Where You Look Is Where You Go

The racing frame

Tobias Lutke's frame for seeing technology shifts early comes from racing, delivered in a context where he and a colleague actually race cars together. The observation is that the single largest difference between a great driver and a mediocre one is where the eyes are pointed. Look at the turn-in point, where the car is about to be, and it will miss the apex; look at the apex, where the car needs to end up, and it arrives there. Every driver knows the corollary: do not look at the tree, or the car will steer into it.¹

Training the eyes to look further ahead than feels natural is difficult in a car, and the claim is that it is equally difficult in business and in life, for the same underlying reason: the instinct is always to evaluate whatever is directly in front of a person.

A worked example told against himself

David Heinemeier Hansson tells his own miscalibration on this exact point as the clearest illustration of the frame. Early AI coding tools were autocomplete-style assistants that repeatedly offered to take over the keyboard, which he experienced as an interruption machine, describing it at the time as an open office reproduced inside his own editor. He evaluated that honestly, disliked it, and then locked in on that experience as a prediction of what AI coding was going to be. Lutke, over the same period, was looking roughly two years further down the road, not at what the tools currently did but at the trajectory their improvement implied, and his argument to Hansson was simply that the current state would not last. Hansson's own assessment of the gap between them is unusually candid: he says he is still not entirely sure what Lutke saw, or why he did not see it himself, and describes himself as a little frustrated on his own behalf for lacking the same early conviction.

The CEO framing

Lutke's internal memo on this, written in 2023 or early 2024 and well before AI agents were doing meaningfully autonomous work, contains what is described as the sharpest job description available for what a CEO is actually for. His stated reason for occupying the role at all is to identify discontinuities in the trend line and reposition the company ahead of them. That is a claim about the purpose of the role, not about its typical daily activity: not operating the business, not allocating capital, not even strategy in the conventional sense, but watching for the specific point where a curve breaks and moving a large organization before that break becomes obvious to everyone else.

Hansson observed the result of this from a board seat at Shopify, reporting that the company's internal tooling was ahead of what was commercially available anywhere, citing an internal tool that aggregates customer feedback across channels so product managers can query it directly; he asked whether it could be commercialized and was told no, it was built for internal use only.

The difficulty worth naming

The most useful part of the story is not that Lutke turned out to be right. It is Hansson's own observation about how hard the call should have been to make at all. Most of the people who saw the AI shift this early already worked directly in AI. Lutke was running an enormous company sitting on a meaningful share of world e-commerce, with countless competing demands on his attention, and still kept his focus on a trajectory two years out. Hansson describes being almost shamed into catching up by someone with far less attentional room available than he had himself. The implication cuts against the usual excuse: being occupied by the present is not what prevents foresight, since the person with arguably the most present to be occupied by managed it anyway.

Why it matters

The frame gives a diagnostic anyone can run on themselves: ask whether the thing being evaluated is the artifact currently in front of a person, or the trajectory that artifact implies, since Hansson was doing the former and mistaking it for judgment. It also names a specific and common failure, extrapolating a technology's ceiling from its current, worst ergonomics, since the interruption-prone assistant of 2023 was a real and reasonably evaluated experience, and reasoning from it was still the wrong conclusion. And it gives foresight an actual mechanism rather than treating it as an innate gift, since eyes are trainable in a car, which is the entire reason the metaphor carries useful weight.

Tensions and open questions

The frame is only clearly visible in hindsight; it cannot itself tell a person which apex is the real one, and plenty of people have looked two years ahead at other technologies and steered directly into a tree. No filter is offered for telling the difference in advance, and Hansson admits he still cannot fully reconstruct what Lutke actually saw at the time. Converging on foresight also does not guarantee converging on the same conclusions: Hansson caught up on the underlying capability and landed on a note of caution about AI removing a useful constraint, rather than on unqualified enthusiasm for it, even though both he and Lutke were looking at the identical trend line and arrived early relative to their peers. Everything about the Shopify details here is also secondhand and friendly, since the memo's contents and the internal tool's effectiveness are relayed as one board member's own impressions in an interview, and there is no visible record of technology calls this same method got wrong, which is the population that would actually be needed to judge it.