Hello friends,

Last Friday, I talked separately to two old friends (both writers) about writing with machines. For a proposal, I’d just unscientifically reviewed my writing on uncertainty in the last year and compared it to what I wrote five and ten years ago. The recent stuff is worse, though I’ve spent decades clarifying what I’m writing about.

The problem I’m grasping to define isn’t the mediocrity of LLM-generated prose. It’s that machine-made prose looks too finished right out of the box. When something appears so done, my brain shifts from critical assessment (“is this the argument I’m trying to make?”) to proofreading (“should I break this sentence into two sentences?”). I’ve become increasingly insensitive to when the substance in the prose subtly (or even grossly) misses the point I am trying to make.

Sitting with incomplete and emerging thoughts, starting over, crossing out — that slow, uncomfortable work was where the thinking used to really happen. When an idea on the page wasn’t quite right because I didn’t yet know what I meant, struggling to articulate it was how I figured out what I meant.

Now, when writing with machines, I increasingly prompt them with rambly voice memos which take barely any time to dictate. LLMs transcribe the voice, return some bullet points which I accept or modify, then take those points and turn them into slick prose filled with signposts of conceptual coherence: “moreover” to suggest addition, “however” to signal contrast, etc. These create prose with the appearance of structure arising from reasoning. But I haven’t had to do much to build up that reasoning, so the resulting paragraphs end up with few of the ideas I am (in theory) using them to develop.

The better the machines get at producing clean sentences for me, the harder it is to see where my own thinking needs more work. This problem is thornier for being incompletely articulated.

More on this soon, to pick up threads from previous essays on the inadequacy of output indistinguishability as a criterion for LLM quality, the dangers of the seductive mirage of AI, and the resulting need for AI tools with UX intentionally designed to help users learn to think critically.

Share

Writing

Over the last two weeks, I wrote two essays on how organisations can design themselves for uncertain times (not just risky ones) by focusing on tacit knowledge, better motivation — they’re meant to complement the first in the series (on improving hiring by using negotiated joining and open-ended roles) which I wrote a couple of years ago.

[on tacit knowledge] Learning what matters: The most important knowledge in any organisation—what makes your work distinctively yours—can’t be written down or taught through training programmes. It’s tacit knowledge that must be learned through direct experience. Most organisations try to solve this by writing better documentation, which fails. Organisations that successfully teach tacit knowledge embed learning into everyday work through three mechanisms: making work public and concrete, focusing feedback on outcomes rather than processes, and using concrete examples as anchors. This approach appears chaotic but produces faster, more reliable learning than conventional training. (11 Jan, 2026)

[on better motivation] Desperation by design: Innovation requires uncomfortable unfamiliar work, but people instinctively resist discomfort. Incentives and motivation programmes don’t work for innovation either. Instead, successful organisations use productive desperation: deliberately committing to projects calibrated to be just beyond current capabilities, with real possibility of failure and no escape route. This forces learning, role flexibility, and new work patterns, but requires careful and intentional design in the form of team consultative calibration and rhythmic cycles of desperation and recovery. (11 Jan, 2026)

[on improved hiring] Unfrozen from the start: With conventional hiring methods, employees expect clear and stable roles, creating organisations that are “frozen” and inflexible from the start. Using open-ended roles with negotiated joining is an unconventional way to build an organisation that starts unfrozen and is good at uncertainty work that isn’t well-understood yet, where rules and operating environments are always changing and being disrupted. (17 Oct, 2023)

Elsewhere

1. “… the Smart Brick can sense distance, direction, and orientation of other Smart Bricks near it when you’re building.”

2. “Many people resist being kinded unless they are in dire life-threatening need. But a kindee needs to accept gifts more easily.”

3. “All you need is specs?”

See you soon,
VT

Hello friends,

Happy new year. Recent events compel me to remind you that idk (the world’s first and only tool for learning how to be productively uncomfortable) is once more available in Europe and the US.

I’ve been in various virtual and meatspace places in the last two weeks:

Watching a New Year’s Eve coulibiac being constructed: A coulibiac is a Russian imperial dish of fish cooked inside a pastry shell and insulated by a layer of herbed rice, boiled eggs, and precooked vegetables (on similar principles to a beef Wellington). It takes much time and planning. Not all laborious foods are imperial, but laborious foods are much more common where labour is both cheap and abundant—hence the many Peranakan dishes requiring large quantities of fine knifework, and all tiny-format stuffed vegetables and pastas.

In a big box grocery store observing “baby carrots” being restocked: Industrial processing can’t afford to be discriminating at the margin when handling highly variable products. Agricultural products vary too much for processing lines to trim frugally—slowing down enough to peel every carrot only as much as necessary would make industrial carrot economics unviable. So baby carrots are pieces of big carrots lathed down into tiny cylinders designed to be small enough for all shape and surface variation to be cut away on nearly all carrot pieces. This lowest-common-denominatorisation is the price we pay for standardising outputs when inputs aren’t standardised.

Lurking on a WhatsApp chat of aggressive first-adopters of LLMs: Prompting an LLM well (i.e., to make the LLM generate output conforming to the prompter’s intentions) forces the prompter to make implicit reasoning explicit. This is similar to what happens when teaching a human how to do something (e.g., cooking an omelet or suturing a wound), but with a crucial difference: with an LLM you can’t rely on the learner’s discretion and judgment. When you teach a human, you can hope for them to exercise judgment under ambiguity, but LLMs can’t (yet) be relied on exercise subjective judgment at all (what I’ve called the meaningmaking lens on AI). So, with an LLM, prompters should aim to articulate every step of a reasoning process (what I’ve called a reasoning scaffold) — but even aggressive first-adopters don’t seem to be using this frame to think about what they’re doing.

Driving by a town dump when a truck filled with cheap toys pulls up: A replacement-oriented society defaults to designing objects to be lower-cost because they are intended to be disposed of and replaced instead of being repaired. A maintenance-oriented society defaults to designing objects to be more expensive but easy to take apart and clean or replace broken pieces of (we sometimes call these repairable systems). Maintenance-orientation was the default when our machines weren’t as good as they are today; the stuff made by maintenance-oriented societies was nice in ways we can’t describe but can usually sense.

Share

Writing

Judgment from the ground up: Critical thinking is foundational for making decisions that require subjective judgment. People learn how to do subjective decisionmaking through practice. Unfortunately, organisations increasingly reserve subjective decisionmaking for senior members, so junior members get neither practice nor support in learning how to do it; this also means that they don’t get practice in the kinds of critical thinking needed for good subjective decisionmaking. This leads to succession crises, inefficient bottlenecking in decisionmaking, and poor decisions made by senior members who are far removed from the operational realities that junior members are exposed to. Organisations must therefore redesign themselves to provide junior members with opportunities and tools for learning how to think critically when making subjective decisions. To build capacity for subjective judgment from the ground up, organisations must create low-stakes practice opportunities where juniors can make real decisions with real but limited consequences, learning the same skills that senior people use for high-stakes choices. (3 Jan, 2026)

Elsewhere

1. “‘The biggest misconception is that small business owners in developing countries will be the last to adopt AI. I’m seeing the opposite,’ said Dutt. ‘They are adopting it first because for them, even a small improvement in efficiency makes a big difference.’” (They’re building what I call Boring Tiny Tools.)

2. “… nasal oxytocin is going to be the Ozempic for autism …”

3. “… as every good quarterback comes to know, the discursive, denotative nature of language can impede or get in the way of the process itself. I once, and only once, took a Tai Chi lesson. The instructor looked amazed. “How did you learn these movements?” he asked. “Pitching hay and ice skating” I replied. “Ahh, yes”, he exhaled with a nod, a smile and a wink. I saw no need to come back … Never touch your tongue to a frozen shovel. Keep it firmly in cheek.”

See you soon,
VT

Hello friends,

Crass commercialism: idk, the world’s first and only tool for learning how to be productively uncomfortable, is once more available in Europe and the US! Read more below about why cultivating productive discomfort is essential.

In early June this year, I was trying to write my first software specification for a vibecoding session. I needed to build a tool I needed for a consulting engagement. I’d written product requirements before at Google, decades ago. But this felt uncomfortably different. The task was ostensibly simple: describe what I wanted a tool to do so that an AI could write the code to do it. In practice, the entire frame had shifted. What worked for vibe coding with AI bore almost no resemblance to the PRD discipline I’d learned in a corporate environment.

It would have been easier to hire someone. Modern life encourages us to pay someone else to handle our discomfort. A nontrivial part of the modern economy is built on the premise that people will pay to avoid discomfort. Food delivery so you don’t have to cook. Task outsourcing so you don’t have to learn new skills. Template solutions so you don’t have to think through problems from first principles. This systematic elimination of discomfort from daily life has made us remarkably efficient at avoiding things we’re bad at. But it’s also made us progressively worse at dealing with discomfort at all.

The problem is: Uncomfortable situations are usually the ones with the most potential for learning and innovation. If we continually avoid discomfort—and modern society increasingly pushes us to do exactly that—we lose our ability to do uncomfortable things. The muscle that lets us push through initial confusion, sit with not-knowing, and slowly build competence in unfamiliar territory weakens from disuse.

So I didn’t hire someone to write my software. Instead, I structured the discomfort deliberately. I started with scoping tiny projects with low stakes so I could abandon anything that didn’t work. I built a tool to process receipts for my business accounting. Another to organise research and output production. Another to make my tangle of travel bookings reliably interrogatable. Each tool taught me something about writing specifications that an AI tool could actually work with. And each small success built my capacity for slightly more uncomfortable next steps.

Over the last six months, I’ve built a bunch of stuff. Some of it helps me run my tiny consulting business more reliably. Others I have bigger intentions for. (Such as the prototype of a tool for helping students learn how to think critically while using AI tools). Building all this required pushing through the initial discomfort of not knowing how to do something, and structuring that discomfort in ways that made learning possible.

I eventually realised that I’d implemented a version of my own training programme for productive discomfort. Not all discomfort is equal. Some discomfort is just suffering—pointless, grinding, soul-destroying work that teaches you nothing except that you hate it. But productive discomfort is different. It’s the discomfort that comes from being at the edge of competence, doing something slightly harder than you’re comfortable with, in a context that’s designed to support learning rather than punish failure.

Modern life seems to eliminate both kinds of discomfort indiscriminately. We lose access to the productive kind along with the pointless kind. This is bad, because productive discomfort is how we grow. It’s how we develop new capabilities. It’s how we stop being helpless in the face of unfamiliar situations.

Back in 2021, I developed idk. It’s a series of prompts that help users practice being more comfortable with productive discomfort. Each card presents a small challenge that creates discomfort that is intentionally designed to expand horizons, change perspectives. The people who have bought it seem to like it.

Earlier this year, a catastrophic logistics failure. My third-party logistics provider lost all my stock in a warehouse transfer. (They also got bought by another company and became five times more expensive.) This became a masterclass in productive discomfort because I was in the middle of several other projects. The easy thing would have been to just shut down idk sales and move on. But I decided to push through the discomfort of abandoning a logistics partner I’d worked with for years to track down recoverable inventory, rethink how I distribute and fulfill orders, and rebuild my supply chain from scratch. As of this week, idk is finally back in stock in the US and Europe. Not in time for Christmas delivery, but idk would make a good gift for the new year…

The logic here is the same whether you’re learning to write software specifications, practicing discomfort through card prompts, or navigating a small business catastrophe. We need deliberate practice at being uncomfortable. We need to design systems and situations that put us into situations of productive discomfort which are carefully structured, safe enough to learn from, and which build progressively on our slowly growing capacity to handle uncertainty.

Otherwise we just keep paying for convenience until we can’t do anything uncomfortable at all.

Share

Writing

Boring Tiny Tools — On why digital transformation fails when organisations try to do it with Big Fancy Software instead of Boring Tiny Tools: narrow-scoped tools that fit seamlessly into existing workflows. I argue for a fundamentally different approach to product development that focuses on designing software that only replaces work that requires no human judgment, serves tiny userbases, offers users exactly what they need, and uses only off-the-shelf components. (22/12/2025)

Doing nothing: a reading list — A short and non-exhaustive list of ten things to read before spending time with nothing much to do. In different ways, these things all show how removing one kind of variation gives other kinds of variation space to emerge and become visible. (15/12/2025)

Elsewhere

1. “An increasing number of ‘crazy rich’ clients are commissioning dozens, or even hundreds, of U.S.-born babies with the goal of ‘forging an unstoppable family dynasty,’ he said.”

2. “I’m not going to sit on the money and live from interest. I want to add another zero to my net worth and buy more Picassos.”

3. “For the first time in history, weirdness is a choice.” (I’m not so sure about that, tbh.)

See you next year,
VT

Hello friends,

Back now in Singapore, where it’s surprisingly cool and not unbearably humid.

Last week in Marseille, I wanted a book that I knew was in one of the dozens of boxes scattered around the apartment. (Jasper Morrison’s The Good Life.) I half-remembered last seeing that narrow, pale blue, clothbound spine in a corner of a green and brown carton sitting on tomettes. The tile narrowed it to two rooms, and the carton colour to one of maybe three boxes. This was sense memory combined with spatial memory—an ordinary act of retrieval that depends on our evolved adaptation to working in 3D space.

This kind of situated memory comes naturally in physical environments but is nearly impossible to replicate online. Virtual whiteboards try to mimic navigable physical space, but so far almost all of them do it terribly. VR and AR will eventually solve some of these problems, but the current implementations aren’t good enough yet.

The affordances and constraints of physical vs. virtual space have been on my mind because I recently redesigned—from the ground up—a strategy workshop I’ve run in person for years. Yesterday at 6am, I delivered a shortened, first fully online version for a team in Australia.

The redesign reinforced my growing belief that online workshops can’t simply be in-person workshops transposed to Zoom. Doing them well online requires rethinking the workshop and interaction mechanisms entirely, to both work around the disadvantages of not being in the same physical space and also exploit the technical and interaction possibilities that working online affords. I no longer especially want to focus on making online workshops tolerable substitutes for in-person ones. I’m now much more focused on designing online workshops that are worth doing because they’re online.

This week’s writing is about some preliminary answers I’ve come up with, about what we lose when we work online (spatial memory, body language, the low friction of moving physical objects), but also what we gain (high-speed elicitation, comprehensive documentation, computational analysis of collective thinking).

Share

Writing

Making online workshops great … from the ground up. Online workshops rarely work well, and we usually do them mostly because an in-person workshop would be too expensive/infeasible. This is the wrong way to frame the problem. The right way is to ask how to redesign online workshops from the ground up to be effective not in spite of but because they’re online. Virtual spaces destroy our spatial navigation abilities and make reading participants nearly impossible, but they also enable transformative capabilities: high-speed elicitation, comprehensive documentation, and real-time machine-supported analysis of collective thinking that’s essentially impossible with walls of paper post-its. After years of refusing to run workshops online, I’ve spent the last few months fundamentally redesigning my approach from the ground up—discrete virtual spaces, explicit participation scaffolding, and computational tools that support rather than replace human judgment. The mindset shift we need is for workshop design to start thinking about online spaces as being fundamentally different from meatspaces, and especially in thinking about designing for their distinctive advantages. (28 November, 2025.)

Elsewhere

1. “I can see why all these people without heads wouldn’t be good for a community, and might be a cause for abandonment.”

2. “The design and slicing phases demanded additional time due to the boat’s substantial size and intricate geometries.”

3. “That’s equivalent to about 250 feet on an e-bike, or around five and a half seconds running a microwave.”

4. “Focusing on high-resolution modelling is dangerous not only because we have no answer to the question of when a model is sufficiently realistic.”

5. “Speaking outside his chateau in France, former Brexit Party and UKIP MEP Coburn answered ‘no’ when a BBC journalist asked him whether he had ever been paid to give a speech to promote pro-Russian campaigners.”

See you next time,
VT

Hello friends,

A few weeks ago, I rented an all-electric vehicle to drive up from Marseille into the Massif Central — north on the A7, then west through the Monts d’Ardèche, up onto the Mézenc plateau, and then into the Livradois-Forez. After leaving the A7 with its profusion of charging stations, I entered the mountainous plateau in the centre of France and an entirely different operational regime.

This volcanic plateau is in the middle of a band of low-population departements running from the northeast to the southwest — the empty diagonal. The plateau has few inhabitants and sparse charging infrastructure, barely any of it high-speed.

Renault’s predictive range analytics proved unreliable when confronted with these steep mountain roads, repeated elevation gain and loss, and sharp temperature changes. The range estimates changed so dramatically that I might as well have been planning the route around an infrastructure of phantom charging stations. Under these conditions, optimisation is a mug’s game. I was glad, as a believer in excess capacity in times of uncertainty, to have planned the route around much more range buffer than the charge-planning apps suggested.

EVs change the logic of long-range trip planning. With a combustion engine, you can be cavalier about fuel and optimise routes for speed or distance — it’s hard to run out of fuel because gas stations are everywhere and refueling takes five minutes. But with an EV, refueling locations and times become a central consideration. Charge now at this slow 22kW charger and add an hour to the trip, or use the fast charger 80km (and a 45 minute detour) away that might be occupied or broken? Having a lot of excess capacity is crucial.

The UX of the charging ecosystem is also still astonishingly bad. The charging network apps know where and how fast charging points are, but are terrible for optimising charging around anything else you might want to do while the car charges — like stop in a hardware store or find a swimmable river. The handoff between apps and in-car navigation systems is terrible; planning a route and actually driving it requires juggling several different apps and systems, none of which talk to each other properly.

Here’s an underexploited business opportunity: A decent rural restaurant with a few moderately fast chargers (22kW, which are more likely to be feasible with existing rural electrical supply infrastructure) would clean up. Not fast food or fancy food, and it has to be far from the autoroute — but make it a real sit-down place, a classic routier or bistrot with a formule, where spending 45 minutes to an hour would be a pleasurable obligation.

Share

Writing

1. Public value and public strategy. Why borrowing private sector strategic thinking and values undermines public strategy — and a practical tool for using tradeoffs to develop robust public strategy that serves diverse stakeholders over long time horizons. 13 November, 2025.

2. Now-ish. It’s been over 3 months since I updated the now page. At the end of July, I went back in Singapore for a Centre for Strategic Futures Fellows visit to think about overlooked strategic issues. Adelaide in August, where I installed a temporary exhibition on experiencing not-knowing, and ran strategy workshops for Adelaide University leadership and SA Water. Singapore in September for a top secret futures project (embargoed until early 2026); then Tunis to run a tradeoffs-focused strategy workshop for UNDP Tunisia. Japan in October for business development, to see my old supervisor, and to walk part of the Kumano Kodo; then Berkeley to present my research on reasoning scaffolds. Auvergne in November to retrieve tools from the barn at the old house, then London to run a masterclass on how to think about AI deployment in traditional industries. 15 November, 2025.

Elsewhere

1. “In this work we construct a polyhedron which is provably not Rupert, thus we disprove a conjecture from 2017. We also find a polyhedron that is Rupert but not locally Rupert.”

2. “Thanks to its proprietary magnetic technology, the Kaite Pad also offers full or partial erasing capabilities.”

3. “We can’t get enough of the boneless design and the secure feeling of holding it tight.”

4. “You can securely tighten anything to your drywall now!”

See you next time,
VT