Why Everyone Is Wrong About AI (Including You) | Benedict Evans

KEYPOINTS

• 01

  His 'controversial' take is the centrist one: AI is the biggest thing since the iPhone, but only that.

  Some people say AI is bigger than computing, on the scale of electricity or the industrial revolution. Evans' base case is another platform shift: everything new gets built around it for 10–15 years, it has the same sort of effects on jobs and productivity as previous shifts, and then 'it'll just be software'. Every shift is 'different', like every financial bubble, but that doesn't mean it isn't a familiar pattern.

• 02

  You can be sure a technology is huge and still have no idea how it will play out.

  In 1995 people talked about 'cyberspace' and an 'information superhighway' run by cable and phone companies. Mary Meeker forecast email users separately from web users and expected email to be bigger. Microsoft won browsers, which turned out not to matter, and the value went to search and social 5–10 years later. In 2000, as a telecoms analyst, Evans was sure mobile internet would happen, but no one foresaw the iPhone as 'a small Mac', carriers and Nokia capturing nothing, or a 10-year lag.

• 03

  Incumbents always try to make the new technology a feature, and hindsight makes their fate look inevitable.

  People expected IBM to win PCs, Microsoft to win the internet and Google and Meta to win mobile. Kodak's 1975 digital camera was the size of a refrigerator. Once digital was viable, Kodak went all in and briefly sold the most digital cameras in the US. It lost because digital was a low-margin commodity compared with high-margin film, and because smartphones and social media killed printing.

• 04

  Google's threat isn't one rival; it's a reset in which everyone reconsiders their defaults.

  Evidence from the antitrust trial shows Google is still far better than other traditional search engines. But LLMs reset the product, how it's sold, the org chart and incentives needed to build it, and user habits. Google has big advantages in that reset, but defaults are up for grabs.

• 05

  Data isn't a moat for LLMs, because everyone needs essentially all the text that exists.

  Models train on text, not YouTube. Meta's lawsuit over downloading pirated books shows even giants lack enough long-form prose; social platforms have short snippets. Anyone with a billion dollars can scrape the web or download Common Crawl from AWS, so the data is a level playing field.

• 06

  The models look like commodities, and the consumer winner so far is brand and distribution.

  He bets most people couldn't tell Grok, Claude, Gemini, Mistral and DeepSeek apart in a blind test. Yet ChatGPT tops the app store while Gemini ranks between 50 and 100 and the others don't make the top 100. He compares it to browsers, which have barely innovated beyond tabs and search in the address bar. There's no network effect yet, since more users don't improve the model. Memory is a switching cost, and you could ask one model what it knows about you and paste it into another.

• 07

  Most people still look at chatbots and 'don't get it'.

  Across five US surveys he sees about 10% of people using an LLM chatbot daily, 15–20% weekly, 20–30% every month or two, and 20–30% who tried it and didn't see the point. Fast adoption is an unfair comparison, because it's a free website while a 1980s PC cost about $5,000 in today's money. The real puzzle is why weekly users can only think of something to do with it once a week.

• 08

  A blank chat box puts the burden on the user; products that wrap AI in a specific task get adoption.

  Evans doesn't code, brainstorm, summarize or make images, which are the obvious uses. The things he would want done, like figuring out why a client's ERP system rejects his bank details, LLMs can't do yet. A 'draft a reply' button inside Salesforce gets used because someone else has worked out the use case. He compares it to VisiCalc, where accountants finished a month's work in a week.

• 09

  Whether output that's 'roughly right' is useful depends on whether you can fix it.

  Today he says chatbots are wrong 'a dozen times a page' and have zero value where numbers must be exact. An AI-written conference bio of him was the right kind of bio with all the wrong facts. That was useless to the organizers but useful to him, since he could correct it in 30 seconds.

• 10

  AI rarely produces 'original and good', because LLMs lack a feedback loop for originality.

  AlphaGo could invent moves because every move gets a score. Monkeys at typewriters and Borges' library produce masterpieces with no way to recognize them. For an LLM, variance counts as error. Could a model have known that fans tired of 1970s prog rock would want punk, or that post-war Europe would embrace Dior's New Look? As AI handles the median, value shifts to the curated, the in-person and the unique.