Surgery With My Teeth

My guitar collection rocks. There’s a gold top Gibson Les Paul, a cherry red SG, a couple of Fender Strats, a Cole Clark acoustic made from Queensland silky oak, and a peculiar looking, no-name, deep blue Telecaster. I built that one from raw wood by hand, with a cheap saw and a lot of sandpaper. It was my COVID project, and it took two years.

Before I started it, I watched a ton of YouTube videos about woodwork and guitar making. I ordered the wood from two different timber suppliers. The neck was made from a Canadian maple tree that I chose for the guitar before it was cut down. I bought enough wood for two attempts, because I was certain that I would screw up the first one.

I started with the body of the guitar. The first cut with the saw into the high-grade Mississippi swamp ash felt like I was performing surgery with my teeth. I chopped out a large triangle of wood and, despite being nowhere near the measured outline that I had drawn, I was almost surprised that I’d managed to remove the first piece without irretrievably ruining the whole project. I had seen dozens of videos on how to do exactly the work that I was doing, but I had no way to bridge the gap between what I’d seen and the actual act of interacting with the wood.

“Being ready” is often taken to mean “understanding enough to begin”. But some understanding can only be gained from doing the thing. If we require the understanding before we start, then we have psychologically created an unreachable starting line.

Preparation is a wonderful thing, until we expect it to give us something that requires experience. You can read every book about cooking and never know how to “salt to taste”. The guitar building equivalent is knowing when to stop filing and start sanding. At some point, in most endeavours, an attempt must be made at doing something that carries uncertainty. And from the point of that attempt, the uncertainty you carry starts becoming less vague, and it becomes something you can work with.

The thing in front of you doesn’t have to be a finished product, either. A first attempt at a piece of the thing can still mean that some of the work gets done, even if nothing in it is actually worth keeping. It can show you the boundaries of your actual understanding or capabilities, or help you to crystallise your position where you had only a general feeling before. It gives you something to react to, and because of that, you can usually bring a clearer vision to your next attempt.

I first began learning to use AI because I could sense that it was integral to the viability of my professional skills. I do it now because making things with it is something I love to do.

My first attempt at an app was a document compare tool. It was meant to surface differences between documents in a manner that could enable you to make a decision and preserve it through subsequent rounds of review and iteration.

When I attempted the first app, I didn’t know how to code. I was using ChatGPT, before coding harnesses were mainstream, and I was asking the browser-based chat interface to produce code for me to copy and paste into a code editor. When something broke, I had to attempt to understand the code enough to explain the problem and generate another attempt. It was like drawing in the dark, and I found it extraordinarily difficult.

At some point after starting, I produced something that gave me an interface that let me upload two documents, compare them, and correctly identify similarities and differences. It was basically a worse version of the Word compare function, but it was a working thing on my screen. A representation of part of an idea that I could previously only describe.

That was the moment the world changed for me. I didn’t suddenly form the view that “anyone could do anything”. The shift was more subtle, and for me at least, more dramatic: anyone can do something. I had turned part of a software idea into a working thing without first learning the coding skills to build it alone.

My app failed. I was working towards a deadline, and I didn’t have enough for a coherent proposal when the deadline arrived. Throughout the attempt, there was never a moment where I felt like I was wasting effort.

I got better as I went. I was turning intentions into tasks, and I was having fun figuring out how to do that. The process of arriving at the actual tasks is still mine to own, and I came to recognise that organising the work and judging what to ask for is a genuine skill. Something I started pursuing as a precaution had become something I valued for its own sake.

Judging something as being worth pursuing is different from judging it as being worth having done. A disappointing outcome does not necessarily make the decision to try a mistake, and no lesson needs to be learnt to justify the attempt.

My guitar had a measured outline to cut into, and my first app had a deadline. It is common to start towards a goal with less, and the things we learn on the journey can change the destination. That’s what happened for me. The project ended, but the pursuit did not.

There are things that we know we want but don’t know how to pursue. There are times where our preference for something is exposed when we encounter something we don’t like. And there are things that we are driven to want because we tried them. In each case, the change in our understanding alters the way that we make our next judgement call.

The difficulty is that that same understanding is what we use to decide what deserves our effort in the first place. If we demand a complete justification in advance, we exclude the encounter that would have given us a reason to care. Anthony Bourdain said in a 2017 interview, “I’d rather fail gloriously and foolishly than turn in efficient and adequate work again and again.” He wanted to pursue things whose value was not already guaranteed by his ability to repeat them.

In my essay It gave me a playlist, I argued that as tools got better, the difference between novice and expert AI users shrank, but that knowing what is worth asking for remains a consequential differentiator. That claim places significant weight on arriving with a worthwhile request, as if the judgement had to exist before the work began. It doesn’t. The judgement I was insisting on is partly a product of the work, not a precondition for it. A capacity that cannot be delivered by an AI coworker can still be developed through working with one. Better tools do not supply imagination, but they can enable encounters that shape our judgement.

Sometimes I discover my own preferences only when a model gives me something I don’t like. Figuring out what I’m bothered by is part of the decision process around what to ask for next.

Knowing what to ask for can include knowing what to ask for in order to find out what it is that you actually want.

I’m currently working on a project where this has become very practical for me. I built a functioning workflow, more or less exactly as planned, and when I interacted with it, I realised that I wouldn’t want to use it. My imagined work included too many uncertain variables that made the working tool burdensome to use.

The underlying idea remains a good one. But as I sit with the dissatisfying workflow, I can only tell that I dislike using it, not what needs to change. My original vision no longer serves me as a useful compass.

So instead, I tried something else. I gave the entire codebase to a new model, explained what I didn’t like, and asked for an alternative proposal. It proposed changing the system underneath the interface rather than the interface itself. I think that change could give me a different basis for a redesign I can’t yet picture.

That proposal is being implemented. I haven’t seen the result; the work is substantial and has been running without my intervention. The bet that I’ve placed is that a different starting point will help me identify what I want to do next.

The purpose of that task is to give me the ability to play my part. I don’t expect the agents doing the work to deliver a perfect final product. All I want is for it to give me something that I’m better equipped to react to.

This approach of asking for something to react to can also become a way of postponing a decision. Output that looks good can be a reason to stop looking, and a stream of polished alternatives is an easy way of searching forever. The best test, and the only one I trust, is actual use. The first version of my current project disappointed me when I used it, and I must be open to the next version doing the same.

When I built my guitar, I didn’t know that the neck was any good until it was on the guitar with strings on it. Without the strings, I can’t tell whether the width fits my hand, or the frets feel right. If it had been a failure, the second block of Canadian maple would have been used.

Today I play the guitar for one entirely uncomplicated reason: I love to make music with my own hands. I don’t seek out improvement, and mistakes don’t frustrate me. Recently, I recalled a song I hadn’t played in at least a year, and I just played it. That is enough, and it is not what I was chasing when my parents bought me my first cheap acoustic. I could not have known then what the instrument would come to be for.

The version of me who copied code into an editor could not have known either. Every experience in this essay changed what I was able to recognise, and none of them required that recognition in advance.

AI has made those beginnings more frequent. It gives an uncertain intention somewhere to go, and more ways to find out what I think when the first attempt doesn’t settle it. A completed vision is not a condition I place on starting.

Some of the things I will care about are things I do not yet know how to ask for.