
Why Everyone Is Wrong About AI (Including You) | Benedict Evans
What this covers
This conversation between Shane Parrish and Benedict Evans is a wide-ranging interview built around how to think about platform shifts. Evans uses historical cases to calibrate expectations: the Otis automatic elevator that became just "an elevator," a 1995 vision of cyberspace that left open whether the internet would be centralized or permissionless, and the common misreading of Kodak, which he says lost to margin compression and smartphones rather than by ignoring digital. He also talks about how incumbents tend to treat new technology as a feature of the existing business instead of cannibalizing it.
The AI-specific material is where the positions get sharper. Evans doubts that data volume is a moat for LLM) builders, since everyone needs roughly the same public text. Yet ChatGPT still dominates consumer use, which he treats as a brand and distribution effect, and he uses MySpace to argue that such leads can be fragile. He cites survey figures: about 10% of people use these tools daily, while a sizable share tried them once and saw no value. On the competitive map he is contested but clear, arguing that Amazon is best placed because neither its retail business nor AWS) is directly threatened, whereas Google's risk is a moment when users reset their defaults).
The remaining ground is more philosophical and political. Evans asks whether an LLM must "understand" a problem if its answers are reliably right, and argues that originality needs a feedback loop of external evaluation, as AlphaGo had and text models lack. He calls the story of an AI blackmailing researchers overstated, and he argues AI regulation should target specific uses rather than "AI" as a whole. He also covers Soviet planning and price signals, the view that industrial policy works best by removing barriers, Silicon Valley as a machine for creating startups, Tesla's camera-only bet, and his advice to students to learn how to learn rather than a single technical skill.
Benedict Evans argues that AI is a significant platform shift comparable to the iPhone and internet, but not fundamentally different from past technology transitions—the key uncertainties are not whether it matters, but how value capture will work, whether network effects will emerge, and how incumbents will navigate the disruption.
- Past platform shifts (internet, mobile, PC) were also unpredictable in their specific outcomes despite being obviously important, showing that 'this time is different' is always true but doesn't prevent disruption
- Current LLM commoditization and lack of apparent network effects mean ChatGPT's dominance is fragile and could be displaced by superior products or modalities, unlike past winner-take-all dynamics
- The real question is not whether AI is important but where value concentrates: in models, applications, cloud infrastructure, or device ecosystems—a question that can only be answered by watching how users actually adopt and integrate these tools
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The 1995 depiction of 'cyberspace' by a research firm was deliberately ambiguous about whether the internet would be centralized (controlled by cable/phone/media companies as 'information superhighway') or decentralized and permissionless, reflecting genuine uncertainty about which architecture would prevail.
“I have a diagram I use a lot from, um, 1995. This research firm made a diagram of something called they called cyberspace. Um, because it wasn't clear it was just going to be the internet. It was clear that everyone was going to have some kind of computer thing connected to some kind of network. But remember the phrase information superighway...which sort of conveys that it would be centralized and controlled by cable companies and phone companies and media companies, which is sort of how everything had always previously worked. It wasn't clear, no, it was going to be the internet. It wasn't clear the internet was going to be kind of radically decentralized and permissionless”
Regulation always involves trade-offs between competing values (growth vs. job protection, innovation vs. safety), and policymakers often fail to acknowledge that restricting one thing necessarily has consequences elsewhere—the classic mistake is implementing policies (like housing restrictions) that break price signaling without creating an alternative allocation mechanism.
“When you regulate stuff, there are trade-offs. Like, you learn about this in your first year in economics class. Like, regulation has costs and consequences, and it's not necessarily um you know, there's there's always a trade-off.”
The Otis Autotronic elevator, which introduced 'electronic politeness' (infrared door sensors) in the 1950s, is an example of technological innovation that was genuinely revolutionary at the time but has been completely absorbed into the invisible background of how elevators work—we don't call them 'automatic elevators with electronic politeness' anymore, we just call them elevators.
“There's something I love talking about which is is um the rise of automatic elevators. So until the 50s elevators were manually operated...in the 50s, Otus creates the autotronic, I think it's called the autotronic elevator, which had electronic politeness, which basically meant the infrared thing that stops the door closing.”
Learning how to learn and learning how to think are more important than learning specific technical skills (like coding) for long-term success, because careers and job responsibilities will change, and the ability to adapt to new tools and domains matters more than expertise in any single domain.
“I think you should presume that you will need to be curious and that you'll have many careers and different kinds of jobs. I think you should be focusing on learning how to think”
MySpace demonstrates that early network effect dominance can be fragile and that late entrants can displace leaders if they offer a superior product/experience—MySpace was easy to exit to Facebook, suggesting that even if LLMs develop network effects, subsequent innovations (new modalities like AR/glasses) could displace current leaders.
“It was very easy actually to get people to switch back and forth between MySpace and Facebook and and Bibo and and and Friends Reunited and whatever the um um or all these other things in the early days, right? then you kind of get this this separation out. But then of course then you get then Instagram comes along and then Tik Tok comes along right and so as soon as you have something that's a different proposition that turned out to be extremely easy to pull that away”
Unbundling of incumbent services often follows platform shifts—new platforms enable new business models that break apart integrated services, creating opportunities for both new platforms (that bundle differently) and smaller specialized competitors.
“over time you get new stuff you unbundle both the incumbent's intake and you unbundle existing companies because of something that's possible because of this new technology”
Historical platform shifts (internet, mobile, PCs) were each considered uniquely disruptive in real time, with people claiming 'this time is different,' but the actual outcomes were highly unpredictable—people could not foresee that the browser would become the dominant platform, that social media would emerge 5-10 years later, or that the iPhone would replace PCs as the center of tech.
“What everybody's saying this time is different, which everybody does at each platform shift, I would imagine. And what's the same? Well, that's the there's a famous book about um financial bubbles called this time is different because people always say this time is different and it always is.”
As LLM capability increases and approaches expert-level reasoning, there is a philosophical puzzle about whether the output must 'understand' the problem or 'reason' about it to be valuable—if the answer is consistently correct, does it matter whether the system actually understands or is just matching patterns correctly?
“this is at a high level this is some of the AGI argument that if the thing gets infinitely close to reasoning without ever actually reasoning does it matter like at a certain point the thing if the thing is always right if thing if the thing is only wrong once in a billion years does it matter that it's not that it's not always right”
Incumbents facing platform shifts typically attempt to make the new technology a feature of their existing business rather than cannibalizing their existing revenue—they use automation to do more of what they already do rather than fundamentally transforming their business model.
“when any with any of these sort of fundamental technology changes the incumbents always try and make it a feature and they try to absorb it and the same thing outside of technology um existing companies try and absorb it and they use it to automate the stuff they're already doing and then over time you get new stuff you unbundle both the incumbent's intake and you unbundle existing companies”
The Kodak digital camera example is commonly misunderstood—Kodak did invent digital cameras early (1975) but the technology was not viable as a consumer product until the late 1990s, so the company's failure was not from ignoring the technology but from margin compression in digital (low-margin commodity) versus film (high-margin proprietary), combined with smartphones and social media eliminating the printing business.
“it's like the cliche that people say oh Kodak had digital cameras and they didn't get it or they ignored it or they didn't want to do it because it would destroy their business. But then you go and look at it and it's like, well, that was 1975 and the thing they had was the size of a refrigerator. Like, no, that was not a consumer product.”
Data size and ownership are not significant competitive advantages for large tech companies in LLM development because the generality requirement of large language models means all players need approximately the same enormous volume of text data, and that data (Common Crawl, pirated books, etc.) is available to anyone with sufficient capital.
“everyone's kind of using the same data which is you need such an enormous amount of generalized text that the amount that Google has or that meta has is not actually enough to move to be a kind of fundamental difference in what you can train with.”
Regulation of AI should be framed not as regulating AI broadly, but rather regulating specific applications, externalities, and use cases—regulating 'AI' as a monolithic technology is analogous to regulating 'databases' or 'cars' generally, which doesn't make sense and misses the actual policy questions.
“I think regulation of AI is sort of the wrong level of abstraction. Talking about regulating AI as AI is the wrong level of abstraction. And it's like saying we're going to regulate databases or regulate spreadsheets or regulate cars. Well, we do, but not like that.”
The Infinite Monkey Theorem and Borges' Library of Babel illustrate that generating random text is not innovation without a feedback loop—AlphaGo could produce 'original' moves because it had a scoring system (winning), but LLMs generate text without knowing if it's good, so originality requires external evaluation systems the models lack.
“The have feedback. Yeah. It had a feedback loop because every move has a score. It can evaluate the score every move. The classic um parable of the monkeys and typewriters or you know the Borg's infinite library is there's no feedback loop...The monkeys were typewriters would generate new masterpieces but there's no feedback loop. So there's no way of knowing.”
Soviet central planning failed in the 1960s-1980s not because the idea was wrong but because sophisticated modern economies have too much complexity for centralized planning to process without price signals—adding incentive structures without prices cannot solve the information problem.
“Soviet central planning in the '60s '7s '80s and it's about sort of what happens when you have a central planning that just cannot cope with the level of complexity of a sophisticated economy in the '60s and '70s as opposed to let's make grain and tractors and locomotives in the 20s and 30s and steel which already kind of works”
Despite commoditized underlying models, ChatGPT dominates consumer LLM adoption (top of app store, 50-100 ranking for Gemini, others below top 100, similar pattern in Google Trends and revenue), suggesting brand and distribution create powerful network effects despite product parity at the technology level.
“chat GPT is at the top of the app store rank. Gemini bubbles between like 50 and 100. Um none of the others were in the top 100. Same in Google Trends, same in the usage numbers, same in the revenue...it's like chat GPT is now like the brand. It's the default. It's the Google.”
The concept that 'Anthropic's AI threatening to blackmail researchers' is overstated—when the researchers asked a text-generation model to tell a story about what it would do if blackmail were optimal in a certain scenario, the model generated text about blackmail, which proves nothing except that language models can generate stories following human behavioral patterns.
“Enthropic has done a bunch of things where they say like the AI was threatening to blackmail me...And you read the story and you think, okay, you asked what's basically a story generating machine. Please tell me a story of what you would do if in this situation where most people would probably say X and the machine says probably X. And you say, my god, it said it would do X. It's like, well, yeah, it would blackmail...based on human behavior.”
The primary policy lever for technology success is negative (removing barriers) rather than positive (picking winners)—reducing regulations, lowering costs, and creating favorable conditions for startups matter more than direct government investment or championship selection.
“most of the answer is things like getting out of the way...From a technology analyst perspective, I think of this in terms of a what are you doing that would make this harder and b think of this as just more startups.”
Amazon's position is strongest because it has an incumbent business (retail, ads) that is not disrupted by AI and a cloud business (AWS) that will benefit from selling AI infrastructure, unlike Google and Microsoft whose core search/office businesses face direct disruption from AI.
“Amazon has an incumbent business that doesn't get disrupted by this, at least much less obviously, and a cloud business that will be very happy selling all of this stuff”
Silicon Valley functions as a 'machine for creating startups' where the cultural environment (everyone doing the same thing, access to successful founders, availability of capital) creates conditions where starting a company is assumed as normal, not exceptional.
“Silicon Valley is like a machine for creating startups and still too Many people kind of look and say, 'Well, that was a dumb idea.' And it's like, 'Well, that's the wrong question.'”
About 10% of people report daily use of LLMs like ChatGPT, 15-20% report weekly use, 20-30% report monthly use, and 20-30% report trying it once but not getting it or not seeing value, meaning consumer adoption is still narrow despite the hype around weekly active users.
“You look at survey data and I had did this slide in the last presentation I did of like five different surveys from the US from late last year earlier this year and it's all roughly the same. It's like something around 10% give or take three or four percent of people depending on the survey are using this say they're using this every day. Another sort of 15 to 20% of people say they're using it every week...another 20 or 30% of people who say I use it every month or two and another 20 or 30 people% of people who said yeah I had a look I didn't get it”
In a market where commodity infrastructure and baseline products are supplied by LLMs, value will shift toward curation, personalization, and differentiation at the application layer—the analogy is the shift from 'infinite retail' (Amazon has everything) to boutique/niche experiences where the scarcity is curation and discovery.
“the more that the LLM can do what everybody would probably do or say what everyone would probably say, then the more you push to other places.”
There is a matrix of user types: those with obvious LLM-suitable tasks and who recognize it; those with suitable tasks but who don't recognize it; those good at thinking about new tools; those not good at that—most people don't have the mental model to map their work onto AI capabilities without the tool being deeply integrated into existing workflows.
“there's a sort of I said a moment ago, there's like a matrix. There's a matrix of like who has the kinds of tasks that it's good at obviously. Who is good at has the kind of task that it's good at? Not obviously. Who is good at thinking about new tools for the things that they're doing? Who isn't?”
Google's core threat from AI is not that Bing will become a better search engine, but rather a moment of discontinuity where users reset their defaults and reconsider their baseline assumptions about which search tool to use.
“the very high level threat to Google is that you have this moment of discontinuity in which everybody resets their prior and res reconsiders their defaults. And so it's no longer just the default that you go and use Google and for this search or that search like maybe Bing is 10% better on that search.”
Tesla's advantage lies in autonomous driving capability via camera-only systems rather than LiDAR, but it's unproven whether Tesla will achieve this before competitors like Waymo eliminate LiDAR and achieve better safety—this is a long-standing question that remains unresolved.
“Will Tesla get it working with cameras before Whimo can get rid of the LAR? Uh, we could have had that conversation. I literally was on podcasts seven, eight years ago having those conversations. We don't know the answer. Maybe.”
Google, Microsoft, AWS, and Meta spent approximately $220 billion on capex in 2024 and are expected to spend over $300 billion in 2025, roughly doubling to tripling spending from a couple years ago, primarily driven by competition to train and operate large language models.
“last year the Google Google Microsoft AWS and Meta spent about $220 billion of capex last year and we'll probably spend something over 300 this year and so that's basically more than doubled almost tripled I think from a couple of years ago”
There is currently no apparent network effect in LLMs like there is in search engines or social media—more users don't make the product better because models are not retraining continuously with user feedback, so additional usage does not drive model improvement.
“You can't as as I'm sure you know, you're not they're not retraining the models all the time with the data. So you don't have that kind of runaway effect as like continuous flow and more queries produces you know better results.”
Learning and pattern recognition come from repeated exposure to raw experiences and contrasts (not compressed summaries), and the depth of understanding depends on engaging with primary material rather than consuming others' interpretations and conclusions.
“We have an experience. We reflect on that experience and we create a compression and that compression becomes our takeaway...But what we consume most of the time is other people's compressions. So like when people read your newsletter, they're consuming a compression of the work that you've done uh but not the actual raw work. So in a way it's an illusion of knowledge if if you haven't done the work in that area.”
If integrating LLMs into iOS does not fundamentally change the smartphone experience or create new device categories, Apple could face a 'Microsoft-like' scenario where the device remains valuable but the software and services that run on it shift to cloud-based third-party providers, eliminating Apple's historical control over the experience.
“Does it end up kind of getting Microsofted in the sense that you're going to still for the time being you're still going to buy a smartphone...but you're still going to buy a smartphone...You're still going to buy the nice one with a good battery and the fast chip to run the models and the good screen and the best camera, which will still be an iPhone because Apple still has the best chip team...But everything you do on it will be from someone else.”
Predicting novel cultural shifts (like punk music replacing prog rock, or Christian Dior's 'New Look' replacing wartime austerity) requires understanding when people will desire something genuinely different, which LLMs cannot do because they pattern-match on training data and don't have feedback loops for recognizing which novelty will resonate culturally.
“the challenge is knowing that now people are really fed up of 70s Prague rock and they would really like something else and that something else would be punk and that would work. is knowing that people in the '40s were really fed up of the war and would want luxury and that Christian's new look would work and would express that. Could an LLM do that thing? How much variance do you need? I don't know.”
Tesla is often mischaracterized as 'the iPhone of cars' but is actually 'Android with no iPhone'—an AI/software-powered EV in a market where Chinese manufacturers are producing comparable or superior EVs and are protected only by US tariffs from competition.
“the interesting Tesla point is people always looked at it and said, 'It's the iPhone of cars.' No, it's not. What's happening is that cars are becoming Android with no iPhone. And then Tesla is just selling and is is in that metaphor Tesla is just another android phone maker and they're competing with the whole Chinese industrial policy to make more and there's going just going to be a flood outside the US is they're protected by tariffs.”
Most people cannot distinguish between different LLM models (ChatGPT, Claude, Gemini, Mistral, DeepSeek) when given the same prompt in a blind test because the underlying products are essentially identical with only cosmetic differences in branding and UI.
“it seems to me right now you could do like a double blind test of the same prompt given to Grock, Claude, Gemini, Mistral, Deep Seek. I bet most people wouldn't be able to tell which is which.”
Thinking by writing is a discovery process where writers check whether their thoughts are original by asking 'is this what ChatGPT would say?'—if yes, it's not worth publishing because it's not adding insight beyond the baseline that LLMs now provide.
“I would always would ask in the past is kind of your point about pattern recognition is I look at something and say am I adding value here am I saying something useful am I saying something different am I asking the key question...now I can just say is that I would what I would always would ask in the past is kind of your point about pattern recognition is I look at something and say am I adding value here am I saying something useful am I saying something different am I asking the key question am I pushing further am I pushing the am I asking the next question rather than just answering the obvious questions now I can just say is this what chat GPT would have said and if the answer is this is what chat GPT would have said then I didn't publish it”
The question of whether governments should try to replicate Silicon Valley is largely unanswerable because it's an economist's question, not a technology question—industrial policy successes and failures are context-dependent and cannot be easily codified or transferred.
“A long time ago, I used to get these questions about like how can we replicate Silicon Valley? And I always feel like the answers to those questions as much as possible is like you can't...the idea of trying to create national champions is very hard now that almost kind of becomes an economist's question rather than a technology question. How do you create national champions? Where has that worked? Where has that not worked?”
ChatGPT adoption is faster than PC or smartphone adoption because it is free and web-based (no hardware purchase required) and there are more people online now, which makes raw adoption numbers larger but not necessarily more impressive as a percentage of addressable market.
“It's faster than smartphones. Yes. Because you didn't need to buy a $1,000 smartphone...It's faster than PCs. Yes. Cuz you know what PCs cost in the 80s adjusted for inflation? It's like five grand...It's free for a lot of them...It's free. It's a website. You just go there. Of course, it's got faster adoption and there's way more people online as well.”
Apple's historical strategy is to not be first-mover in consumer internet trends but to do it 'right' once the market is established—they don't have YouTube, car sharing, or grocery delivery services, and similarly, Craig Federighi indicated they don't have a chatbot yet.
“Apple's in a slightly different position in that they have always sort of taken the position that they don't want to be first. they want to do it right and that they don't need to be doing whatever the latest consumer internet thing is like they don't have a YouTube I think Craig Fid Federigi said in an interview after WWDC like we don't have a YouTube he didn't call it YouTube like we don't have a YouTube we don't have a car sharing service we don't do grocery delivery we also don't have a chatbot okay”
Calibration—understanding what quality looks like by seeing many examples and contrasts—is a valuable skill developed through experience (looking at many startups, art, investments) that helps distinguish good from mediocre work instinctively.
“I think something else that I learned was um calibration. There's a sort of again another metaphor I always think of which is that if you go to like a really great art gallery...there is a masterpiece. If you go to a smaller weirder art gallery...You see it glowing across the room...it's the same when you see lots and lots of startups like, 'Oh, that's why he's Max Lechin. Oh, no, that's why this is bollocks.'”
OpenAI's reporting of 'weekly active users' is a misleading metric (similar to how earlier internet metrics like 'hits' and 'registered users' were meaningless) because it includes people who use the product once a week, which is not the same as meaningful regular adoption.
“Funny thing about social is people when social happened people would talk about registered users. You know in the early days of the internet people would talk about hits...So that's seven hits. So it's was meaningless and you had to switch to page impressions and then it's registered users and then it was monthly active users and on social people said well hang on if you're using Instagram once a month you're not using it in daily active users or nothing and weekly active users we don't like either. Now OpenAI is doing weekly active users and Sam Wolman was a social media startup founder. He knows this. It's a [ __ ] number.”
The value of LLM output (whether it's 'right enough') depends entirely on the use case—a biography that's structurally correct but factually wrong is useless for the conference organizers who generate it but useful for Evans who edits it, because he gets the structural scaffolding and corrects details.
“I had very early example of this and I was going to speak at an event at the beginning of 2023 and the conference people had asked me for a long biography of myself...It was always the right kind of biography it was the right kind of degree the right kind of university right kind experience, the right kind of jobs. It just wasn't actually right things. But B, I could take that and fix it...For them, it was useless...For me, it was completely very useful as I spend 30 seconds fixing it instead of spending an hour scratching my head”
Different people have different modes of thinking: discursive/associative modes that spiral through ideas, and analytical modes that break things apart and identify 2-3 core structures, and the analytical mode is useful for understanding what something actually is by testing against data and structures.
“I have a mode that's sort of discursive and slightly rambling and free associating and I'll kind of spiral off in different directions and hopefully come back to the point. Um, and then there's a mode where I want to try and pin the thing down and break it apart and say what are the two three four things that are happening here”
Amazon's actual revenue and profits come primarily from AWS (cloud infrastructure) and advertising, not from e-commerce itself—advertising generated $50-60 billion in revenue last year and is often overlooked in discussions of Amazon's business model.
“basically all the money comes from AWS and the ads. People who complain about AWS don't haven't realized the ads make it Amazon did 50 billion $60 billion of ad revenue last year.”
Meta spent $15 billion on Scale AI (acquiring 49%) and appears interested in pre-revenue AI labs like Anthropic's Safe Superintelligence (SSI) and Thinking Machines, both valued at many tens of billions without products or revenue, indicating a strategy to acquire AI talent and capability at any cost.
“meta bought scale half 49% at scale.ai for $15 billion. Apparently looked at at at both of the other recent OpenAI spin out safe super intelligence and what's the other one? Thinking machines which are both basically pre-product pre-revenue labs with a somebody from Open AI at multiple tens of billions valuations.”
Success is defined as living without rockets landing on one's head, children not dying from disease, not worrying about bank closures, and doing interesting work one enjoys that pays the rent—and Evans achieves this by traveling and giving presentations.
“We live in the luckiest time. You know we are not worried about rockets landing on our heads. We're not worried about our children dying from diseases. We're not worried that the bank might be closed tomorrow and all of your money is gone. We're doing something interesting that we enjoy and that pays the rent that we want to be able to pay. I get paid to fly around the world and give slides for money. So, I think I'm doing okay.”
LLM capabilities are improving exponentially in some domains (math, coding) while remaining stagnant in others, and the intersection of LLM capability curve with human expert capability curves will happen at different times for different fields—intern-level capability in some domains might take years to reach master's level in others.
“the chat GBT level of insight...is increasing at a faster pace than most people and eventually those slopes inter and it's probably intercepted or or intersected already with you know maybe up to intern level and next year it might be master's level or it might even far surpass that and it hasn't some domains in terms of math um the year after it might”
Evans' personal use cases for ChatGPT are limited—he doesn't write code, brainstorm, summarize, or generate images—and most practical uses for LLMs (code generation, email drafting, image creation) don't align with his work, making it difficult for him to discover valuable applications that would require actively imagining new workflows.
“I don't write code. I have zero use for something that will write code for me. I don't really do brainstorming. Okay? I don't do summarization of things. I don't do the things where it's sort of out of the box, easy and obvious...I don't do those things. That's I don't create images.”
LLMs currently have zero value for quantitative analysis because the error rate is still too high (producing wrong numbers frequently), unlike qualitative analysis where 'roughly right' outputs can be helpful and corrected by users.
“I think today it has zero value for quantitive analysis...Because if well let me let me qualify that. Do the numbers need to be right or roughly right...the problem today is it's not wrong once in a billion years it's wrong a dozen times a page”
Google's Gemini models and broader AI capabilities are currently strong and improving across the board, making Google competitive in model quality, whereas Meta's Llama 4 was perceived as an 'embarrassment' and Meta is 'scrambling to catch up.'
“you know objectively Google is clearly like firing on all cylinders now and is doing very is making great models. Llama seems to felt llama 4 was seems to have been an embarrassment and so met kind of scrambling to catch up.”
OpenAI's CEO Sam Altman is perceived unfavorably by those who have worked with him, though opinions are characterized as 'fairly unanimous' and 'fairly negative,' and he is described as a 'polarizing figure'—though the formulation suggests the opposite (that opinions are consistent rather than polarized).
“Sam Alman is I was going to say polarizing figure, but actually opinions about him tend to be fairly unanimous um and tend to be fairly negative. Like everybody who's ever worked with him quit.”