Oscar Barlow

Zones of want

2026-04-14T00:00:00+00:00

Some things you want, and you can just have them. You want to read a magazine, so you go to the shop and buy one. Then you take it home and read it. Done. Easy. A little humdrum.

Other things you want, and you can’t have them right now. Actually, you can’t even have them if you make a sound plan and immediately take a few steps forward. You want a promotion, not your boss’s job but their boss’s job. A want, once identified, becomes a source of discomfort. To get the job you want, you first need to get your boss’s job, or something like it. And so this discomfort is productive.

Or perhaps you don’t have to do that first? Perhaps you can find a way to just get the thing you want? Worth exploring. Moving things between zones of want is a great trick, if you can pull it off.

Next, you may want some things that are just impossible. You want to have made different choices ten years ago so you can be a professional musician now. You choose to engage with reality here, or not. Why didn’t you make those choices back then? When you find out the reasons, you might continue to want this thing - or you might not. Once you engage with the reality of the want, you discover that you do not have this want.

This is actually similar to refusing to engage with the reality of the want. In this case, you do not actually want something. Sticking with this want will remain uncomfortable for you. It’s peculiar that you can end up at the same destination by two opposing routes (you engage, the want dissolves; you don’t engage, the want is unreal). But here we are.

This brings us to our final zone. You want something, it’s impossible, you’ve engaged with the reality of the want, and you still want it. I want climate change to be fixed - this belongs in this zone. As before, this unfulfilled want causes me discomfort, and this is discomfort that I choose in keeping wanting this. Amor fati.

You can respond to wants by taking action or changing your attitude. As we progress through these zones of want, first to last as I’ve laid them out, I’d suggest you dial up your focus on attitude and dial down action.

The above is somewhat informed by Covey’s Circles of Concern and Influence exercise. But as for eudaimonia, you’ll need to look elsewhere. These zones are pragmatic, not normative.

Henry Ford, paper stacks, and three buckets of AI work

2026-03-03T00:00:00+00:00

Ford’s workers hated his assembly line¹ and in their diaries expressed how unnatural it felt, and how they resented their deskilling. It took a lot of trial and error to arrive at the assembly line technique². Ford’s continuous experiments with production techniques almost bankrupted him several times over³.

Comparisons between software engineering and manufacturing production are common enough⁴, so let’s focus for a moment on the experience of the workers: how bizarre, and contrary to common sense and good practice the assembly line felt.

Now consider this on digital transformation - the quote is about government services but it’s broadly applicable:

Much of the economy has changed in the last 70 years or so, but broadly speaking, [public] service delivery looks a lot like service delivery did in, say, 1950. We had a paper-based workflow then, and the government is pretty good, relatively speaking, at architecting workflows where: we need a floor of clerks, we need a front office, the front office passes paper to the floor of clerks, they get 10,000 papers on Monday, we successfully drain that stack within six weeks, and everything is operating as designed. And if you asked in 1950, “Give me a real-time count of how many papers are in process,” they’d say: one, that’s a crazy thing to ask for, but two, if you want an approximation, go look at the physical floor and count how many inches of paper there are in the to-be-done box. That’s your count. These days we have computer systems that are pretty good at doing real-time counts. But because the institutional memory of the system is that tracking the progress of an in-flight application is just strictly beyond the bounds of what’s materially possible — and I can’t even ask a web app to look at how many inches of paper there are because the paper doesn’t exist anymore — I’ve actually lost fidelity by moving to the computer system. The management of these programs often doesn’t do things that are pretty table stakes for private industry, which came up in a post-computer age. Private industry writes SQL queries because it has business analysts. It’s not that the government has no business analysts, but I often think there is a crushing undersupply of them relative to need, and a management culture that doesn’t really make data-driven decisions — in a number of places, both in the management of the agencies and in the legislative branch⁵.

Emphasis mine.

Again I want to focus on the experience people involved in this work - we’re thinking about managers rather than labourers now. By adding computers to their business processes, they believe they have lost visibility and so control of those processes.

Here are three ‘buckets’ categorising where generative AI can be integrated into an organisation:

Personal productivity
In the loop in back-office processes, such that teams have an ‘AI teammate’
Into customer-facing products or important internal systems by teams of engineers and data scientists

We’re totally fine with 1 & 3. We know how to train people to use Deep Research, or wordsmith their emails with Gemini in GMail. Likewise, any organisation that has been shipping data science projects to production for any amount of time should not have difficulty shipping an AI project (and many of them have).

1 & 3 also don’t need to challenge our production techniques. It is possible to use AI to do your work pretty much like you used to but faster, and 1 & 3 are where we see that⁶.

We have no idea about 2. We have no idea what coordinating the work of groups of humans who are each collaborating with one or several peculiarly intelligent machines looks like. We can be pretty sure, however, that to the people doing it it’s going to feel bizarre, contrary to common sense, and like a loss of visibility and control.

I suspect it will emerge out of 1, and we should amp that up as much as possible.

Eschner, K. (2016, December 1). In 1913, Henry Ford introduced the assembly line: His workers hated it. Smithsonian Magazine. ↩
Weber, K. (2013, October 1). The Moving Assembly Line Turns 100 ↩
Hounshell, D. A. (1984). From the American system to mass production, 1800-1932: The development of manufacturing technology in the United States (pp. 230–273). Johns Hopkins University Press. ↩
ICYMI: writing software is (was?) a creative pursuit so full of irregularities that it cannot be automated, but delivering software artefacts to a live environment - we even call it ‘production’ - sure as heck can be, and doing so propagates many beneficial changes back through the software delivery lifecycle. ↩
McKenzie, P. (Host). (2026, February 26). Understanding government procurement, with Luke Farrell [Audio podcast episode]. In Complex Systems. ↩
It’s possible to work pretty much like you used to, but it seems likely that 3 will change profoundly in the near future. See: Tane, B. (2026, February 20). The software development lifecycle is dead. Boris Tane. ↩

At some point, you need to do formal reasoning

2025-09-17T00:00:00+00:00

When you are building software you are building a formal system, and that means at some point you need to do formal reasoning.

Let’s start with the stops vs. legs example from Eric Evans’ 2019 talk ‘What is DDD?’ If you haven’t seen the talk it’s worth watching on its own merits.

Just to briefly recap: building a software system for logistics around the concept of ‘cargo’, ‘stops’, or ‘legs’ is a significant choice, and has big ramifications for the code you produce. Actually, different parts of the system are highly likely to treat different concepts as primary, which can create complication when you have to translate between them. And that’s before we’ve come to the messiness of the real world - both the general rules of how cargo/stops/legs are to be treated, and likely numerous exceptions to these rules, what developers normally call ‘edge cases’.

Bringing this back to using AI in software development, here’s a quote from Evans’ talk:

We’re used to just hearing words and, you know, we get the gist of what’s being said, you get the meaning of what I’m saying but [we’re] not paying enough attention to the words themselves…

When you look at a popular SaaS product that’s been cloned by someone prompting a coding agent, that’s what you’re looking at: some code that has been produced by a machine that gets the gist¹ of what you say. The code is functional but in reality not yet useful, because it is only partially grappling with the general rules, and not at all with edge cases. That’s as far as gist can take you.

To move beyond ‘gist’, GitHub have published an open source toolkit for spec-driven development:

Where we’re headed

We’re moving from “code is the source of truth” to “intent is the source of truth.” With AI the specification becomes the source of truth and determines what gets built.

Grand, but the emphasis on intent is not quite right in my view. Intent can be better or worse defined. If you are going to define your intent well - really well - you are writing a formal specification. A written specification that allows you to precisely define your intent may as well be code. There is no getting around writing detailed formal specifications for formal systems, if those systems are going to have contact with the real world. And if you’re writing a formal specification, you are doing formal reasoning.

Let’s connect this to skills loss. A commonly-voiced anxiety about LLM-assisted coding is that developers will forget how to write code. This is probably true, but it’s not worrisome. A precise, formal spec interpreted by an LLM is just another level of abstraction. As developers have begun working with higher-level languages, they’ve forgotten things about how lower levels of the stack work. This is a normal part of technological development. And supposing the developers do need to work in lower levels of the stack in future; they will learn how to do so in much the same way as they do today - including, by using LLMs.

Let’s take a moment to acknowledge that although in this context, I’m presenting this negatively, it’s still miraculous. ↩

Risk, uncertainty and innovation

2025-09-02T00:00:00+00:00

These 3 articles from Vaughan Tan ¹ about (formal) risk, uncertainty and innovation are great:

They are well-considered pieces, dense with ideas but nonetheless accessibly written.

Précis:

Situations of formal risk as defined by Tan, are quite rare. If you think about a situation of uncertainty only in terms of risk, you will misunderstand it. This may make you complacent, or it may mean you miss opportunities. Financial services professionals should be particularly interested in this point.
Attempts to stimulate innovation usually understand innovation very narrowly, and in particular don’t pay enough attention to the notion of context in innovation. Innovation and uncertainty/’not-knowing’ are intrinsically linked, and so the best way to stimulate innovation is to develop organisational comfort with not-knowing.
Uncertainty/not-knowing can be usefully broken down into 4 types (about actions, outcomes, causation and value). Knowing which type of uncertainty you’re dealing with enables you to use appropriate mental tools.

Tan decries organisations that ape structures and processes from large tech firms as a means to produce innovation - I agree - and recommends instead developing an ‘uncertainty mindset’. He’s authored a book to help you do this.

However I’m reminded of something from my days as a marketing manager. Marketing is also described as a mindset, but clearly there is more to how marketing works than just mindset. Marketing managers who move between firms will find familiar processes and structures, etc. Likewise, it seems to me probable that there’s useful management work to do on innovation beyond cultivating mindset. There may even be processes and structures that are transferable between firms.

What we are talking about here is practical epistemology, the techniques we use to come to know things. Sifting through an uncertain and changing morass of information is a skill you learn studying the humanities. This is the thesis of the entertaining and informative blog A Collection of Unmitigated Pedantry - I’d recommend Collections: What Do Historians Do? – A Collection of Unmitigated Pedantry to see this expounded, or Collections: How Many People? Ancient Demography – A Collection of Unmitigated Pedantry to see it in action.

I came to Tan via the email summary of How to Run Smart Experiments When You Just Don’t Know - Commoncog. I expect the Speedrunning the Idea Maze course, offered by CommonCog, engages with these ideas some more. ↩

The cosmic weevil is in the wrong place on your organogram

2025-08-13T00:00:00+00:00

Imagine a lone botanist studying high-altitude lichens on a remote, windswept plateau in the Atacama Desert. One day, she spots something impossibly out of place: a beetle. It’s breathtaking, with a carapace that swirls with the deep blues and purples of a distant nebula, dotted with tiny, shimmering silver flecks that look like stars. It has a long, elegant snout, a key feature of a weevil. She nicknames it the ‘Cosmic Weevil’.

The botanist consults with colleagues and they form the certain opinion that this is a new species from the genus Cyrtotrachelus, the “Giraffe Weevils.” This genus is known for its members’ unusually long snouts and their diet of bamboo shoots. While finding a bamboo-eater in the world’s driest desert is bizarre, the morphology is undeniable—the antennae placement, the leg structure, and especially that signature snout are a perfect match for Cyrtotrachelus. The working theory is that it must be a relic species from a time when the Atacama was a lush bamboo forest.

Unfortunately, the genus Cyrtotrachelus is looking pretty full, and can’t really accept new species. The botanist and her colleagues don’t want their scientific discovery to languish, so they take a pragmatic decision. They learn of the genus Pandeleteius, a group of hardy weevils often found in arid environments. This provides a crucial link: while the Cosmic Weevil’s diet is a mystery, its discovery in the Atacama Desert aligns with the habitat of Pandeleteius.

The problem, of course, is the snout. Pandeleteius weevils are broad-nosed and have short, stout rostrums—the complete opposite of the long, graceful snout of the Cosmic Weevil. But while the snout is highly anomalous, other features like the leg structure and the texture of the elytra are not entirely inconsistent with Pandeleteius. The botanist and her team convince a Pandeleteius expert to co-author a paper with them, that places the new species in this genus. The Cosmic Weevil is eye-catching, and the Pandeleteius expert is pleased to be involved with it - a species such of this can help attract grant money and sympathetic public attention, as well as being scientifically interesting in its own right.

An automobile parts manufacturer, with a long history of process-driven efficiency, decides to create a new portal for their wholesalers. The project is a genuine strategic opportunity: the goal isn’t just a new front-end; it’s to use predictive analytics to anticipate wholesaler requests. The project lead assembles a brilliant cross-functional team of data scientists, AI engineers who specialise in edge computing, and a couple of contract front-end developers. They propose using LLMs hosted on edge nodes to accomplish this—and maybe even to extend the project to allow the LLMs to autonomously take decisions, once the core proposition has been proven. It’s a technically innovative approach that promises to boost revenues in a strategically important value stream.

Unfortunately, the company dictates that all projects are categorised by their primary output. Although this looks a bit like a data science project, the fact that it involves a front-end means that it must be categorised as a front-end project. And the front-end team is beyond swamped right now.

The project lead announces to the team that the project must be postponed, indefinitely. It simply would not be proper for this project to be allowed to proceed outside of its proper place in the org chart.

When you choose to understand something as belonging to taxonomy or as part of a portfolio, you are making a trade-off. Taxonomies are highly legible, but inflexible. I needed to introduce the absurd concept of a genus being ‘busy’ to make the story of the cosmic weevil work¹. Portfolios are flexible, they are less constrained by organising logic then taxonomies. But the more a portfolio tends towards being a grab-bag, the harder it is for outsiders to make sense of. My particular interest is in organisational contexts, where I suspect projects are handled as part of a taxonomy when they should be considered as part of a portfolio. It can land you in absurdity. Where is the cosmic weevil on your organogram, and what’s the cost of leaving it there?

I wrote this post after reading Make an Org Chart You Want to Ship — Advice from Linear on How Heirloom Tomatoes Should Inspire Team Design and letting it roll around my head for a while.

There is such a thing as taxonomic revision, as in the case of Scrophulariaceae, but is rather different from what I described. It’s a lot more robust, for one thing. ↩

Subset saturation law of group chats

2025-06-18T00:00:00+00:00

Suppose there is a social group that likes to communicate via group chat. We’ll call its members Akira, Bilal, Chiara and Dan and refer to them as A, B, C and D for brevity.

The subset saturation law of group chats states that eventually, group chats will splinter off the group chat corresponding to all possible subsets of the main group. So in this case given that A, B, C, D is the main chat, the following chats will eventually appear:

B, C, D
A, C, D
A, B, D
A, B, C
…¹

Splinter chats form to arrange a surprise birthday for one of the main group members, arrange an outing when one or several members aren’t available, organise an activity in which only some members of the group have an interest, and so on. In the example given, the main group has only 4 members but this is theoretically extensible to chats and social groups with large numbers of members.

One might take the degree of subset saturation as an index of the strength of ties between members of the group. Larger groups would probably score lower on such an index compared to smaller groups, unless the large group had been around for a very long time. That makes sense - one would expect that smaller groups, or well established groups, would be more tight-knit. Alternatively one might take

1-1 chats should be included in this list but have been omitted. They aren’t what people normally call a group chat, and the list is illustrative enough as it is. ↩

Characteristic change in business processes from AI transformation

2025-05-01T00:00:00+00:00

Scott Werner notes in The Coming Knowledge-Work Supply-Chain Crisis that business processes for managing knowledge work ill-suited the amount of throughput it looks like we’ll get from using AI:

… our tools aren’t designed for the volume of work AI can generate. In AI Programs While I Sleep, you can see that I am already underwater with hundreds of AI-generated PRs to review. Our code review tools are designed for reviewing at most 5-10 PRs a day, not 50. You can also see a similar pattern emerge in the other videos having to do with managing user stories, doing product acceptance, and test case validation. Our tools are designed for orders of magnitude less work.

Making ‘orders of magnitude’ more work happen is the productivity boost that many expect from AI. But as well as volume increasing, I think our outputs are about to get a lot more varied - our processes are going to get fuzzier.

Consider a process like employee onboarding. CRUD operations against a relational database are a reasonable mental model for this kind of process. For each task in the employee onboarding flow, you can imagine a table in a database with its having its fields updated; induction_training_completed_date, department_fk, and so on. To be clear, I’m not arguing that every business maintains such a database, but rather that this is what business processes are like. This is how we think about getting things done. And this isn’t surprising, given that software systems and business management practice evolved alongside each other during the latter half of the 20th century.

This kind of process promises a lot of regularity. Of course there’s some variation - some employees take longer to complete the induction training, while others never complete it at all. Most employees get assigned to a department, but exceptionally some employees don’t. Businesses (and database designers) expect this and know how to design their processes and software systems to manage this kind of variation.

Adding more volume to this kind of process certainly poses challenges as described in the blog quoted above, certainly. But now let’s consider adding an AI tool in to our imaginary employee onboarding flow. In this case, someone has decided to create a NotebookLM notebook to help employees onboard instead of putting them through induction training. Instead of watching training videos and answering quizzes, now new joiners chat with an AI, answering questions as they come up. Each person will ask different things and get different answers. At this point, it becomes a lot harder to know when the induction training has been completed. Actually, does the question even make sense any more?

You could still track when the notebook was assigned or first accessed, but that’s not really what you’re interested in. In the CRUD-esque world you had some reason to believe that all new joiners went through a regular process - you don’t any more. The outcome of this particular process has become less tractable, less defined, and what the process consists of - what new joiners actually encounter - is a lot more varied too. This is a characteristic change in our processes, not just a volume change.

Of course, this kind of change won’t be limited to induction training. It’s just a useful example.

LinkedIn and solipsistic peacocking

2025-04-23T00:00:00+00:00

Emily Gorcenski observes about LinkedIn:

I’ve been on the internet a long time. Usenet. Forums. Email lists fighting over moon landing conspiracy theories. Facebook. I had 100k+ followers and a checkmark in the heady days of Twitter. I’ve become an expert in the alt right social media ecosystem. And yet, LinkedIn is by far the absolute worst when it comes to misinformation and low information literacy.

For all the try-hard thought leadership being self-promoted here, the rhetorical and critical thinking skills on display are at an absolute low point. At least in debating the moon landing conspiracy theorists would read your sources and try to come up with a half-baked counterpoint. Here, you can literally spoon feed someone a whole buffet of counterfactuals and they breeze right by as if they don’t exist.

LinkedIn has changed a bit over the years - some users observe that it now feels like Facebook used to.

Some things haven’t changed. As you can see in Em’s post, LinkedIn is still for solipsistic peacocking - showing yourself off to an audience you barely acknowledge exists. It’s hard to change your DNA.

Generative AI does not make art because it is not mortal

2024-12-17T00:00:00+00:00

There’s been some debate recently about AI art, whether it will ever count as real art or be worthy of interest. A lot of the ‘no’ side of the argument has focused on the choices and creativity that creators bring to their work. I think Erik Hoel’s piece is informative. Humans are, on the whole, not terribly original. And AIs can be said in some sense to be making creative choices just like humans do. Hoel concludes by identifying a slim¹ margin of originality and control of distribution channels.

I think this is largely true but the analysis of choice is incomplete. The choices that human artistic creators make - that you make - matter because we are finite beings. It is the human condition. You will decay, and one day you will die. You will never live this day again, and that is why it matters what you do today. That is why your choices matter. That is why we explore lost love that can never be restored, battles that must be fought and won, and other such things, in art.

And this is why it’s reasonable to dislike a text when you know it was written by AI (if it wasn’t already obvious from how bland it is). The AI isn’t mortal and doesn’t even know it. It doesn’t have skin in the game², and that’s why it’s boring.

Perhaps one day an AI will exist that understands (its own?) mortality, and perhaps it will choose to explore this through, let’s say, literature. That would be a fascinating book to read. It would certainly be art.

and vanishing ↩
life ↩

AI reintermediation - the answer to consumer shadow work

2024-11-19T00:00:00+00:00

Being a consumer in 2024 is work. I think AI will help via reintermediation.

It’s been like this for a long time - many digital businesses work by disintermediation. A travel booking site allows you to search many, many providers of flights, hotels, etc, which has some benefits to you the consumer (price, variety), but like I said, it’s work. This kind of work is sometimes called ‘shadow work’, and there’s a lot of it in our lives.

I think there are opportunities all over the place to make this shadow go away, using AI. In general terms what this would look like is, rather than (e.g.) searching travel sites yourselves, you tell the AI what kind of a trip you want to take and then it does the searching, co-ordinating, and booking for you.

Ed Prentice, founder of AmigoAI, asks how this might affect advertisers and disintermediators. In theory, making it easier for buyers to access goods/services they want by filtering information for them should make markets more efficient. But we’ve already seen AI optimisation services show up here and there, so it’s a bit unpredictable. And no-one really expected the ‘few titans, many dwarves’ feature of disintermediated markets, so - as always - the future is unpredictable.

Disintermediators and aggregators should be well-placed to reintermediate with AI provided they can accomplish the technical lift, which not all of them will, creating opportunities for new entrants. In this vein, Perplexity has (at the time of writing) been successfully competing with Google Search.