Westra on AI

Contact

Wed, 24 Jun 2026 09:00:00 +0200

The easiest way to contact me is through my LinkedIn profile.

Ideas for a good future

Wed, 24 Jun 2026 09:00:00 +0200

In an earlier piece I deliberately kept my conclusions cautious. We do not know how far the automation of knowledge work by AI will go, or how quickly. Let me set that caution aside for a moment: suppose it succeeds. Suppose that within the foreseeable future, machines take over a large share of paid intellectual work. What then?

The first reaction is usually fear and unease — about money, but above all about the question of what people are still for. We have tied work, income and identity together so tightly that we can barely imagine a dignified life without paid employment. That is precisely where we need to start.

Work is a bundle

A paid job delivers more than a wage. It gives your day a rhythm, a place among others, and the sense that you are contributing to something. Pay, structure, connection and status are currently bound together in one package, attached to the labour market.

Automation cuts primarily the first thread: the market no longer needs your labour to make money. But for the person, little changes: they still need money to live on, a daily rhythm, a sense of belonging, and the knowledge that they matter.

No shortage of valuable work

We fear a world without work, while valuable work lies everywhere undone. Elderly people who deserve more attention than there are hands to give it. Children who would benefit from smaller classes. Neighbourhoods that need maintenance, nature that needs restoring, knowledge and culture that need to be passed on. The list is long and the needs are real.

There is no shortage of valuable work. There is a shortage of paid work — a fundamental difference.

The obvious objection is: if that work is so valuable, why doesn’t the market pay for it? The answer is that its value is hard to monetise. The value is diffuse, it materialises only in the long run, or it accrues to people who cannot pay: a child, a sick person, an ecosystem, a future generation. It is public value, and we do not organise that through the market but through shared choices. Education, healthcare and culture exist because we have collectively decided they are worth it, even if they generate no profit. It is work that resists automation, because it revolves around presence, human interaction and relationships. It is work that people do for people.

Care and interdependence as an organising principle

We can therefore think about work differently. As long as we equate “work” with “paid job”, a future with fewer jobs looks empty. The moment we look at “work” as “contribution to society in a broader sense”, a world of possibilities opens up.

A group of British researchers has developed this idea under the heading of care as an organising principle, at every scale: from the household to the community, the economy and the natural world. The core insight is that we are interdependent in thousands of ways, whether we find that comfortable or not. Nobody manages alone — not as a child, not as a sick person, not as an elderly person, really not at any point in our lives. A society that acknowledges this puts care at the centre of everything it does.

What is the economy actually for?

The economist Kate Raworth has a model for this, which she calls the doughnut. In her model the goal is not to make the economy as large as possible, but to meet everyone’s needs within the limits of the planet. A social floor that nobody falls through, and an ecological ceiling that we do not break through. Between those two lies the space in which a society can flourish.

What is appealing about this image is that it takes growth off its pedestal without throwing prosperity overboard. Growth becomes a means that is sometimes necessary and sometimes not, rather than an end in itself. The measure shifts from how much we produce to whether people and the living environment are genuinely better off. From “how do we keep production going” to “what is that production actually for?” Amsterdam was the first city to embrace this model as a compass for policy, in 2020.

A basic income or a job guarantee

So much for the direction. But a direction is nothing without options for implementation. If we decouple income from paid labour, two practical questions arise. How do people get an income? And how is all that valuable but unpaid work financed and taken seriously? Here the answers diverge.

One much-heard answer is unconditional basic income: the government gives everyone a fixed amount, regardless of work. It is simple, it gives people freedom, and it keeps the state out of the question of what counts as valuable work.

Another answer is a job guarantee, most fully developed by the economist Pavlina Tcherneva. The government offers everyone who wants to work a job at a decent fixed wage, precisely in the work the market neglects: care, community, nature restoration. Anyone who cannot find a place in the private sector can come here. Such a job guarantee also works as an automatic stabiliser. In bad times, when businesses shed workers, the programme grows. In good times, when the private sector picks up, it shrinks again naturally.

I lean towards the job guarantee, for two reasons. The first is that a basic income primarily turns people into consumers. The money keeps the economy moving, but it does not cause the work we actually want done to get done. Care, attention, restoration — someone still has to do that. A job guarantee arranges income and work in one move, and gives people what a benefit payment does not: a place in society and a chance to contribute to it. The second reason is political durability. A bare money transfer is vulnerable. It is easy to cut, easy to frame negatively as free money, or to quietly erode. Work from which people and neighbourhoods visibly benefit is much harder to abolish.

And the costs? They form a less hard constraint than is often assumed. The school of thought behind the job guarantee — Modern Monetary Theory (MMT) — points out that a state that issues its own currency cannot simply run out of money the way a household can. The only real limit is what is available in people, materials and environment. It shifts the question from “can we afford this?” to “do we have the people and resources to do this?”

The job guarantee also has its weak spots: work that disappears the moment the economy picks up fits poorly with care and nature restoration, which require continuity above all.

What is of value?

And who actually decides what counts as valuable work? With a job guarantee, that answer lies partly with the state — and that is a real risk. Whoever draws up the list determines a great deal. But perhaps that is not a flaw to be engineered away. Perhaps it is the heart of the matter. In a world where machines can do most paid work, the question of what human effort is worth is a political and moral question — and it belongs to all of us, not to an algorithm, a market or a ministry alone.

Alongside strengthening the safety net, we need to work out together which work we consider worth preserving, defending and growing. We are barely having that conversation now. We would do well to start it before technology makes the choice for us.

Links:

Kate Raworth, “Doughnut Economics: Seven Ways to Think Like a 21st-Century Economist” (2017)
Pavlina R. Tcherneva, “The Case for a Job Guarantee” (2020)
The Care Collective, “The Care Manifesto: The Politics of Interdependence” (2020)

Wed, 24 Jun 2026 09:00:00 +0200

Follow this blog using an RSS reader. Copy the feed URL below, or paste this page’s URL into any feed reader that supports automatic feed discovery.

RSS feed →

What is RSS?

RSS is an open standard for following blogs without an account or algorithm. Subscribe using a feed reader like NetNewsWire (Mac/iOS, free), Reeder (Mac/iOS), NewsBlur (web, Android, iOS) or Feedly (web). New posts appear automatically in your reader — like email, but without the inbox.

The plan is to automate almost all knowledge work. Where is the preparation?

Mon, 22 Jun 2026 09:00:00 +0200

Imagine a seventeen-year-old choosing a course of study right now. She is weighing up law, something in administration, or a degree in design. She does what we have been advising her to do for years: pick a direction and invest seriously in an education, in money and in time. The adults around her quietly assume that by the time she graduates, an entry-level job will be waiting for her, and a career beyond it.

But that assumption is beginning to look increasingly shaky. The companies building the world’s most powerful AI systems have written down what their goal is. And that goal is not “AI as a handy assistant” — though that is often how they present it in media appearances. The goal is to take over the work itself.

OpenAI describes its mission as building Artificial General Intelligence (AGI): “highly autonomous systems that outperform humans at most economically valuable work.” Autonomous systems that surpass humans at most economically valuable work. That is the sentence with which one of the most influential companies of our time defines itself. It is worth letting this sink in. Outperforming human labour is the explicit mandate these companies have given themselves, and hundreds of billions of dollars are being raised to make it happen.

The sheer scale of investment in AI is itself an argument. The sums now flowing into data centres, chips and energy are so large that they are hard to justify with the story “we are building a handy assistant.” They only make sense if you assume that the real goal is not to speed up a task here and there, but to take over large portions of knowledge work from people. The business model of these companies depends on succeeding at exactly that.

GDPval — canary in the coal mine

A counterargument you often hear is: tech companies always exaggerate. That is true enough. But how can we assess what is actually happening?

OpenAI has developed a benchmark called GDPval that tests AI performance on real professional work. Think of drafting legal advice, building a financial model, writing a nursing care plan, producing a quote. The test consists of real assignments from 44 professions, completed by both an AI model and a human, then evaluated by experienced professionals who do not know whether they are looking at the human’s work or the model’s. If the AI’s output is judged to be better or equal to the human’s, that counts as a point for the AI.

When the study was conducted in 2024, the score of the best model rose from around 12 per cent to nearly 50 per cent in roughly a year — where a score of 50 per cent means that in half of the tasks the AI delivers work that is as good as or better than the human’s. By late 2025, a newer model was already at around seventy per cent equal or better. This concerns knowledge work, not physical work, and a draw counts in AI’s favour. But the direction is clear, and so is the speed.

In other words: for a growing share of well-defined knowledge tasks, an AI model now produces output that an experienced professional can no longer distinguish from human work. That does not mean entire jobs are disappearing yet. But jobs are made up of tasks, and the entry-level roles in which young people learn a trade often consist of precisely those kinds of tasks.

Are we seeing a wave of mass layoffs due to AI? No, not yet. But that is also not what you would expect: labour markets are notoriously slow to respond to new technology, because introducing new systems takes years. The computer, for example, took decades to really make its way onto the shop floor. So there is no reason for panic — but there is every reason to pay close attention and take preparatory steps now.

The difficulty is that we are poor at imagining a world that is fundamentally different from the current one. We unconsciously project today’s labour market into the future, just with an AI veneer on top. The possibility that the structure of the labour market itself might change — that the relationship between work and income might look different — falls outside our frame. And what we cannot imagine, we cannot prepare for. We need to stretch our imagination.

Two futures, both achievable

The gloomy picture is easy to sketch. A labour market where the bottom rungs of the ladder have been removed, where young people find no foothold to learn a trade, where the gains from automation flow to a small group and the costs fall on everyone else. A society wealthier than ever, yet declaring whole groups of people redundant. That is the outcome you get if you let the current incentives run their course and do nothing more.

But there is another outcome, and it is equally achievable. If machines take over a large share of necessary work, something is freed up that has historically been scarce: time and human attention. Time for care, for education, for each other. For restoring nature, for work that currently goes undone because it is not profitable. A society that decouples prosperity less rigidly from paid employment, and gives people the space to contribute in ways the market poorly rewards.

Three measures

Because nobody knows how AI will develop or how quickly, it makes sense to start with measures that are worthwhile whatever the future looks like. Here are three.

The first: a solid safety net. We already have systems — like unemployment insurance — that work as automatic stabilisers the moment people lose their jobs, regardless of the cause. As the economist Martha Gimbel puts it: don’t get too creative with your solutions for an AI-disrupted labour market. Strengthen the systems that already exist, because they are flexible, they work automatically, and they have proved their worth in previous technological transitions. Whether the shock turns out to be large or small, a stronger safety net is never wasted investment.

The second: measure what matters. We cannot respond to displacement we do not see coming, so we need a system that makes shifts in the labour market visible before they become a crisis. Invest in reliable, detailed, and timely information.

The third: start thinking now about preparing an alternative fiscal and social architecture, so that we do not have to improvise in a panic if change arrives suddenly. There is plenty of input for this — more on that in a future post.

None of these steps requires you to believe that all work will vanish next year. But they are useful first steps if you take the possibility of an AI-seriously-disrupted labour market seriously enough to prepare for it.

Links:

OpenAI Charter (AGI mission definition)
OpenAI, “Measuring the performance of our models on real-world tasks” (GDPval, Sept. 2025)
Martha Gimbel, “Don’t get fancy with your labor market fixes for AI” (The Argument, Dec. 2025)
Axios, “Economists, investors pitch Washington on AI-driven job loss safety net” (March 2026)

About

Sun, 21 Jun 2026 17:00:16 +0200

I’m Mark Tiele Westra, and in this blog I think out loud about AI: the technology, the implications, and what it means for how we work and live.

My background: a PhD in Theoretical Physics and a Master’s in Philosophy of Science, Technology and Society. Over 15 years of experience in software development, including data analysis tools used for water and sanitation projects in more than 20 countries, and machine learning projects with the World Bank. I currently hold a dual appointment as Practor Applied AI at Firda and Associate Professor of Applied Sciences of Applied AI at NHL Stenden.

I find the technology behind AI equally endlessly fascinating and genuinely unsettling. This blog is a reflection of my search to better understand this remarkable phenomenon.

Note: the English articles were machine-translated from the Dutch originals.

AI and the five stages of grief

Sun, 07 Jun 2026 09:00:00 +0200

With AI we have realised something we regarded as uniquely human — our intelligence — inside a machine. And now we stand looking at it. What we see surprises and unsettles us: perhaps intelligence is not quite what we thought it was, and perhaps it is not ours alone.

This kind of unsettling has happened before. Copernicus showed that humanity is not the centre of the universe; Darwin showed that humans are animals among other animals. Sigmund Freud described such shifts as “wounds” to human self-love — painful decentrings in which humanity must let go of deeply held beliefs. And that pain calls for grief.

In 1969, the Swiss psychiatrist Elisabeth Kübler-Ross wrote about grief. She identified the now-familiar five stages: denial, anger, bargaining, depression and acceptance. Perhaps we can apply this framework to how people respond to AI? In an essay from 2024, technology philosopher Benjamin Bratton writes about precisely this — not as stages that follow one another, but as a typology. Something like this:

Denial — “it’s just autocomplete, mere statistics, not truly intelligent and certainly not truly creative.”

Anger — “AI is an existential threat; it endangers work and human dignity and must be stopped, by force if necessary.”

Bargaining — “with the right rules, ethics committees and alignment techniques we can keep AI under control.”

Depression — “it may already be too late to save humanity from AI; this is going to end badly for us.”

Acceptance — “this was always coming; it is the next step in a long evolutionary process.”

All of them are understandable, human responses. We have seen all of them in the news recently. Each contains a kernel of truth, and each is also a simplification. Can we do something with this? Can this framework help us look toward the future — and move beyond the stages of grief?

In his essay, Bratton invites us to search for other ways of thinking about the nature of artificial intelligence: ways that are neither optimistic nor pessimistic, neither utopian nor dystopian. So that we can look with open eyes into the mirror that AI holds up to us, and learn something about ourselves and our place in the world.

Digital Makers

Tue, 05 May 2026 09:00:00 +0200

In the article “More AI, less enrolment is a dangerous combination”, Anita Bosman and Klaas Brongers describe a worrying trend: young people are choosing not to study ICT because of AI. Meanwhile software is rapidly penetrating every corner of society, and the shortage of ICT professionals is structural.

Their suggestion to make ICT education more attractive is well-taken. But it is easy to miss the mark. Adding a healthy dose of AI to the curriculum and explaining that ICT really does matter for society is not going to be enough. Because the underlying fear is left unaddressed: that AI makes programmers redundant. As long as that perception persists among students, parents and careers advisers, little will change.

What we need is a new narrative. Not “learn to code despite AI,” but “build the future with AI as your tool.”

At Firda we are working on that narrative. We call it the Digital Maker: a creative problem-solver who uses a broad set of tools — code, neural networks, language models, and their own creativity — to build solutions alongside AI across all kinds of sectors. From a robot that picks apples to a healthcare app that helps patients. Someone with deep ICT knowledge, combined with experience and interest in other fields: healthcare, agritech, logistics, robotics.

AI does not replace the programmer. AI makes the programmer more powerful and more versatile. But you have to tell that story, and it has to match what young people actually encounter in their training.

A plea for messy imperfection

Mon, 06 Apr 2026 09:00:00 +0200

Pieter processed address changes for a large insurance company. Predictable, straightforward work. Useful and important — for him and for his employer. Until customers could submit their own address changes online. After that, there was no suitable work left for Pieter.

Now Pieter has “a certain distance from the labour market.” But has Pieter drifted away from the labour market, or has the labour market drifted away from Pieter?

We treat efficiency as though it were a law of nature. As though it is self-evident that work disappears the moment a machine can do it more cheaply. But that is a choice. And every choice has winners and losers.

Pieter’s work did not become unnecessary. Address changes still need to be processed — only now a system does it. His work was optimised away. The question is not only can technology take over this work, but also: should we want it to? And if we do, what do we offer Pieter?

With the arrival of AI, this pattern is accelerating. Not only administrative work is disappearing, but also what we until recently called “knowledge work.” What remains is largely supervision: maintaining oversight, correcting course, making judgements. Senior-level work.

The labour market is not shrinking as a result, but it is narrowing. The broad base of executional work is contracting. What remains is a top tier of knowledge workers and a growing bottom layer of insecure, poorly paid work. The unemployment figure does not tell that story. It is also about the quality of work. About dignity.

There was nothing wrong with Pieter’s work. It was predictable and manageable. He was good at it, and it gave him a place in an organisation, in society. That is a basic human need.

We have lost the ability to value that kind of work. Work that is not “smart,” not “innovative,” not “disruptive.” In a culture obsessed with optimisation, there is ever less room for it. But paid work is more than an economic instrument. It is how people participate in society. When that pillar falls away, more falls away than income alone.

Hence a plea for messy imperfection. For a labour market that deliberately leaves room for people. That does not optimise everything. That accepts that human labour is sometimes less efficient, and is fine with that.

Messy imperfection means: work that keeps a place for Pieter, even if a system could do it faster. A society that values care work. Care for people, for the community, for nature. Not always the cheapest choice, not always the fastest. But perhaps the most human one.

Maybe it is time for the labour market to come back to the people. Messy and imperfect. But with room for everyone.

Anthropic vs. the US government

Sat, 28 Feb 2026 09:00:00 +0200

Anyone interested in AI ethics should pay close attention to what is unfolding in America right now.

The Department of Defence (given the pointed new title “Department of War” by Trump) is demanding that Anthropic (maker of Claude) lift all usage restrictions on its AI models. Specifically at issue are two safeguards Anthropic had written into its contract: a prohibition on mass surveillance of American citizens, and a prohibition on fully autonomous lethal weapons that select and kill a target without a human in the loop.

Anthropic has refused, and that has brought the wrath of the American government down upon it. Trump has ordered all government agencies to stop using Anthropic’s AI models. Defence Secretary Hegseth has labelled Anthropic a “supply chain risk” — a designation normally reserved for foreign adversaries and never previously applied to an American company. At the same time, the administration is threatening to invoke the Defense Production Act, a law that only applies when a product is essential to national security. So Anthropic is simultaneously a security risk and indispensable. A curious contradiction that reveals this is above all a power struggle.

Meanwhile OpenAI (maker of ChatGPT) has stepped into the breach and struck a deal with the Pentagon. They claim to uphold the same red lines, but leave the assessment of what that means to the Department of Defence. Convenient.

Anthropic is no saint: it works with surveillance firm Palantir and was the first AI company to operate on the government’s defence computer networks. But this confrontation does show that significant differences exist between AI providers when it comes to drawing ethical lines.

By now almost all of us use AI in some form. This episode shows how high the stakes are: the ethical limits on AI are under enormous pressure, and a favourable outcome is far from certain. This is a strong argument for digital sovereignty — our own European and Dutch AI models, on our own terms, grounded in our own ethics.

Links:

Software 1.0, 2.0, 3.0

Sat, 07 Feb 2026 09:00:00 +0200

The profession of software developer is not going to disappear — it is going to become richer than ever.

As a programmer you conceive something — a website, a game, an algorithm — and the computer executes it. That is the essence of the craft: translating a human intention into a working computer system. That does not change. But the toolbox has exploded over the past few years.

Andrej Karpathy (former AI director at Tesla, co-founder of OpenAI, and the person who coined the term “vibe coding”) has a clear framework for this. He distinguishes three kinds of software:

1.0 — Classical code. Instructions that tell the computer exactly what to do. Think of your word processor, your banking app, or the software that guides a rocket to the moon. Until recently written by humans, but increasingly written by AI or in collaboration with AI.

2.0 — Neural networks. Systems that are not programmed with rules but trained on vast quantities of examples. A self-driving car does not recognise a pedestrian because someone wrote rules describing what a pedestrian looks like — it recognises one because the system has seen millions of examples.

3.0 — Language models such as Claude, ChatGPT and Gemini. These are complex systems built on neural networks. What makes them remarkable is that you can direct them in plain language: you “program” them with a prompt.

To handle the enormous variety of the real world you sometimes need software 1.0, sometimes 2.0, sometimes 3.0 — or a combination, as in AI agents.

From now on we will increasingly work alongside AI agents that write the bulk of the code. The developer retains control: specifying the desired outcome, verifying the code — without compromising on quality. Karpathy calls this “Agentic Engineering”: agentic because as a developer you mostly no longer write the code yourself but deploy agents to do it, engineering because it remains a genuine discipline with its own depth and expertise.

But what does this mean for the profession? Does every developer need to master all three forms of software, or will new specialisations emerge? How does the balance between designing, building and testing shift? At Firda and NHL Stenden we are fully engaged with these questions. Because if the craft changes, education must move with it.

Links:

Karpathy presentation

AI in the Coalition Agreement

Sun, 01 Feb 2026 09:00:00 +0200

The new Dutch Coalition Agreement pays considerable attention to AI. Rightly so, I think, because AI is going to have a major impact on our lives. This, to my mind, is the key sentence: “We are building a digitally resilient society through an ecosystem approach, working together with education, business and government to take a step forward so that everyone can fully benefit from the opportunities of artificial intelligence.”

Good idea! But how do you do that? How do you ensure that “everyone can benefit from the opportunities”? How do you make sure that deploying AI strengthens human autonomy? That AI contributes to a good life? How do we know who benefits, and who bears the risks?

One important ingredient here is AI literacy — the ability to understand, use and critically evaluate AI. Under the EU AI Act, organisations have been obliged to work on this since last year. How do we give that shape in education?

One of the best resources I have come across in this area is the UNESCO AI Competency Framework. There is one for students and one for teachers. It is well-grounded, evidence-based material. For students, the competency dimensions are:

A human-centred attitude
Ethics of AI
AI techniques and applications
Designing AI systems

Each dimension comes in three levels: understand, apply, create. This way, each aspect can recur several times throughout a curriculum, allowing knowledge to stick and deepen.

If you are working on this topic, this framework is well worth your time. What appeals to me most is that ethics and a human-centred attitude form the foundation for everything else. And that awareness is something we are going to need urgently to hold our own in the AI storm ahead.

Relevant links:

My software estimation intuition no longer works

Sat, 17 Jan 2026 09:00:00 +0200

Throughout my career I have regularly encountered one of the more elusive elements of software development: estimating how long it takes to build a given piece of functionality. It is a kind of dark art in which intuition, experience and unfounded optimism combine to produce a wild guess at “how long this will take to build.” Among programmers it is a rich source of humour, usually with a kernel of truth. There is, for instance, the “pi rule”: take your estimate and multiply it by pi (roughly 3).

Why do I bring this up? Last week I had an experience that showed just how dramatically the craft of software development is changing. I wanted to build a demo application for an Agent Constitution Framework I am developing at NHL Stenden. Think: HTML/CSS/JS frontend with 10 tabs, Python backend, Ollama language model, SQLite database. Had I been asked to estimate this work in my previous life as a developer, I would have said two to three weeks.

On Monday at 8:00 I started writing the specification. By 8:20 it was done. I then asked Claude Opus 4.5 to write the code. By 8:30 the code was ready. And it worked first time. Clean user interface, well-structured code, genuinely nothing to criticise. That is roughly 200 times faster than I would have estimated a few years ago. I stared at my screen with considerable astonishment.

Of course, programming is not only about writing code. Reaching consensus on what needs to be built is at least as important. And naturally this was a small, greenfield application that only needed to run locally on my laptop — all the complexity of DevOps, hosting and security was absent. Even so. The craft is changing at breakneck speed.

Exponential growth – deceptively slow at first, then alarmingly fast

Fri, 07 Nov 2025 09:00:00 +0200

Exponential growth: deceptively slow at first, then alarmingly fast. Place one grain of rice on the first square of a chessboard, two on the second, four on the third, eight on the fourth, and so on. How much rice do you need in total for all 64 squares? Take a moment to imagine it. What does your gut say? Now the answer: you need 370 billion tonnes of rice — roughly 670 times current global production.

Halfway across the board, the total is “only” 43 tonnes. Futurist Ray Kurzweil uses the phrase “the second half of the chessboard” to describe the point at which the speed of exponential growth becomes truly felt in technological development. That is where we are now with generative AI. It means our intuition fails us. That developments move so fast that an opinion on AI has a shelf life of at most six months. That our collective imagination is entirely inadequate to see what is coming, leaving us purely reactive in the moment.

For me, 2025 was the year the AI acceleration became tangible. Language models like Claude, Gemini and ChatGPT grew steadily more powerful, and AI agents emerged. Software developers saw their daily work change completely; freelancers in the creative sector — translators, photographers, designers — began losing commissions.

So what are we to do? How do we ensure this revolution works for people, society and nature? That its costs and benefits are distributed fairly? Too much is at stake to let this wash over us passively. It demands reflection — in education, in businesses, across society at large.