--- url: 'https://workflows-of-ai.com/blog/where-are-we-on-the-sigmoid-curve' title: 'Where are we on the sigmoid curve?' author: name: 'Marcus Johansson' url: 'https://workflows-of-ai.com/about/marcus-johansson' date: '2026-11-10T10:51:40+01:00' updated: '2026-11-10T10:51:40+01:00' type: article summary: 'In February 2023 Tom Scott said AI was on a new sigmoid curve and that he had no idea how far along it we were. Almost four years later, I try to answer him: for capability, for adoption and for the money.' tags: - 'Marcus Johansson' - AI image: 'https://workflows-of-ai.com/sites/default/files/previews/4e2a898cd96be6ae.png' files: - 'https://workflows-of-ai.com/sites/default/files/podcast/c2fc732b372aa4d4.mp3' - 'https://workflows-of-ai.com/sites/default/files/slides/1d4b0e2aa95935ca.mp4' - 'https://workflows-of-ai.com/sites/default/files/spoken/a27f8d35db2a2cf4.mp3' published: true og: site_name: 'Workflows of AI' 'image:width': '1200' 'image:height': '630' 'image:alt': 'An 8-bit developer with a laptop climbs the steep middle of a blue S-shaped curve, past a pink flag that says You are here, towards the flat top.' --- AI # Where are we on the sigmoid curve? In February 2023 Tom Scott said AI was on a new sigmoid curve, and that he had no idea how far along it we were. Almost four years later, I try to answer him. Marcus Johansson November 9, 2026 · 14 min read If I could make everyone watch one video about AI, it would be a 15 minute video from February 2023. It was made a couple of months after ChatGPT came out, by someone who just wanted to fix his email backup. It explains the promise of AI, the fear of it and the possible bubble better than anything I have read since. [Watch on YouTube: I tried using AI. It scared me. (Tom Scott)](https://www.youtube.com/watch?v=jPhJbKBuNnA) ## Who is Tom Scott [Tom Scott](https://en.wikipedia.org/wiki/Tom_Scott_(presenter)) is a British YouTuber and former web developer. For ten years he put out a video every single week: short, well-researched explainers about places, language, history, science and technology. If you have watched educational YouTube in the last decade, you have probably seen him. He is about my age. When he says in the video that he was lucky enough to live through the mass adoption of the internet, and that he was on the cutting edge of it, I know exactly what he means, because I lived it too. ## The sigmoid curve The core of the video is one idea. Useful new technologies follow a sigmoid curve, an S shape. There is a slow start while the technology is invented and the worst problems get sorted out. Then growth explodes: people find it useful, it gets better and better, competitors pile in and everyone finds more real uses for it. Then it reaches the limit of what is possible with that technology, and progress flattens out again. That is why very few people camp outside Apple Stores for the new iPhone any more. For digital things, I have always, a bit stupidly, been at the start of the sigmoid curve. I once had the address marcus@algonet.se. [Algonet](https://sv-wikipedia-org.translate.goog/wiki/Algonet?_x_tr_sl=sv&_x_tr_tl=en&_x_tr_hl=en) was Sweden's first internet provider for private people, and Marcus is the most common name of my generation in Sweden. Getting your bare first name at a Swedish provider says less about me being clever and more about my father's interest in new technology, and about how few people cared about the internet when Mosaic was still the go-to browser and Netscape was seen as something disruptive. Tom picks Napster, 1999, as the moment he should have seen the storm coming. It wasn't the first sign, but it was the first big one: millions of non-technical people using something by choice, the first to get mass-media attention and legal problems. In 1999 [about 41% of Swedes were online](https://data.worldbank.org/indicator/IT.NET.USER.ZS?locations=SE). It was the start of the steep part, not the start of the curve. I remember it as a change in what I did every day. This is how getting a new song changed for me, from the record shop to Spotify: How I got a new song Before 1999 A trip to the shop and a whole album June 1999 Search, download and wait 1999 Download, then copy it over a cable July 2001 The same, from another program 2002 Download, copy, wear it October 2008 Search and press play 1. Before 1999 ### The record shop Music came on CDs. To get the one song you liked, you bought the whole album and carried a Discman if you wanted it with you. Getting a new song A trip to the shop and a whole album 2. June 1999 ### Napster Suddenly almost any song was one search away. You still waited for every download on a modem, but most of the time you only got the songs you wanted (sometimes an unexpected virus came along too), and millions of people who were not nerds did it by choice. Getting a new song Search, download and wait 3. 1999 ### Creative Nomad My first MP3 player, imported from the US through eBay. The sound quality was still bad, and it was almost as big as a CD player. It had hardly any storage, 64 MB or about an hour of music, so you picked the songs on the computer before you left the house. Getting a new song Download, then copy it over a cable 4. July 2001 ### Napster shuts down A court shut Napster down. It didn't matter: by then there were KaZaA, LimeWire and the rest. In 2003 the Swedish site The Pirate Bay opened, and I sailed high on it too. Nobody went back to buying albums for one song. Getting a new song The same, from another program 5. 2002 ### mpio DMK A tiny player on a cord around the neck, and better in almost every way. You could run with it, it was minimal, and it had enough storage for anything you did outside, like walking or running. My music didn't live on CDs any more. Getting a new song Download, copy, wear it 6. October 2008 ### Spotify Every song, legally, one search away, paid for with a monthly fee or with ads. Sweden and much of the western world already had the internet built for it. Downloading stopped making sense, and this is simply how people listen to music now. Getting a new song Search and press play None of those steps felt dramatic while they happened. Each one was just a bit easier than the one before, and then one day the old way was gone. That is what the steep part of the curve feels like from the inside. Then he says the thing that makes the video so good: > I think we're on a new sigmoid curve, and I have no idea how far along that curve we are right now. — Tom Scott, February 2023 He lays out three options. If we are near the top, programmers and artists get some new tools, but nothing at a human level, and not many jobs change. If we are in the middle, we get very impressive tools soon that still need humans to work them. And if we are at the start, at the Napster point, then everything is about to change, just as fast and just as strangely as it did in the early 2000s, and this time he is on the wrong end of it, like the music executive in 1999. He ends by saying that at some point he will look back at the video and, with hindsight, easily see where we were on the curve. It has been almost four years, so let me try. ## It is three curves, not one The reason the question is so hard is that "AI" isn't one curve. There is how good the technology is, how many people use it, and how much money is being bet on it. They don't move together, and most arguments about AI are people talking about different curves. ### Capability: still climbing In the video, ChatGPT writes a Google Apps Script that is a bit wrong, Tom fixes it with plain English, and he decides it codes about as well as he does with his out-of-date skills. The next day he gives it a problem that needs a logical leap, and it confidently fails in several obvious ways. So he isn't worried about his job. "Not yet, anyway." That second problem is the one that has changed the most. In 2026 I don't ask a chatbot for a script, I hand a coding agent a task and it reads the code, runs the tests and fixes its own mistakes. I configured [this website](/blog/how-i-vibe-configed-this-website) that way, and I had an agent document every Drupal 11 module for [the module knowledge base](/blog/theres-a-module-for-that). It still fails, sometimes confidently, but at a size of task that was science fiction in February 2023. My own way of writing code went through the same kind of steps as music did: How I wrote a new feature Before 2022 Search, read and type it December 2022 Ask, copy, paste and fix 2023 Start typing, press Tab and verify 2025 Describe it, then review it 1. Before 2022 ### Autocomplete and Stack Overflow The editor knew the function names. Everything else came from the documentation, a search or an answer on Stack Overflow, and you typed it in yourself. You understood every line, because you wrote every line. Writing a feature Search, read and type it 2. December 2022 ### ChatGPT You described the problem in a chat window and copied the answer into the editor. Half of the time it was wrong, and you needed to modify it. Writing a feature Ask, copy, paste and fix 3. 2023 ### GitHub Copilot The suggestions moved into the editor. It wrote the rest of the line, then the rest of the function, in grey text you accepted with Tab, and I verified what it wrote. Writing a feature Start typing, press Tab and verify 4. 2025 ### Coding agents I describe the task, and Claude Code reads the code, writes it, runs the tests and fixes its own mistakes. I review the code instead of writing it. Writing a feature Describe it, then review it There is a measurement for this. [METR](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/) tracks how long a task, measured in human time, the best models can finish half of the time. In March 2025 that length had doubled about every seven months for six years. Their [update from January 2026](https://metr.org/blog/2026-01-29-time-horizon-1-1/) puts the doubling since 2024 at about 89 days. A curve that gets steeper is not one that is flattening. Benchmarks only measure what can be measured, and the top of an S always looks like a straight line from below, so I can't promise it keeps going. But I don't see the bend yet. ### Adoption: possibly the steep middle Tom's barber had used ChatGPT to write a formal email. For Tom, that was the Napster moment: not the nerds, but everyone. Pew Research asked Americans in February 2026, and [about half of US adults now use AI chatbots](https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/) (49%), up from a third in 2024. 44% have used ChatGPT, up from 18% in 2023. About a quarter use a chatbot every day. That is fast, even for a sigmoid curve. Not every new technology spread faster than the one before, but the curves have been getting steeper. Here is how many years it took from launch until half of the US had it: Years from launch until half of the US had it - Telephone 70 years · 1876 to 1946 - Electricity 43 years · 1882 to 1925 - Radio 11 years · 1920 to 1931 - Colour TV 20 years · 1954 to 1974 - Mobile phone 19 years · 1983 to 2002 - Internet 9 years · 1993 to 2002 - Smartphone 6 years · 2007 to 2013 - AI chatbots 3 years · 2022 to 2026, at 49% Share of US households, from Our World in Data (Comin and Hobijn and others). AI chatbots: share of US adults, Pew Research Center, February 2026, still just under half. The launch years are my choice of starting point. Years from launch until half of the US had it | Telephone | 70 years (1876 to 1946) | |---|---| | Electricity | 43 years (1882 to 1925) | | Radio | 11 years (1920 to 1931) | | Colour TV | 20 years (1954 to 1974) | | Mobile phone | 19 years (1983 to 2002) | | Internet | 9 years (1993 to 2002) | | Smartphone | 6 years (2007 to 2013) | | AI chatbots | 3 years (2022 to 2026, at 49%) | And the US is not where the enthusiasm is. People in emerging economies are much more positive about AI than people in Europe and the US. In a [2025 study by the University of Melbourne and KPMG](https://mbs.edu/faculty-and-research/trust-and-ai) of 48,000 people in 47 countries, three in five people in emerging economies trusted AI systems, against two in five in advanced economies. The [Stanford AI Index 2025](https://hai.stanford.edu/ai-index/2025-ai-index-report/public-opinion) shows the same split in an Ipsos survey from 2024: 83% in China, 80% in Indonesia and 77% in Thailand say AI products have more benefits than drawbacks, against 40% in Canada, 39% in the US and 36% in the Netherlands. A large part of the world is not hesitating at all. Then there is Germany, where I live, the country I would pick as the example of always being off the curve on digital things. As Angela Merkel said: ["Das Internet ist für uns alle Neuland"](https://netzpolitik.org/2013/das-internet-ist-fur-uns-alle-neuland/), the internet is new territory for all of us. #Neuland became a meme. During the pandemic, some health offices [still reported COVID cases by fax](https://correctiv.org/faktencheck/hintergrund/2021/11/12/meldung-erfolgt-per-fax-warum-in-deutschland-niemand-weiss-wie-viele-geimpfte-an-covid-19-erkranken/). And in 2024 only 36.8% of German households could get fibre to the home, against 69.2% in the EU as a whole, according to the [EU's Digital Decade report](https://digital-strategy.ec.europa.eu/en/factpages/germany-2025-digital-decade-country-report). I had fibre when I left Sweden in 2008... From a digital point of view, I seem to have picked the worst country to settle in. Outside of the red tape, most other parts are fantastic, though. Half is the middle of an S, and the middle is where it is steepest. For the internet in Sweden, that was around 2000. So for adoption we are past the Napster point. We are in the part Tom was afraid of, where it changes everything, fast and strangely. This is what that curve looked like for the internet in Sweden, the one I lived through: The internet in Sweden 2024 96% of Swedes online 0% 50% 100% 1992 2000 2010 2020 2024 AI chatbots, US adults, 2026: 49% 1. 1994 · 3% ### Algonet I get marcus@algonet.se. Hardly anyone I know is online, and the ones who are use Mosaic. 2. 1997 · 24% ### The web arrives In three years the share of Swedes online goes from 3% to almost a quarter. Still mostly nerds, and still on modems. 3. 1999 · 41% ### Napster People who are not nerds go online by choice. The slow start is over. 4. 2001 · 52% ### The dot-com crash The money crashes, Napster shuts down, and for the first time more than half of Sweden is online. The crash didn't slow it down: the next year was the steepest of all. 5. 2003 · 79% ### Two years after the crash 79% of Swedes are online. The Pirate Bay opens. 6. 2008 · 90% ### Spotify 90%. The internet is built, and new services can just assume it is there. 7. 2024 · 96% ### The top About 96%. Nearly everyone who can be online is, and nobody gets excited about it any more. 8. 2026 ### And AI? Half of US adults use AI chatbots. On the internet curve, that is Sweden in 2001: the year of the crash, right in the steep middle. Individuals using the internet in Sweden, % of population, World Bank (IT.NET.USER.ZS). The dashed line: US adults using AI chatbots, Pew Research Center, 2026. The internet in Sweden | Year | Share | |---|---| | 1992 | 1.5% | | 1993 | 1.7% | | 1994 | 3.4% | | 1995 | 5.1% | | 1996 | 9% | | 1997 | 23.7% | | 1998 | 33.5% | | 1999 | 41.4% | | 2000 | 45.7% | | 2001 | 51.8% | | 2002 | 70.6% | | 2003 | 79.1% | | 2004 | 83.9% | | 2005 | 84.8% | | 2006 | 87.8% | | 2007 | 82% | | 2008 | 90% | | 2009 | 91% | | 2010 | 90% | | 2011 | 92.8% | | 2012 | 93.2% | | 2013 | 94.8% | | 2014 | 92.5% | | 2015 | 90.6% | | 2016 | 89.7% | | 2017 | 93% | | 2018 | 89.2% | | 2019 | 94.5% | | 2020 | 94.5% | | 2021 | 94.7% | | 2022 | 95% | | 2023 | 95.7% | | 2024 | 95.5% | ### Money: running ahead like headless chickens And then the bubble. According to [J.P. Morgan Asset Management](https://am.jpmorgan.com/dk/en/asset-management/institutional/insights/market-insights/investment-outlook/technology-and-ai/) in June 2026, analysts expect five US companies to spend about 697 billion dollars on capex this year, 173 billion more than they expected in January. AI capex went from 33% of the big cloud companies' operating cash flow in 2023 to an estimated 93% in 2026. That is a bet that the curve keeps going up, paid for before the revenue is there. And Nvidia's Jensen Huang might have just popped the bubble by [breaking another commitment](https://nvidianews.nvidia.com/news/a-message-to-our-company). ## A bubble does not stop the curve The dot-com crash came in 2000 and 2001, right in the middle of the steep part. Companies went bankrupt, the Swedish start-up Boo.com among them, and a lot of people lost a lot of money. And internet use just kept going up. Sweden went from 46% online in 2000 to 79% three years later. The crash was about the money being early and spread over too many bets, not about the internet being useless. Much of the fibre laid by companies that went bust was used later anyway. That is my answer to the bubble question. It can be a bubble and a sigmoid curve at the same time. The money curve looks to me like it is ahead of the other two, the way it was in 1999, and it may well have to wait for them in a painful way. But a correction in the money would not move the capability or the adoption back to 2023. Nobody uninstalled the internet in 2001. Just don't forget how to develop applications yourself: Claude Code subscriptions will not last forever. ## So where are we? Using Tom's three options: not near the top, because the capability curve isn't bending yet. For adoption, we are in the middle, at the steepest point. And the Napster moment he was afraid of was already behind us when he made the video. ChatGPT was the Napster, the barber was the sign, and we are now somewhere around 2000 in internet years: the steep part, maybe with a crash in the money ahead, and with the strange changes still to come. So my summary is that I still feel the way I did when that video came out. It is impossible to say where this ends, but we know that 2023 was not the end, at least. If the curve flattens soon, I think it will be because of electricity rather than the models: data centres need power, and power grids take decades to build, not months. I'm as excited and as scared as I was back then. If you want one very tangible proof that we have moved somewhere between 2023 and 2026, look at Will Smith eating spaghetti. The first one is from 2023, the second from 2026: 2023 ## Will Smith eating spaghetti An AI video from 2023, made with ModelScope. [Source on YouTube (opens in a new tab) ↗](https://www.youtube.com/watch?v=XQr4Xklqzw8) 2026 ## Will Smith eating spaghetti An AI video from 2026. [Source on YouTube (opens in a new tab) ↗](https://www.youtube.com/watch?v=xnY-j43ih_0) And the more that gets released in AI, the more I'm convinced that trying to figure out the future is impossible. One could hope for better people or governments leading this revolution, but the revolution won't stop. *My views are my own and not those of my employer.* Filed under [AI](/blog?category=9) This content was produced by AI and edited by a person. How was AI used?AI takes my rambling notes, structures them and writes a first draft. I then prompt it to use interactive components, and at the end I read through it and rewrite it until it says what I want it to say. Marcus Johansson Tinkers with AI and Drupal [About Marcus](/about/marcus-johansson)[Drupal.org profile](https://www.drupal.org/u/marcus_johansson)[Drupal AI](https://www.drupal.org/project/ai) ## Podcast: Where are we on the sigmoid curve? [Listen: Podcast episode (4:53)](/sites/default/files/podcast/c2fc732b372aa4d4.mp3) Listen in your podcast app - [Overcast (iPhone)](overcast://x-callback-url/add?url=https%3A%2F%2Fworkflows-of-ai.com%2Fpodcast.xml) - [AntennaPod](https://antennapod.org/deeplink/subscribe?url=https%3A%2F%2Fworkflows-of-ai.com%2Fpodcast.xml&title=Workflows%20of%20AI%20Podcast) - Transcript AI disclosure This podcast episode is created and spoken by AI based on articles written on workflows-of-ai.com. Kwame Kwilson Today's post asks where we are on the sigmoid curve. Back in February twenty twenty-three, Tom Scott made a video saying AI was on a new S curve, and that he had no idea how far along it we were. Marcus tries to answer him, almost four years later. Erin, do you believe anyone can actually answer that? Erin Milleu Not with certainty, no. The top of an S always looks like a straight line from below. But I'll give him this: he doesn't pretend to know. He splits the question up, which is rather the sensible thing to do. Kwame Kwilson Let's start with the curve itself. Tom picks Napster, nineteen ninety-nine, as the moment he should have seen it coming. And Marcus points out that in nineteen ninety-nine about forty-one percent of Swedes were already online. Jane Olson Which is the bit I like. Napster wasn't the start of the curve. It was the start of the steep part. Erin Milleu Yes, but be careful with that. Tom's Napster for AI was his barber writing a formal email with ChatGPT. One barber isn't data. Jane Olson Fair. That's why the post brings numbers for it later. One barber is the sign, not the proof. Erin Milleu Then I'm happy. Mostly. Kwame Kwilson So the post says AI isn't one curve, it's three. Capability, adoption and money. Erin, capability first. Erin Milleu This is the one I'd watch. METR measures how long a task, in human time, the best models can finish half the time. In March twenty twenty-five that length was doubling about every seven months. Their January twenty twenty-six update puts it at about eighty-nine days. Jane Olson And you can see it in the work. Marcus had an agent document every Drupal eleven module for a knowledge base. In twenty twenty-three you'd ask a chatbot for one script and fix it yourself. Erin Milleu Exactly. And a curve that's getting steeper isn't one that's flattening. He's careful to say benchmarks only measure what can be measured, which I appreciate. Jane Olson He also has a little timeline of how he writes code. Stack Overflow, then ChatGPT, where half the time it was wrong and you had to fix it, then Copilot, then agents. Erin Milleu Each step a bit easier than the last. That's what the steep part feels like. Kwame Kwilson Adoption next. Erin Milleu Pew asked Americans in February twenty twenty-six. Forty-nine percent of US adults use AI chatbots. That was a third in twenty twenty-four. Jane Olson And the US isn't even where the enthusiasm is. In an Ipsos survey, eighty-three percent in China say AI products have more benefits than drawbacks. In the US it's thirty-nine. Erin Milleu Quite a gap. Half is the middle of an S, and the middle is where it's steepest. On his internet curve for Sweden, forty-nine percent is roughly twenty oh one. Jane Olson The year of the crash. Which is a nice segue, Kwame. Kwame Kwilson It is. The money. J.P. Morgan says five US companies are expected to spend about six hundred and ninety-seven billion dollars on capex this year. And AI capex went from thirty-three percent of the big cloud companies' operating cash flow in twenty twenty-three to an estimated ninety-three percent this year. Erin Milleu That's a bet paid for before the revenue is there. No other way to read it. Jane Olson Agreed. The post says the money is running ahead like headless chickens. Erin Milleu Rather vivid. It also says Nvidia's Jensen Huang might have just popped the bubble. I'd leave that one as a might. Kwame Kwilson Jane, does a bubble stop the curve? Jane Olson That's the post's best point. The dot-com crash came in two thousand and two thousand one. Companies went bust. And Sweden went from forty-six percent online in two thousand to seventy-nine percent three years later. Nobody uninstalled the internet. Erin Milleu And a lot of the fibre laid by companies that went bust got used anyway. Jane Olson Though not in Germany, where Marcus lives. In twenty twenty-four only about thirty-seven percent of German households could get fibre to the home. Some health offices reported COVID cases by fax. Erin, I assume you'd fax your takeaways in, if it were allowed. Erin Milleu Only in triplicate, Jane. Properly stamped. Kwame Kwilson Speaking of takeaways. One each. Erin first. Erin Milleu Capability isn't bending yet. Watch that curve, not the share prices. Jane Olson Mine's from the post itself. Don't forget how to develop applications yourself. The subscriptions won't last forever. Kwame Kwilson And Marcus ends where Tom started. Excited and scared, about as much as in twenty twenty-three. Thanks for listening. ## Narrated slides: Where are we on the sigmoid curve? [Watch: Narrated slides (3:52)](/sites/default/files/slides/1d4b0e2aa95935ca.mp4) [Present the slides yourself](?presentation=true) ## The post in three sentences Tom Scott's February 2023 video about trying ChatGPT is still the best short explanation of the promise, the fear and the possible bubble of AI, and it ends with him not knowing where on the sigmoid curve we are. Almost four years later, my answer is that there are three curves: capability is still climbing, adoption is at the steep middle where the internet was around 2000, and the money is running ahead of both. A crash in the money would not stop the curve, just as the dot-com crash didn't stop the internet.