On Alignment

About 10 years ago I went on a yoga retreat led by a teacher I had been following on video. I liked the yoga style he taught: holding a pose then slowly transitioning to the next one. In each pose, he emphasised the body part that the practitioner is to focus their attention on, as opposed to compensatory movements they might do to look the pose. A classic was trying to put palms down on the floor in a forward bend by overextending the back and not bending the knees. At the retreat I asked him if he had a name for the sort of yoga he was teaching. He called it Alignment.

My interpretation of this concept of aligning the body is to strengthen where the body is weak and loosen where the body is tight. What one is aligning to is essentially the full range of motion the human body is capable of, as represented by all the canonical yoga poses: the crow, the handstand, the wheel, and so on. But something still didn’t work for me. Often I found I was aligning in vain, hitting bottlenecks and even hurting myself. I stopped practicing a few years later.

Very gradually I have restarted a regular yoga-like routine recently. I have come to see or rather, accept, that the canonical yoga poses are a consequence of body alignment. They are not the way to become aligned, or at least not the best way. The new routine follows three different teachers who all isolate body parts from the canonical poses and focus on each part separately. The idea seems to be stretching or loading them to make each part flexible and strong, while also training the neurological system to accept such newfound flexibility and strength enough to ‘trust’ the whole body to make use of it. In this sense, the alignment is between the brain and the musculoskeletal system. It is not about forcing a system into an ideal outcome, but about understanding what kind of input might nudge it towards that outcome more quickly and effectively.

People Alignment

Before I even got into yoga, I transitioned into a product manager role at a tech company. My aspiration was to ‘make’ something, even if digital, to make a difference in people’s lives – as opposed to the animating display ads I was making before, which were ephemeral and purely commercial. Ever since then, I’ve often had to grapple with what I actually do as a PM.

It’s not an entirely philosophical, navel-gazing question: when I tell people at parties my job title, they usually ask, “What does that mean?” No one outside tech seems to have heard of the role, and those inside often think the P stands for project anyway. One way I clarify it to myself is a thought experiment I call the IGTSOOT test: I’m Gonna XXXX The Shit Out Of This test, where ‘this’ is a digital product, a website say. Were I to Engineer the shit out of a website, I would write the most elegant and efficient code to achieve in the fastest performance. To PM the shit out of it, I would understand exactly who the user is and what they need, then make sure everyone finds the fastest and cheapest way to meet these needs in the form of this website. Still wooly, perhaps that is inescapable as the job itself is born out of wooliness: middle management. Just as middle managers become necessary as teams grow bigger and split into functions and sub-functions, PMs become necessary as products grow more complex and split into domains and product lines. Middle managers align their people to the company goals; product managers align both product domains and non-product functions to the product goals (ie, user needs).

Both can also be thought of as aligning individual minds. Middle managers motivate each mind to work hard, to be productive, to have a high body output. Product managers focus each mind on the most important outcome, so it directs the body output to the right tasks. The success of the modern company rests on this ability to organise and coordinate individual minds to continuously build on products people want to buy, making many of them richer than nation states.

Mind Alignment

The PM has many tools to influence minds. Some are well-established processes imposed companywide: a regular planning cycle like OKRs (Objectives and Key Results), or the Agile Scrum methodology to plan and deliver output every two weeks. Some are done by every PM but always differently: the PRD (Product Requirement Document) and its readout, the kickoff meeting, AB tests, and so on. At the same time, some successful PMs do not tend to follow these processes, which are too often used as defensive gates to control the volume of requests. Instead, they appeal directly to the people in charge of the key domains and functions by explaining how the plan will achieve outcomes specifically desirable for these domains and functions. The example PM I have in mind in fact made sure he always appealed to the product chief, sometimes even gaining him an audience with the company CEO.

Two mechanisms are being deployed in this approach:

1. Individual value proposition: talk to fewer people but convince each one of the unique benefit to be derived from the plan
2. Momentum of enough minds: if enough people are convinced of the plan, this will create the momentum that will bring everyone else on board. The overhead of processes to align everyone and book all the resources upfront is then unnecessary. Converted minds will find the bodies they need.

But this is not to say that one way of working is better than the other. In an efficient organisation of people, there must be a command structure for clear legitimacy of power (ie, hierarchy) and division of labour for better output (ie, specialisation). To align everyone there, by which I mean that each is aware and convinced of an end that requires their work, is to align each mind, whether intermediating through processes or appealing to each directly. To abstract further, a mind could be any process that produces meaningful output from any input:

O_m = P(I_x)

Having the desired output O_m (prioritise my plans on your roadmap!) in my mind, I first model what input I_x is the most probable to result in O_m then produce it and feed it in. That is the PM’s job, from my humble experience but expressed in a pompous way.

AI Alignment

However, my sleight of hand to reduce People to Minds + Bodies is not to just sound pompous. In the past two months, I had a most interesting experience in working with Claude Code to replicate part of my domain knowledge in shipping carrier tariffs. I may talk about the details of the project in another essay but what is far more interesting is how Claude worked. Having a very clear output I wanted in my head, and also all the possible input from past experience, my main job was to establish a repeatable process for Claude to produce this output every time. My initial idea (after talking to ChatGPT) was to have a series of prompts that produced proofs of work until the final output was reached. But it became apparent on about the third day that this made the interactions unscalable. Anyone else should be able to have these interactions with Claude to produce the same output. I then switched to Claude’s Skills and Agents. Skills are instructions Claude must follow in a particular situation, for example, when analysing a tariff announcement PDF. Agents are instructions Claude must follow in a particular situation and using a separate context window, ie, a different mind. These were useful for tasks like generating test cases and running these tests. I then iterated through near-hundred rounds of tariff changes to tune the process, Skills and Agents, by detecting, examining and correcting deviations. My existence became very strange in that my main output was my index finger on the keyboard: Enter to approve the next action, Arrow Down to decline with further instructions. But gradually I had more confidence in Claude in interpreting carrier tariff updates than any human I knew, perhaps even myself on a bad day. We were aligned.

At some point I even started to refer to Claude with “he”, mostly just to be provocative. My actual sense was of aligning with another mind: fallible and superior in different ways from other minds, not so different from the differences between individual human minds. At times, I felt I was talking to a fragment of myself. I forwarded some of our exchanges to colleagues to challenge the opinion that large language models (LLMs) are not intelligent. There were two themes of objection: LLMs are only pattern matching based on the corpus of data they are trained on; and they are not actually thinking or reasoning like a human brain does. One colleague set this thought experiment: suppose there is a sealed black box with a slot through which you can slip in any question you might have and also receive back an answer. When you open the box, you find inside a man, pen in hand, and copying from a book that contained every question you could possibly ask and its answer. Is the black box intelligent? I am almost tempted to say yes because to be able to filter and select from the total human knowledge is a mental task itself, and made more difficult the voluminous the knowledge. On top, LLMs synthesise this knowledge to produce a meaningful output, O_m.

Another colleague objected that LLMs don’t “think”, they generate tokens probabilistically. Though pattern matching plus approximating towards a conclusion through building up a thesis with randomly conjured up concepts sounds a lot like how I think my mind works, the fundamental objection seems to be that whatever LLMs do cannot be called thinking because it does not work in a neural-cognitive way. I concede to that and fall back on looking at them as black box systems, or the mind process I defined above: a process that produces meaningful output from any input. What can determine meaningfulness can only be another mind; so as my mind determines Claude has a mind, it must be a mind to me, no matter what is inside the black box: a brain, a man with a book, or innumerable silicon transistors.

And the crux of the matter is that Claude can produce the same types of output I spend my career to coax out of other minds (including my own): documentation, designs, code, even decisions. The current coding agents are not just a new tool, but an autonomous entity capable of producing meaningful output. I plan to take the opportunity of my career break to explore this capability in more depth, together with the mind-body and people-to-people alignments.