

• Tesla is just another Android phone maker

• Competing with dozens of Chinese manufacturers

• Half a dozen existing Model cars

• With test drivers

• Doing geofenced drives

• Everyone else was doing this 10 years ago

• Is this the breakthrough, or more of the same?

The Philosophical Questions: What We Still Don't Know

Can AI Be Truly Original?

• Made "original" moves no human had tried

• But: had an external scoring system

• Every move has a score (closer to winning)

• Feedback loop tells it the move was good

• Infinite monkeys would eventually type Shakespeare

• But there's no feedback loop

• No way to know which output is the masterpiece

• The Borges infinite library contains masterpieces - but which ones?

• Variance is bad

• Originality gets a lower score

• Trained to match patterns, not break them

• How would it know something is "different but good" vs. just different?

• Easy to generate more stuff that sounds like Pink Floyd

• Easy to make more Grateful Dead

• ("What do Grateful Dead fans say when they run out of drugs? 'This music's terrible'")

• But how would AI know people are fed up with 70s prog rock and want punk?

• How would it know Christian Dior's "New Look" would express post-war desire for luxury?

• Is knowing something is "original and good" just pattern matching at a longer frequency?

• If you zoom out enough, is it still just following patterns?

• Does it matter if it's "really" reasoning or just right 99.9999% of the time?

• Is this even the right question to ask?

The Boutique Renaissance

• Sells only one book

• Changes monthly

• You don't choose - they choose for you

• But you have to know it exists

• Amazon: has everything, but how do you choose?

• Can't just say "what's a good lamp?" when there are 10,000 lamps

• The boutique: curated, individual, unique

• But requires discovery

• LLMs might suggest the unique individual thing

• But would LLMs also create the unique individual thing?

• The more LLMs do what everyone would do...

• The more valuable the truly unique becomes

• Scott Galloway does different stuff than Benedict Evans

• Mary Meeker does different stuff

• Some is about who you are, your story, authenticity

• Some is about saying interesting stuff regardless of who you are

• Some is recommendation algorithms

• All have value in different contexts

The Department Store Parallel

• Creates department stores through force of will

• Invents fixed prices (so you can have discounts and loss leaders)

• Invents mail order

• Invents advertising

• Puts slow-moving expensive stuff upstairs

• Puts food and makeup downstairs (impulse buys)

• "Have you seen what that maniac's doing?"

• "Selling hats and gloves in the same shop!"

• "He has no morals!"

• "He'll be selling fish next!"

• Nothing new under the sun

• People have freaked out about mass-produced products before

• People have freaked out about "too much content" before

• Erasmus was supposedly the last person to read every book

• "Too much AI slop" - but how many books were published in 1980? Did everyone read them all?

• Every generation thinks their disruption is unique

• It is unique - but it's also familiar

• Just different scales and contexts

The Practical Reality: Why Most People Don't Use AI

The Weekly User Problem

• Someone tries ChatGPT

• They "get it" - they see the value

• But they only come back once a week

• Why can't they think of more to do with it?

• You have to actively think: "What could AI help me with?"

• Most people don't naturally think that way

• It's cognitive overhead

• Easier to just do things the way you've always done them

• New tools require new habits

• Habits are hard to form

• Especially when the tool is a blank screen waiting for you to think of something

• Compare to: phone buzzes, you check it (easy habit)

The Use Case Matrix

• Coding assistance

• Brainstorming

• Summarization

• Image generation

• These people are heavy users

• Requires imagination

• Requires understanding what AI can do

• Requires trial and error

• These people are occasional users

• Naturally curious about workflows

• Constantly optimizing

• Find creative applications

• These people find uses even without obvious tasks

• Just want to get work done

• Don't naturally think "how could I do this differently?"

• Try once, don't see the point, abandon

• This is most people

• You need to be in multiple quadrants to be a heavy user

• Most people are only in one or two

• This limits adoption

The "Roughly Right" Problem

• Conference organizers used ChatGPT to write his bio

• Didn't tell him

• Everything was the "right kind" of thing

• Right kind of degree, university, experience, jobs

• Just not actually right

• But for him: completely useful - spent 30 seconds fixing it instead of an hour writing from scratch

• For them (needing accurate bio): useless

• For him (needing a starting point): very useful

• "Right or wrong" depends on why you wanted it

• Evans thinks AI has "zero value" for quantitative work today

• Because the numbers need to be actually right

• Not "roughly right"

• You don't want π to be 3.1

• (Though it depends how big the circle is...)

• Not wrong once in a billion years

• Wrong a dozen times per page

• Can't just output it and give to someone

• Requires checking everything

• Often easier to just do it yourself

Why Evans Doesn't Use AI Much

• Doesn't write code (no use for code generation)

• Doesn't do brainstorming exercises

• Doesn't need summarization

• Doesn't create images

• Thinks by writing, not by editing

• "Is this what ChatGPT would have said?"

• If yes, don't publish

• Not because people can get it from ChatGPT

• But because it means he's not adding value

• Anyone would have said that

• Works at consultancy

• Needs pencil sketches of concepts

• Uses Midjourney to generate them

• Perfect use case

• Doesn't matter if the person in back has three legs (doesn't anymore anyway)

• Could Photoshop it out if needed

• Things AI is good at: not what Evans does

• Things Evans does: AI not yet very good at

• Things where "roughly right" would help: he doesn't do those things

• This is true for millions of people

• The use cases don't map to their actual work

• Or the quality isn't there yet

• Or both

The Future Scenarios: What Might Happen

The Search and Discovery Revolution

• Google sends you to websites

• Websites have ads or sell things

• Publishers get traffic

• Ecosystem works (sort of)

• You ask: "What mattress should I buy?"

• LLM just tells you

• No website visit

• No publisher revenue

• No ads (or different ads)

• What is SEO for LLMs?

• How do products get discovered?

• How do publishers survive?

• Where does advertising happen?

• Who captures the value?

• Infinite retail options

• Infinite media options

• Infinite everything

• How do you choose?

• LLMs could solve this

• But they could also create new gatekeepers

The Differentiation Question

• All LLM chatbots feel the same

• Input box, output box

• Different colors and icons

• But fundamentally identical experience

• Browsers have been commodities for 25 years

• Same rendering, same basic UI

• Only innovation: tabs and merged search/address bar

• Winner was about distribution and defaults, not product

• Photo sharing is a commodity

• But Instagram beat Flickr decisively

• Product and experience mattered

• Network effects mattered

• Which model does AI follow?

• Browser-like (distribution and brand matter most)?

• Or social-like (product and network effects matter)?

• We don't know yet

The Memory and Switching Costs

• AI remembers your previous conversations

• Knows your preferences

• Builds context over time

• Probably not

• More like a switching cost

• But you could probably ask one AI what it knows about you

• Then tell another AI

• So maybe not even a strong switching cost

• The model getting better because more people use it

• Self-reinforcing cycle

• More users → better product → more users

• We don't see this yet

• Might never see it

• Or might see it in ways we can't predict yet

The Integration vs. Standalone Question

• People go to ChatGPT/Claude/Gemini

• Like they go to Google

• Becomes the new default

• Brand and distribution matter most

• AI wrapped into existing products

• Salesforce button: "Draft reply"

• Photoshop: "Remove this object"

• Excel: "Analyze this data"

• People never see the underlying model

• Integration will drive most adoption

• Because it removes the mental overhead

• You don't have to think "what should I ask AI?"

• The button is just there when you need it

• Model providers might not capture much value

• Application layer captures value

• Like cloud infrastructure: valuable but commoditized

• Or like the browser: necessary but not where the money is

The 10-Year View: What Comes Next

The Platform Shift Timeline

• Confusion about what matters

• Dozens of competitors

• Unclear business models

• Lots of experimentation

• Early leader might not be final winner (MySpace effect)

• Consolidation begins

• Clearer use cases emerge

• Business models solidify

• Some companies break out

• Others fade away

• Dominant players established

• Network effects solidified (if they exist)

• Becomes "just software"

• Next platform shift starts emerging

• Mature market

• Incremental improvements

• Everyone's moved on to talking about the next thing

• AI is just part of the infrastructure

• This has happened with PCs, internet, mobile

• No reason to think AI is different

• Except in the specific ways it's different

• Which we won't fully understand until later

The Employment Impact

• Won't destroy all jobs

• Won't create mass unemployment

• Impact similar to previous platform shifts

• Will change what jobs exist

• Will change how work gets done

• Will create new categories

• Will eliminate some roles

• Accountants thought it would destroy their jobs

• Instead: could do more work, more complex work

• Some jobs changed, some disappeared

• But accounting as a profession grew

• Some jobs automated

• New jobs created

• Most jobs transformed

• Net effect: probably positive, but with disruption

• Winners and losers, like always

What Comes After AI?

• There will be something else

• We'll all be talking about that

• AI will be "just software"

• Like we don't say "mobile internet" anymore

• Like automatic elevators are just elevators

• Until the 1950s: manual operators

• Otis creates "autotronic" with "electronic politeness"

• Revolutionary at the time

• Now: just a lift

• Nobody thinks about it

• In 2035, nobody will say "AI-powered"

• It'll just be software

• The next generation won't remember when it wasn't there

• We'll be worried about whatever comes next

• This feels revolutionary because we're living through it

• Everything feels revolutionary when it's happening

• Then it becomes normal

• Then something else comes along

The Final Takeaway: Measured Optimism

What Evans Gets Right

• Not "AI will kill us all"

• Not "AI changes nothing"

• Not "this is the singularity"

• Not "this is just hype"

• Biggest thing since the iPhone

• Will reshape industries

• Will create new winners and losers

• Will raise new questions

• Then will become normal

• Helps cut through the noise

• Focuses on actual impact

• Acknowledges uncertainty

• Avoids both panic and complacency

The Questions That Matter

• Where does search revenue go?

• How do publishers survive?

• What are the real use cases?

• Why don't more people use it regularly?

• Can error rates be controlled?

• Where is value captured?

• Do network effects emerge?

• How does this change work?

• What new categories get created?

• Who are the winners and losers?

• Does this change computing fundamentally?

• What comes after smartphones?

• How does this reshape industries?

• What's the next platform shift?

The Honest Uncertainty

• "I don't fucking know"

• "We don't have answers yet"

• "It depends"

• "We'll see"

• Most people pretend to know

• Certainty sells better than uncertainty

• But honesty is more useful

• Helps avoid bad decisions based on false confidence

• We didn't know how the internet would work

• We didn't know mobile would replace desktop

• We didn't know social media would matter

• We figured it out as we went

• Same will happen with AI

For the Rest of Us

• Focus on real problems

• Don't assume AI solves everything

• Don't assume it solves nothing

• Find where it actually adds value

• Be prepared to pivot

• Understand the trade-offs

• Don't bet on certainty

• Diversify across scenarios

• Remember: early leaders often don't win

• Value capture might be in unexpected places

• Stay curious

• Learn how to learn

• Don't assume your job is safe

• Don't assume your job is doomed

• Develop skills that complement AI

• Don't ignore this

• Don't panic about this

• Experiment thoughtfully

• Focus on real use cases

• Accept that you'll get some things wrong

The Bottom Line

• Not hype

• Not nothing

• Genuinely transformative

• Has limitations

• Has trade-offs

• Will take time to play out

• Will surprise us in unexpected ways

• Platform shifts are always confusing

• We figure them out as we go

• Some things we expect don't happen

• Some things we don't expect do happen

• Eventually it becomes normal

• Stay curious

• Stay skeptical

• Stay flexible

• Don't believe the extremes

• Focus on what's actually happening, not what people say is happening

• In 10 years, we'll be talking about something else

• AI will just be software

• This too shall pass

• And that's okay

Key Quotes to Remember

"This is like the biggest thing since the iPhone, but I also think it's only the biggest thing since the iPhone."

"Why is it that somebody looks at this and gets it and goes back every week, but only every week?"

"Is this what ChatGPT would have said? If yes, don't publish it."

"If you make it really hard and expensive to build houses, houses will be more expensive. You can choose that, but you can't then complain."

"I don't fucking know. You don't know what you're going to be good at. Try to create options for yourself."

"In 10 years time it'll just be software."