Unpredictable Patterns #135: On work, structure and artificial intelligence

The centaur knowledge worker, tempo and structure in work and why work - or action - is likely to persist through waves of technology

Dear reader,

Shorter note today on work - and how it is structured. This is a key interest of mine, and I know I share it with many. We think about work, personal productivity and thinking incessantly, and so it is surprising to me that I have not seen more writing on how to recreate work with AI, concretely. There are many articles on the usefulness of the tools offered by AI, but the roadmap to any more collaborative working style still is missing - and is badly needed. This is not it, but it is a start.

Crafting work

Since I recently left my corporate job, and instead have decided to pursue my own projects, I have been struck by how interesting and hard it is to do knowledge work that really feels satisfying.¹

Good work, in my view, is work that makes some kind of difference in the world, and at the same time gives the person doing it a sense of meaning. That meaning can be based on the classical motivations outlined by Daniel Pink: purpose, mastery or autonomy, but the key thing is that there is some kind of meaning to what we do.²

Pink’s triad, often quoted in discussions on work, is interestingly reframing some of the earlier findings from what is called Self-Determination Theory (SDT). In SDT, however, autonomy is replaced by relatedness.³ This is no small change, and as I have been thinking more about work as something I need to shape and craft, I have been wondering about whether both autonomy and relatedness may actually need to be included in the motivation for work.

There are also other theories of good work that focus not so much on what the work creates for us as workers, but instead focuses on the sense of making progress. I find these models and theories intriguing, and suspect that there is a lot of truth to this: making progress on a goal may be a large part of what makes something feel like good work, at least.⁴

Taken together, these theories give us the basics for understanding if we are likely to deem a day well spent in work or not - but what they do not really suggest is how to structure the time we work. This is especially hard if you are working alone - and I don’t mean just in physical solitude. Many of us work alone, even if we go to an office: we sit beside other people, and we discuss with them over lunch, but when we work we sit in front of our computers engaged in writing, doing email or coding. That work is lonely work - even if it is done in the midst of an office.

Authentic team work is probably quite rare, and can be exhausting. If you have spent a full day in a workshop discussing something, or worked in a design setting you will know this - but you will also know that when it takes off, it is really satisfying and rewarding.

Now, lonely work can also be enormously satisfying and rewarding, but it is much easier to doubt yourself when you are working on your own. A day spent with writing or reading can feel unreal, especially when the work is digital and there are no traces of it anywhere.

The live debate about the importance of writing longhand and the use of notebooks is an interesting example of this: there is no end of articles suggesting that writing longhand is better than writing on a keyboard. It is hard to not get the impression that this research is born out of a secret need for this to be the case - we want to feel that we are choosing a better alternative for writing when we we write in a notebook with a traditional pen, when, in all honesty, it may be that the tangible results of our work and thinking is what we really crave, and that may be reason enough! The feeling of good work well done is, in itself, a deep source of satisfaction.

Working with AI

The advent of AI-models offers an intriguing possibility here, and might change the nature of lonely or solitary work. We say, admittedly with not a little handwaving, that AI should augment workers and not replace them - something that has led to the idea that the future will see the emergence of centaur work. Such work will be done in partnership with AI, and should in theory help us focus on what we are good at and use the model to help with the rest.

What we do not know yet is just how to structure such centaur work - what does it look like on a daily, or hourly, level? Experimenting with this is really interesting, and I think there are a few things here that are worth exploring. The following are my key takeaways from working and thinking with an AI-model for a while.

Don’t use the model for drafting. A draft sounds like something you can start with and then bring your superior intellectual skills to perfect - but the real cognitive work, the work you want to practice and deepen, is in the initial shaping of a subject and an idea. Drafts are terribly important from a creative standpoint - they force shape onto a problem, and if you consistently accept external shaping of a problem, you lose your ability to engage with raw reality and shape it with patterns that are uniquely you. Writing is not “content production”, not at all - and this is such a horrible term for it. Writing is deep sensemaking. You make sense of the world in drafts, iterations, sketches - and offloading that to a model will reduce you.

Use the model as several critics. This also reverses some of the common wisdom, where the idea is for the human to be the auditor or critic of the model and ensure that it is not off target. The better use is to have the model engage dialectically with what you do after you have reached a point where you are satisfied that you have a sense of direction in your work. Since self-criticism is a dangerous thing, you should prompt the model to discuss your work, not just criticise it - and suggest improvements and ideas. If you are up for it, putting your work on trial with a defender and an attacker is not a horrible idea.

Create commitment. Starting out by setting out what you want to accomplish in a given day and making sure that you get back to the model and report back on things is actually surprisingly effective. A former colleague noted that this is one of the greatest, hidden benefits if you work in a corporation: when you reach a certain level, your day is created for you by calendar invitations, meetings, and email. “All you need to do is to surf that wave well”, he noted. I found that a profound insight - and I think that is true.

What Coase misses in his analysis of the firm is that it is not just about cost, but a firm creates a tempo. As you have to create that tempo on your own, the use of pomodoro timers and commitment chats with a model turn out to be surprisingly helpful. The model can be tasked to ask you to report on your day as well, and suggest improvements — and if you implement a “captain’s log” habit of chatting to it at the end of the day you have an excellent opportunity to reflect on your day candidly.

Leaving a corporate job is leaving a rhythm machine, a temporal architecture that you now have to create on your own.

The what’s next? If you have a proper list of things you need to do and connect that with your model, it can suggest tasks that correspond to what you want to do. Even just connecting a Google doc to a model with a list of projects and tasks is a nice way to get some advice on what is next, and how to work on something new. You can review the list and suggest what you want to get done, discuss the possible next steps and blockers. If done well, and over time, this will also unearth work patterns and habits for you to shape.

Goals. One thing that has struck me is that if I describe my goals to the model, that is often enough for it to suggest something I can do today to make progress on the goal. This way of working - with a set of goals, and an assistant that suggests how to reach the goal according to your energy levels or time allotment, allows for not just task tracking, but task discovery. The model breaks down the things that make the goal more likely, and helps you see more than one path to the goal — and then tracks what you have done. I have found that the model is often better at breaking down the goal into multiple paths, and even when I disagree with the proposed tasks the discussions are interesting.

All of these are quite simple ideas, but they are interesting in that they bring two key values to work: structure and tempo - and this may well change lonely work in ways that are highly productive. As I continue to experiment I am sure I will learn more - but I am also interested in how others think about and work with AI.

There are some intriguing emotional and psychological aspects of this as well — a sort of tangent to the ongoing discussion about AI as a therapist, and what it means that we are engineering our own psychologies with these new technologies. Work is human action, and in some way what we are doing is creating patterns that shape action through persuasion and interaction — and this psychological coding is underexplored and potentially really important - as we code the technology, we now use the technology to code ourselves.

AI is likely, here, to become the key foucauldian technology of the self⁵, not just a way to know ourselves, but a way to take care of and shape ourselves — from the gnothi seauton to the epimeleia heautou.

When you work with work you work on yourself.

Thanks for reading,

Nicklas

1. Oh, and let’s be clear - what follows is a discussion of a small part of the labour market. White-collar, western work is very different from the literal need to make a living that still pervades much of human life, so in a sense the following discussion will be explored from a position of privilege.

2. See Pink, D.H., 2011. Drive: The surprising truth about what motivates us. Penguin.

3. See Ryan, R.M. and Deci, E.L., 2000. Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American psychologist, 55(1), p.68.

4. See Hsee, C.K. and Abelson, R.P., 1991. Velocity relation: Satisfaction as a function of the first derivative of outcome over time. Journal of personality and social psychology, 60(3), p.341 and to some degree Kivetz, R., Urminsky, O. and Zheng, Y., 2006. The goal-gradient hypothesis resurrected: Purchase acceleration, illusionary goal progress, and customer retention. Journal of marketing research, 43(1), pp.39-58.

5. Foucault, Michel. "Technologies of the Self." In Technologies of the Self: A Seminar with Michel Foucault, edited by Luther H. Martin, Huck Gutman, and Patrick H. Hutton, 16-49. Amherst: University of Massachusetts Press, 1988.