At this stage, people who still claim that generative AI doesn't work and is all bullshit are showing an amazing level of wilful self-deception. blog.google/innovation-a...

Also unaffected: Santa Claus, tooth fairy.

I hate AI because it makes confident assertions it hasn’t checked and also I saw a headline that says a study I haven’t read claims only 5% of AI pilots work

The way that ‘MIT study’ that supposedly claimed ‘95% of AI pilots fail’ (none of which was really supported by the analysis or methodology) still gets cited everywhere ironically tells us quite a lot about the error rate in human reasoning

The interesting thing about OpenAI’s ‘code red’ is that it involves pausing product and doubling down on models, and not the reverse. OpenAI’s core strategic problem is that it has no unique product and no distribution.

I am not entirely surprised that someone who rage-quit Google after publishing a paper based on made-up energy consumption data would support a book based on water consumption data that turns out to be wildly wrong.

This McKinsey chart does a pretty good job of capturing the problem in asking people or corporations "Do you use AI?" without defining 'use' or 'AI'. The US government does a survey like this. It is meaningless. www.mckinsey.com/capabilities...

One effect of the splintering of Twitter is that the people building AI are mostly still there, but most of the people who think that AI is evil and stupid left to go to places like mastodon or Bluesky, where now they just talk to each other in decreasing circles of irrelevance

Twice a year, I produce a big presentation exploring macro and strategic trends in the tech industry. New in November 2025, ‘AI eats the world’. www.ben-evans.com/presentations

Judge agrees it was really dumb for the FCC to claim Meta doesn’t compete with TikTok. This case was originally thrown out because the FCC just… forgot to say what it meant when it claimed Meta was a monopoly. Market definition is the first step in a competition case and they just didn’t bother.

Who says Europe has no hyper-growth companies?

Almost done

Idea: a book where each chapter explains why something lots of people believe without question is wrong. Facebook sells your data, buybacks are bad, rent control is good, $100 Nikes cost $1 to make, foreigners pay our tariffs, you can fix deficits by cutting waste/taxing billionaires…

"Facebook sells your data" is such a strange little cult. Not only do you persuade yourself that "sells user data" is a good way to describe a system in which no-one gets user data - you persuade yourself that this bizarre pretzeled logic is so obvious that one who says 'um, what?' is a moron

It’s interesting how consistent survey data on LLM usage is. Something like 10% of people are using this every day or so, another 15 or 20% every week or two, and then about 50% have looked at it and not come back.

This is a good expression of the bemusement that people who build companies have for the whole 'enshittification' idea (if they thought about it at all). Do you want to build something that's used a lot by a lot of people, or are you running an unambitious lifestyle business?

It’s kind of amazing how many people think that a typical company has enough profit margin to absorb an import tax of 25-54% without raising prices. Almost no company has that money.

Say what you like about Trump and his tariffs, but he's making stocks much more affordable for ordinary Americans.

It’s always interesting to read threads by people taking about ‘generative AI stealing from artists’ and then go to look at art where the concept of ‘stealing’ or ‘copying’ is meaningless (unless a cat-burglar is involved). Two totally different concepts of what ‘art’ means.

I’m trying to keep track. Does genAI mean that no one will have to write code or that everyone will be writing code?

Britain: "voting for Brexit is the greatest act of national self-harm in a generation" America: "Hold my beer"

It's worth re-reading this to understand why so many people in tech thought Biden's tech policy was so bad. Uncritical acceptance of Tristan Harris of all people, presumption that DC should control everything, and in general a tone of panic and negativity. www.politico.com/news/magazin...

From November 2023, the column I wrote about Humane for subscribers to my paid newsletter, in which I compared it to General Magic. mailchi.mp/30c174e4f48f...

OpenAI’s Deep Research is built for me, and I can’t use it. It’s another amazing demo, until it breaks. But it breaks in really interesting ways. www.ben-evans.com/benedictevan...

Congratulations to Grok on spending a lot of money to be yet another company on the leaderboards for a commodity technology with no defensibility that we know of.

LLMs are very bad at knowing if they are wrong (a deterministic problem), but very good at knowing if they would probably be wrong (a probabilistic problem).

“We’re going to kill McKinsey” startups are the new “we’re going to kill Bloomberg” Those companies do difficult valuable things that people pay for, and there are graveyards full of failed competitors. It might be harder than it looks.

This is hilarious

The dream of AI 'computer use' is that the GUI is the new API, just as English is the new programming language. (Sadly, in tech the reality often turns out to be more complicated than the dream).

OpenAI: we got a step change in results for massively more money. Deepseek: we matched existing results for massively less money and open-sourced everything. OpenAI: we want to borrow $500bn. 🤔

I wrote something wondering - are better models better? Every week there’s a better AI model that gives better answers. But a lot of questions don’t have better answers, only ‘right’ answers, and these models can’t do that. So should we change what we expect? www.ben-evans.com/benedictevan...

I don't know what we should make of an announcement of a '$500bn' investment when none of the names involved so far have anything like that kind of money. Oracle barely has $10bn FCF. OpenAI has none. MGX was last heard of 'targeting' $100bn AUM. Softbank?

There are so many dark patterns in Substack that I really wonder about the trade-offs between an eye-catching total subscriber figure (a vanity metric) and the open rate (your actual audience). I have stayed on an old-school stack with no third party network effects, but I also have a 50% open rate

How many people in tech think that one of these is worth the other? That AI executive order was really dumb. Amazingly, terribly dumb. But.

The whole 'LLMs will kill SaaS because now you can just ask' thesis conflicts with the fact that every no-code app hits a hard ceiling and bottom-up enterprise adoption almost never works. People don't know what tools they need and they don't know what they're doing that could be automated.

Reading a policy document arguing for an entirely new tech ecosystem in Europe, covering everything from chip design to a new smartphone OS GLENDOWER I can call spirits from the vasty deep. HOTSPUR Why, so can I, or so can any man; But will they come when you do call for them?

2005: another office suite. It does most of the stuff every office suite did a decade ago, ‘but it's Open Source!' 2025: another photo sharing app. It does most of the stuff that every photo sharing app did a decade ago, ‘but it's decentralised!' 2045: a new AI assistant…

It occurs to me that the kind of people who argue that Meta should no longer try to police anti-social behaviour generally don't think that the San Francisco police should ignore anti-social behaviour. And vice versa

One of the challenges in talking about content moderation, a little like talking about newspapers, is there’s almost nothing you can say that wasn’t said at great length five or 10 years ago, and anyone who thinks it’s easy and simple at this point hasn’t been listening and doesn’t want to hear

Looking at predictions that AGI will happen this year, I am reminded of the theologian's joke that the problem for Jews is that you wait and wait for the Messiah and he never comes, where the Christian problem is that he came and nothing changed.

A gentle suggestion: any creative or intellectual work that can be replaced by a pattern-matching machine producing the average of what people would probably say might not have been very creative or intellectual to begin with.

Apple doesn’t serve well-targeted ads based on text you type, describing exactly what you’re looking for, in the search box in the App Store, but a million gullible idiots believe they’re serving uncannily accurate ads based on snippets of random conversations secretly recorded from across the room.

I sometimes come across clusters on Bluesky and LinkedIn of hundreds of people telling each other that LLMs are just piracy-based nonsense machines whose only purpose it to churn out content. Yes, “someone on the Internet is wrong” - but it’s such a strong echo chamber, and so utterly uncurious.

JOURNALIST REFRESHER ABOUT INTERVIEWING ARTIFICIAL INTELLIGENCE. Hi. When new LLM bots drop, you can't interview them like people who understand the process that went into creating them. They may give you an answer, but they don't actually know the answer. You can't ask "who created you?"

It isn't news that Microsoft plans to spend $80bn on capex in the 12m to June this year... but it is worth pointing out the context. If only there was one kind of pattern-recognition system that could tell you how charts like this tend to play out.

It looks like Deepseek might have snuck in just under the wire with the biggest and most consequential AI story of 2024.

It is possible to believe a lot of people will lose their jobs because of the current wave of AI-led automation and know this is painful… And also point out this has happened over and over again for 150 years and worked out better than the alternative.

The history of automation is full of stories of individual groups of privileged workers who lost out… but it’s extremely hard to see automation overall as ever having been bad for workers.

‘Please shut your window blind, on a daytime flight, and sit in darkness for 5 hours, because someone on the other side of the cabin wants to sleep.’ It’s three o’clock in the afternoon. You have a blindfold.

It is trivial to make chatGPT o1 fail confidently and repeatedly on data retrieval. But, is that the test? If I tell it half a dozen times how do do something, shouldn’t that work? Or is this like complaining that an Apple II can’t replace a mainframe?