On 19 February 2024 I opened a WhatsApp group for readers of Artificial Insights. These were the first messages:
Welcome to the group! I considered creating one when launching the newsletter last year but opted against it. Now I've changed my mind and want to feel out what a group can be useful for. Happy to have you here!
Without sounding too much like an Altman fanboy, I really recommend this interview from the World Government Summit. Definitely going in the newsletter next week.
The interview with NVIDIA's founder from that summit also had a few useful comments worth thinking about. I particularly liked his phrasing of industry of intelligence, quite insightful.
We worked with Omar before he became Minister of AI, for the World Government Summit. I was skeptical when they announced his position, but proven wrong 😂
Oh my… dying to know what you mean by that
As in: I was skeptical a nation needed a minister for AI and automation, but in hindsight it's obvious more countries should have one? Or at least a strategy?
This weekend I exported the chat: 17,551 messages and 3,804 links. I read all of it, with some help, and counted what I could.
I did not plan the timing. I decided today to look back at the group, and Claude, which helped me count, noticed that today, 15 November 2026, is exactly 1,000 days since the group opened. So this post is about those 1,000 days.
This post is for the people in the group. You wrote most of it.
The shape of it
Only 38 of the 1,000 days had no messages at all, and the group never went silent for more than three days in a row. The group never settled into a steady rhythm. It spikes when something big ships and then goes quiet for a few weeks. The four peaks are Apple Intelligence in June 2024, DeepSeek R1 in January 2025, GPT-5 in August 2025 and OpenClaw in February 2026. The busiest single day was last Thursday, and the cause was a meme, not a model.
We passed 300 members in December 2024, which I called twice Dunbar's number. We passed 400 in April 2025 and sit around 450 now. 645 people joined over the 1,000 days and 241 left.
Who does the talking
338 people have posted at least once. I wrote 19% of all messages, mostly links with a line of context. Two members wrote another 24% between them. The ten most active people wrote 62%, and the fifty most active wrote 90%. 121 people posted exactly once.
So the group is mostly readers. What surprised me is that the quiet members say they like it that way. When someone joked in June about "a mole" mining the chat, a few quieter members said they read everything and post little, and that this works for them. A newsletter has the same shape, you just never see it.
When we talk
The chat wakes up at 8 in Amsterdam and peaks at 2 in the afternoon, with a second wave in the evening as the Americas come online. Thursday is the busiest day. Saturday has half as many messages as Thursday, and Sunday has a third. This is a work-week group, even for something many of us do for fun.
What we talked about
For the first two years, ChatGPT was the default noun. People said "I asked ChatGPT" the way they say "I googled it". That changed in the first quarter of 2026, when Claude overtook ChatGPT in our messages. It has stayed ahead since. I can even point to the week: on 28 February, during Anthropic's standoff with the Pentagon, a member ran a poll: "Are you cancelling your ChatGPT account?" 17 of 20 voted yes.
The wider shift is from talking about AI to building with it. Mentions of agents nearly tripled. Claude Code and vibe coding went from nothing to everyday words. In October 2024 I asked how you use AI most. Ideation won with 38 votes. Coding got 13. Two years later, a lot of the chat is people showing each other what they built over the weekend.
Some topics faded. AGI comes up 40% less often than in 2024. Education, a big theme in our first year, dropped by more than two thirds. Midjourney has disappeared. China went the other way, from a curiosity to a constant, with DeepSeek in 2025 and the open models from Kimi and GLM this year.
Every month also had its own obsession. For each month I took the word that stood out most against the rest of the chat:
Not everyone liked the shift. In August 2025 one member asked why the group now talks about models instead of the "why". That is a fair note for a newsletter that set out to explain consequences, not releases.
Where the links come from
YouTube is the main source by far, with 683 links, or 18% of everything shared. X, Substack and LinkedIn follow. arXiv sits much lower, with 30 links in three years. We are a group that watches talks and reads essays, not one that reads papers. That fits the newsletter, which leans on interviews and explainers. It also explains the most useful piece of feedback I got. A member said they had stopped clicking because so many links were 90-minute videos, and asked me to add the running time. I have done that ever since.
Most links appear once. Four were posted by three or more different people, often months apart:
- AI 2027, by four people between April and September 2025. I posted it first and called it a bit like sci-fi. It kept coming back.
- Benedict Evans' yearly presentation, by four people, once for each new edition.
- The Dwarkesh Patel interview with Leopold Aschenbrenner, by three people in three weeks of June 2024.
- Google Stitch, by three people between May 2025 and April 2026.
Our predictions, graded
Last week we graded Envisioning's 2012 forecasts. The chat is full of forecasts too, so here is a short scorecard. I kept names out, except mine.
- Right
- May 2024: 12 of 15 voted that GPT-5 would be a smaller leap than expected. Most of the group read it that way in August 2025.
- May 2024: I called Humane's valuation delusional. HP bought what was left of it in February 2025.
- October 2025: a member wrote "the ads are coming" about ChatGPT. They came in January 2026.
- Partly right
- January 2025: a member said DeepSeek's effect would show within six to nine months. Open models caught up, after eighteen.
- Wrong
- April 2024: I said a general home robot was "12 months away, maybe less".
- December 2024: I said "we will probably never see GPT-4.5 / 5". Both shipped in 2025.
- October 2024 and February 2025: two members dated GPT-5 to March, then May 2025. It came in August.
- Not yet
- February 2025: a $20 bet that a humanoid would load a dishwasher in a real home within a year. Nobody in the chat has collected.
The pattern matches the 2012 map. We are good at reading software and bad at timing atoms. Every robot forecast in the chat was early.
What I keep coming back to
Reading my own messages over three years was humbling. I changed my mind on a lot of tools, but a few positions stayed the same.
One is effort. In June 2024 I wrote: "If they don't want to write it, then why would you want to read it?" In October 2025 I wrote nearly the same sentence again: "If nobody wrote it, why should you read it?"
Another is building. In May 2024 I said that with GPT I could finally build working pilots after a lifetime of not being able to code. By July 2026 I wrote that the learning curve for coding has flattened. The group watched that happen in real time, and many of you went through it too, in the Coding Club and the workshops.
The third is the group itself. In April 2024 I said the group had made the newsletter a hundred times more rewarding to write, because I stopped guessing what to write about on Mondays. That is still true.
What's next
The chat taught me three things about the newsletter. Keep it short, because several of you said it is too rich to read in one go. Say why a link matters, not only that it exists. And make room for the people who build, because that is where the energy went.
Thank you for 1,000 days. If you are reading this and rarely post, you are in good company: 288 of us wrote less than a tenth of the messages, and the group would not be the same without you.
If you read the newsletter and are not in the group yet, you can join from the community page.
Method: I exported the chat on 14 November 2026 and counted messages, posters, links and keywords with a short script. Topic counts are messages that mention a term at least once, per 1,000 text messages. The topic of each month is the word that month used most often relative to the whole chat, named by me. AI agents read the chat in eight periods to find threads and forecasts, and I checked every quote against the export. The charts use counts only, no names and no message text.
