Martin Källström
knowledge / stories

Work is machine simulation

Over sixty episodes of Co-creating with AI, Martin built a private vocabulary for explaining what language models are, what they are for, and what they are not. Most of the terms were coined on air, in reply to a worry from his co-host Rasmus Adler Wahlberg or a guest, and most have stuck, in Multiply's product decisions and in the design of Kindship. The vocabulary carries an argument: much of what we call work is a human pretending to be a machine, the skills we fear losing were often installed by bad technology, and a model is a mirror and a completion engine before it is anything like a mind. What follows is that glossary, roughly in the order the terms appeared.

Deficient technology

The founding exchange is in episode 5. Rasmus confessed that "Using Google Maps actually makes my sense of direction, like, makes me lose my sense of direction" and, watching his daughter abandon a push-along train for a motorised one, worried about "amputating parts of our skill set". ▶ 12:35 ▶ 13:00 Martin's reply introduced the term: "the things that we are amputating are that stuff that has been artificially implanted at some point by deficient technology, like us remembering phone numbers". ▶ 14:36 A phone number exists because the phone could not take a name; "The reason why we need the sense of direction on the road network is a deficiency of the road itself." ▶ 15:02 Maps that know the traffic on every street give you something you never had: "So you're augmenting your mind rather than amputating your mind." ▶ 15:28 The test, whenever a lost skill is mourned, is whether it was human or a workaround.

Machine simulation

The same answer ran straight into the thesis this page is named for. What Martin wants from work is to "interact with other people, relate to other people. That's what I want to do and create, like have the sense of creation". ▶ 15:54 Most tasks are not that: "a lot of work we do has forced us to simulate machines, and it's not too bad to give that back to the machines to do." ▶ 16:20 The phrase reframes automation from loss to return; the archetype, named a year and a half later while watching Claude fill in a web form, is "we have to fill in a form and the information is in some system. So we find it and we copy it into the form". ▶ 11:11

The infant that has learned to talk

In early 2023, when people complained that ChatGPT could not code reliably, Martin objected to the expectation rather than the model: "we expect AI to be really good at coding out of the box. But we don't expect an infant baby to to already know how to code." ▶ 14:35 "what we can see ChatGPT as is that infant baby that have learned to talk", and the job now is to teach it to read, to use tools, to structure what it knows. ▶ 15:00 Scaffolding like LangChain, which "creates an inner life for the AI, that it actually thinks before it speaks", is the schooling; see AI engineering practice. ▶ 13:00

Copy-and-paste master (and monkey)

By mid-2023 the infant could talk, and Martin had discovered what that did to him. "I find myself having become a copy and paste master." He was "copy and pasting between browser windows and apps like a madman", which was not the promise of AI. ▶ 20:56 The term darkened in the next episode. Decades of specialised interfaces exist for a reason and "the chat interface maps very badly to all of those"; without Multiply, "I feel myself being reduced to a copy and paste monkey". ▶ 10:26 ▶ 17:16 It is machine simulation again, inflicted by the machine that was supposed to end it, and the clearest statement of why Multiply exists (AI-native interfaces).

Pineapples in a suitcase

Martin tracks technology with what he calls a canary: "I've often had this kind of canary bird approach to keeping track of technical development." For years it was whether the Beatles were on Spotify yet. ▶ 29:37 For language models it became one question put to every new one: "how many pineapples can fit into a suitcase and make reasonable assumptions about that?" "GPT-4 is the first one to reliably get around, get the answer right. Before that, the answer could be anywhere between 2,000 and 10,000 or like 26,000." ▶ 30:28 He keeps a second canary for autonomy, still unpassed: "I look forward to the day when I'm in a meeting and an AI sort of interrupts the conversation to say, hey guys, I actually think that you are talking about the same thing." ▶ 30:54

Completion engine

The most-used term of the later episodes is a corrective. When newcomers report that AI did not work for their stuff, Martin says they have the category wrong: "AI is not really a smart Tool or something like that". ▶ 1:58 "the foundation is a completion engine that it continues just keeps on writing from where you left off. So if the completion is not smart, then maybe it's because what you put in is not smart either", a mirror many people miss. ▶ 2:23 Learning a model is therefore empirical: "doing things that is supposed to break it or some way, like pushing the limits, that's a big way to learn for me". ▶ 9:36 It does not follow that everyone must become a prompt engineer; Martin had already insisted "we should— not force users to become prompt engineers just to co-create with AI". ▶ 15:19

System 1 and System 2

Kahneman's pair entered the vocabulary as an engineering pattern: cheaper models assembling context for a state-of-the-art reasoner, modelled on a person, since "We have famously System 1 and 2 responsible for thinking fast and slow". ▶ 14:03 ▶ 15:22 Its origin moment came when GPT-4o shipped in May 2024. After months alone building voice for agents ("since I've been solo devving on voice for the past few months, it's such a huge relief for me to get this"), he decided OpenAI had built the fast layer for everyone, "so it's like they've built sort of system 1 of AI where the think fast", leaving him to "worry less about the mechanics of just having a UI to the user". ▶ 9:40 ▶ 10:57 ▶ 11:34 By late 2024 it was the organising metaphor of Kindship, "where I explore artificial consciousness through this System 1 and System 2 metaphor", and after guest Elia Merling observed that agents need a denser channel among themselves, he extended it: humans talk to an AI's System 1 while "their deeper thinking in System 2 can be a shared subconscious among the AI agents". ▶ 38:03 ▶ 38:28 (The same demo later cost him a project; see Clone the repo before you believe the demo.)

English is the last programming language

Borrowed, with attribution, from Andrej Karpathy: "English is the last programming language that we will need", Karpathy's standing joke about coding almost everything in English inside Cursor. ▶ 5:08 Martin's version is blunter: with multi-file editing "you no longer need to actually know how to code"; you can build a Mac, web or iOS app "just from knowing English and being really picky about what you want". ▶ 12:42 Picky is the operative word: English is a programming language only for those who write it precisely.

The whisper game

This one is Rasmus's, adopted on the spot. Martin was explaining why chains of agents fail: if a quality check is weak, "further down the line, other agents are going to trust that input". ▶ 10:26 Rasmus: "it's in Swedish called the whisper game", the children's circle where a sentence is passed around and transformed. ▶ 11:48 Martin's fuller statement is from a few weeks earlier: "there's a compounding of errors as soon as one LLM gets a slight hallucination or error into the workflow", which is why agent frameworks stayed a research topic and why "the demo videos on YouTube are actually cherry-picked". ▶ 21:29

Delivery robots and the flamethrower dog

Martin's test for whether an agent is real is whether it touches the world. In San Francisco he met delivery robots that were merely cute until, inside a conference, "they opened the lid to give people soda cans, then all of a sudden they took on a purpose". ▶ 6:19 Robot dogs bored him, inspection use cases and all, until "I saw a robot dog with a flamethrower on its back, and it was walking around in the forest, like, setting things on fire autonomously." A silly use case, but "when it was actually affecting the world, it became completely different". ▶ 7:35 ▶ 7:09

Boring AI

The last entry is a question: "AI is very glamorous. When do we see the first wave of unglamorous AI companies?" Rasmus supplied the label, Boring AI, and wondered whether the domain was taken. ▶ 32:37 It is where the glossary lands: if work is machine simulation, the money is in giving the simulation back, form by form, in industries nobody makes demo videos about. For the historical reflex behind all of this, see Old fears, new machines.

Worth remembering