Martin Källström
knowledge / stories

Watching AI land on other people

Most accounts of the generative-AI wave are written from benchmarks and launch events. The Co-creating with AI podcast, which Martin recorded with his Multiply co-founder Rasmus Adler Wahlberg from early 2023, is a different kind of record: two founders reporting, week by week, where they had seen AI land on someone who was not them. Schoolchildren, a goldsmith, friends with privacy scruples, a marketing agency, a lawyer's colleagues. This page keeps to what the two of them say they observed, says who observed it, and follows the arc from novelty in March 2023 to habit by 2025. Martin's conclusion is about where adoption happens: at the kitchen table and in the specific workflow, long before it shows up in the statistics.

March 2023: the classroom got there first

The earliest field report is Martin's. Recording in early March 2023, he said generative text was taking over private life, work and school at once, and offered his evidence: “my daughters report that some classmates have ChatGPT open every class they're in and just continuously chatting with and actually staying on topic”, using it to understand what the teacher was saying. ▶ 2:21 ▶ 2:46 His inference was that if kids in Sweden were doing it, kids everywhere were, and his framing of what he was watching became the podcast's name: “no work is disappearing to AI, and it's just that we are starting to co-create with AI.” ▶ 3:37

The numbers, when he went looking a few months later, were humbling. He cited a report that 2% of the US population had tried ChatGPT, and that 15% of those found it a game changer, “a tiny, tiny click” of the population. ▶ 29:42 Who were they? In his estimate, “the well-educated, high-salaried people that just want to increase their productivity”, plus kids doing homework and entrepreneurs writing investor pitches. ▶ 30:24 Rasmus had a matching anecdote from the street: an Instagram account had gone around asking people about ChatGPT and “they had no idea what it was”; he guessed that the share of repetitive knowledge work actually done by AI at that point was a “very, very, very, very small percentage”. ▶ 27:09 Martin's reading of the gap was that it was binary rather than gradual: “either you are in it and you quickly realize the value, or you are just not even touching it”, and the second state can last forever. ▶ 26:01 He went as far as wondering aloud whether some AI doomers wanted to keep the advantage to themselves by spreading fear, and conceded that this “could actually be a viable strategy”. ▶ 26:26

The moment it clicks is when it is built in front of you

Selling Multiply to agencies gave both hosts a front-row seat to the second adoption happens. Martin noticed that when he demoed a workflow he had written himself, prospects liked it but were not very impressed; when he asked them to type part of the prompt, they were, because “it's not a canned demo”, and it clicked that the tool was for anything. ▶ 16:23 He generalised: it clicks much faster when the workflow is the agency's own suggestion rather than one of three existing examples. ▶ 17:23 Rasmus reported the customer line that had become Multiply's wow moment: “I just said this and you just built it in, like, 5 minutes in front of me”, which he compared to writing formulas in a spreadsheet, except the formulas are prompts. ▶ 17:53

The clearest stories of value are mostly Rasmus's. A friend in a heavily regulated field uploaded the entire body of regulation into a no-code chatbot builder, restricted it to answering from that data, and built it in a couple of weeks; when he demoed it, people in the field said they had been hunting for one particular answer for ages, he typed it in, and the answer came. Rasmus allows that it may have been a lucky shot. ▶ 19:31 ▶ 20:23 A marketing agency trying Multiply's research agent had lost the celebrity who was to be the face of a client's brand and needed someone with the same appeal, not A-list, with an existing following; out of about ten suggestions from the first couple of prompts, “There was like 3 really good ones.” ▶ 19:28 ▶ 20:27

Martin's counterpart is Prosperous Planet, a Swedish company helping chains like Pressbyrån choose suppliers with a smaller climate footprint. As a side project, before custom GPTs existed, they fed each customer's relevant global media into a vector store and generated a daily summary of news affecting that customer's purchasing; it turned out “hugely popular with the clients” and was becoming part of the main business. ▶ 15:44 ▶ 17:53 ▶ 18:18 At the enterprise end, he relayed OpenAI's showcase case, Moderna: a six-month trial in which around 40% of staff created their own GPT, 750 GPTs in total, and a legal department that built two, for compliance and agreements, after which nobody there worked without them. ▶ 27:18 ▶ 27:43 He noted it was the one case OpenAI was “very proud of,” not a trend.

Not every landing was soft. Their guest Elia Merling, who ran AI-coworker workshops at a Linköping conference, reported that “when we made AI too prominent within these creative processes, then many became passive”: people stepped back from the table when someone who had read the whole internet sat down at it. ▶ 12:57

From novelty to habit

By late 2024 nobody was impressed anymore; they were just using it. Martin's example was domestic. His girlfriend, a goldsmith, was on the couch asking ChatGPT who her competitors were, “not Googling them,” because she already consulted it about strategy so often that it knew to search for Swedish goldsmiths making hand-forged jewellery. ▶ 2:41 ▶ 3:07 His conclusion was about context rather than model quality: “Google doesn't know anything about her.” ▶ 3:32

The more uncomfortable observation came in the episode on AI memory. Martin said the value of giving an AI your memories is so high that people drop their initial boundaries as soon as they feel it: “I have friends who are very privacy concerned with ChatGPT who have, thanks to the memory function, just let go of that”, and in a couple of cases had made ChatGPT their personal therapist, probably telling it more than they tell most people in their lives. ▶ 20:48 ▶ 21:14 Rasmus's response was a warning rather than a celebration: a lot will go wrong because the thing will fool people that it has a right brain, that the AI girlfriend is real, that the therapist cares. ▶ 25:16 His fix was literacy of the kind social media forced on everyone: “this is a research agent and this is a therapy agent. It's not a therapist.” ▶ 26:17 See AI adoption & society.

The audience had changed too. Martin came back from a 2024 AI meetup describing a real mix of enterprise people and techies, universities and startups, and more than 50% women in the crowd; there were artists, and he got a LinkedIn request afterwards from a professor of political science. ▶ 0:31 ▶ 0:56 ▶ 1:37 Rasmus's gloss was that conversation-shaped technology reaches people the web never did, the way podcasts had, with truckers listening on the road. ▶ 1:55

Who pushes back, and why Martin thinks it matters

The resistance he saw was strongest where the tool touched identity. In an episode on AI coding he described a social-media exchange: one engineer talked about using AI to code, and another replied, in Martin's paraphrase, “since you're no longer an engineer, I don't have to take your opinion into regard.” He filed it as evidence of a denial among engineers that he also saw among creators in the AI-art debate. ▶ 16:55 ▶ 17:21 Rasmus, meanwhile, was watching regulators: he expected the line to be drawn at “AIs pretending to be human”, and thought it should be. ▶ 5:39

Put together, the observations point one way. Adoption did not arrive through the 2% statistic; it arrived one classmate, one goldsmith, one compliance GPT at a time, when the tool did something specific for the person in front of it. That is the argument behind Multiply's move from single-player to multiplayer AI ▶ 15:08, and behind the podcast's founding bet that co-creation, not replacement, was the mode that would stick. ▶ 21:24

Worth remembering