Martin Källström
knowledge / stories

Kill early, never abandon nearly done

Martin has stopped projects at every stage of their lives, and he grades those decisions very differently. Two kills he still defends: a classifieds site shelved in 2006, and a fintech scrapped in 2017 before launch. One he calls the worst decision of its year: a voice-AI product dropped two months before beta in 2024. The pattern is not "persist" or "quit" but a rule about timing. Kill a thing cleanly while it is cheap to kill; once it is nearly done, finish it. This page argues the rule out of the cases, including the ones that cut against it.

A classifieds site, nearly finished (2006)

In a 2008 talk to the Lysator computer society, Martin explains where Twingly came from. Alongside their consulting, the four founders had built a Blocket-style classifieds site with no geographic limits, on the thesis that Sweden's home-grown services had been ahead of their time but never multiplied out into the world. ▶ 4:35 "Vi hade utvecklat en Blockit-site nästan klar." (We had developed a Blocket site, nearly finished.) Then, in spring 2006, blogs started being talked about, in Sweden and abroad. ▶ 5:25 "vi tog ett jättesvårt beslut och helt enkelt la hela det stora projektet" (we took a really hard decision and simply put the whole big project) on the shelf, and built a worldwide blog index instead. What makes it a clean kill is the next sentence: "då slängde vi inte hela den gamla kodbasen utan vi fick med oss väldigt mycket där av språkidentifiering och språkhantering" (we didn't throw away the whole old code base; we took a lot of the language identification and language handling with us). ▶ 4:04

So the first case already complicates the slogan. A nearly finished product was abandoned, rightly. But it was a copy of a known model whose window was closing, dropped for a market that was opening, and the code travelled. That is a swap, not a stop.

A fintech, six months in (2017)

The second case is the one Martin narrates as a rule. Weeks after Narrative went bankrupt in September 2016 he was in Bangkok, building a fintech for Europe and Asia with a spring launch planned. His justification for jumping sectors: "Det jag kan är att bygga internationella bolag" (What I know is building international companies); the domain expert could be hired, as the camera expert had been. Breakit ↗ The product would link European investors to Thai real-estate projects paying 12 to 15 percent, and investor interest was strong. Then the Thai regulatory framework, which should already have been in place, slipped to the following winter at the earliest. "Vi ville kunna gasa på direkt, inte gå och vänta på ett beslut som vi inte kan vara säkra på ska komma. Då insåg vi att det var bättre att inte dra igång alls" (We wanted to hit the gas straight away, not wait around for a decision we can't be sure will come. Then we realised it was better not to start at all). Breakit ↗

Six months of work went with it, and he refused the word wasted: "Det här är inte på något sätt bortkastad tid" (This is in no way wasted time). The sentence that states the rule follows immediately: "det är bättre att vi tar det här beslutet nu och inte tvingas ta det senare när vi kommit ännu längre" (it's better that we take this decision now than be forced to take it later, when we've come even further). Breakit ↗ The price of stopping rises with progress; pay it while it is low. Four years later Martin filed Bangkok, along with New York, under "a bit shorter journeys perhaps". ▶ 9:03

The counter-case: the company that never decided

The rule is about projects. Applied to whole companies, Martin's record points the other way, and he knows it. With Twingly the strategy was to take 10 million kronor in 2007 and ask "Hur ska vi få de här pengarna att räcka så länge som möjligt?" (How do we make this money last as long as possible?). His verdict on that, given in 2017: "det värsta som kan hända med ett företag är egentligen inte att man misslyckas utan att man inte kommer någon vart" (the worst thing that can happen to a company isn't really that it fails but that it gets nowhere), chewing on year after year "utan att varken växa eller kunna ta ett beslut om att göra någonting annat" (without either growing or being able to decide to do something else). ▶ 12:55 Narrative's go-big-or-go-home strategy was the reaction: bet big so you find out fast whether you are wrong. ▶ 11:47 Breakit ↗

Yet the company that never killed anything is the one still alive. In 2021 he concedes that Twingly "got a very slow growth curve" but is "a super stable company that is still producing value", ▶ 12:19 while the company built to "make the market tell us if the company should survive or not" was left with very small margins and did not survive. ▶ 12:59 His conclusion is unusually unresolved: "neither of those was right. So I'm guessing it's something in between." ▶ 13:50 And the regret he names for Narrative is the reverse of killing early: they ran the initial strategy all the way, because "att byta själva grundstrategin för ett företag är ett svårt beslut" (changing a company's basic strategy is a hard decision). ▶ 26:12 For companies the sources give a different test, set out in Failure & resilience: you stop when no plan you can believe in remains. ▶ 22:36

Two months from beta (2024)

The case that gives the rule its second half is told in the Co-creating with AI podcast, in Martin's own words: "I can dive into like the worst decision I made this year based on pure hype". He had spent the winter on "a good unique take on conversational AI" and flew to San Francisco with a demo to get feedback. ▶ 11:19 There he found well-funded teams describing their projects in exactly his words, and "when I came home, I felt that some of the energy and excitement in my own work had been drained by that trip to San Francisco". ▶ 12:37 Then OpenAI showed its advanced voice demo: "what they announced was the proper way of doing conversational AI". ▶ 13:27 "that was for me the nail in the coffin of, of my, like, the serious attempt of making something unique in voice AI". ▶ 13:52

The reckoning comes a few sentences later. "I was probably something like 2 months away from a beta when I made a pivot. And that's something that I really regret doing now", because "none of those companies, including OpenAI, has delivered on that hype". ▶ 14:19 The trip was wonderful, "but I should not have gone there. I should just stay home and finish my own code." ▶ 14:45 His co-host Rasmus Adler Wahlberg names the mechanism: "hype by definition is, um, not reality", and the demo was not widely available. ▶ 15:12 By the year-end episode Martin reports that OpenAI's voice API exists but is "not adopted yet because it's so extremely expensive", and that ElevenLabs had shipped a conversational API doing the same job. ▶ 18:09 The thing he quit for arrived late, dear, and from someone else.

When the ground moves anyway

The pair looks cleaner than it is, because sometimes a nearly finished thing is overtaken through no decision of yours. The morning after OpenAI's first Dev Day in 2023, Martin described the split: either "Holy fuck, I'm going to do so much more things now with AI" or "holy fuck, my business is in ruins". ▶ 1:12 Rasmus explained what had changed for Multiply: a 128K context window, "300 pages of text in one prompt". ▶ 2:28 They had just pushed Claude to production for its 100K window; "But now, I mean, we will just throw that work and just go with GPT-4." ▶ 8:03 They had also built chunking and summarising to keep GPT-4 from being overwhelmed, a bet Rasmus calls a mistake in not acting on trends they had themselves predicted. "But we actually invested in compression." And yet: "we're not throwing it out because it can still happen." ▶ 9:32 Martin had flipped the other way months earlier, from expecting context management to be "solved by larger context windows" to wanting smaller, better-curated ones. ▶ 3:24 ▶ 5:56

The difference from the voice case is instructive. The compression work was finished and running, it was shelved rather than abandoned mid-build, and what overtook it had shipped that morning, with a price list. The voice pivot was triggered by a demo that shipped nothing.

The rule, as the cases state it

Read together, the cases say something more precise than the title. Kill early and cleanly when the blocker sits outside your control, a regulator's timetable or a closing market window. The cost of stopping rises with every month, so "bättre att vi tar det här beslutet nu" (better that we take this decision now). Breakit ↗ A nearly done thing dropped for a competitor's shipped product is arithmetic; dropped for a competitor's promise, it is the regret he still carries. ▶ 8:03 ▶ 14:19 Whatever you stop, carry the code. ▶ 4:04 ▶ 9:32 And the man who wrote the rule admits the company-level version is unsolved: the cautious company lived, the fast one died, and the honest answer is "something in between". ▶ 13:50 What makes it bearable is a habit he states elsewhere: "I've come to associate pain with learning." ▶ 43:56

Worth remembering