Hello friends,
For the last year, I’ve been talking to people (many in government and public policy) about using AI as a tool or building tools with AI — very often in the context of what happens to people when AI tools make them superfluous. How do we decide what makes a tool good and what kinds of tools we should aim to make?
Answering these questions is particularly important when there’s already a whole frothing mass of magical-seeming AI tools and more showing up every day. Which makes it tempting to uncritically think of all these tools as being generically good. So the stuff in this issue has been sort of floating about at the back of my mind for a while. It took a few days snowboarding with not very good rented equipment this week for it all to crystallise and fall out of solution.
Sliding down slopes in Savoie
I’m in the mountains benefiting once more from generous friends renting an accommodating apartment from which I can walk onto a major access lift into the Trois Vallées, which claims to be the world’s largest ski area connected only by lifts and slopes.
2009 was the last time I’d snowboarded before last January. As I drove up the switchbacking mountain roads last year, I realised with a frisson that I’d forgotten both how to get on a lift and which foot I habitually lead with on the board. Would I spend 3 days tumbling down icy slopes being chortled at by expert snowsports enthusiasts?
Luckily, whatever paltry snowboarding skill I had developed decades ago on the icy slopes of the American Northeast came back quickly — in fact, within minutes of coming off the first lift of the day. After only a few embarrassing faceplants involving no permanently debilitating injuries, I was happy as a clam coming down blue slopes. After that, this year was a doddle, so I became concerned instead with finer points of control.
Control
Here is a theory of snowboard control. The snowboard rests on the snow surface. You, the rider, are connected to the board by boots which are locked to the board by bindings. You move your body, and the boots and bindings transmit your motions to the board, changing its orientation to the snow. Snowboard control — adjusting how fast, and in what direction, you slide down the slope — is done by changing how the board and its edges are oriented to the snow surface.
Skill
Writing out the theory of snowboard control like that has almost no connection to the act of actually riding a snowboard. This knowledge is inherently tacit and somatic — understood by the body as a whole — instead of understood cognitively as a symbolically represented system of explicated knowledge. It is something we can know even if we cannot say or write it down.
(Michael Polanyi’s original illustration of tacit knowledge is the act of riding a bicycle. This can be unpacked into a symbolic representation: a bunch of math describing how the bicycle stays upright while moving forward. But that symbolic representation of bike-riding is useless for actually riding one. Nearly everyone who actually ride a bike does so without knowing the math behind it — their bodies know it somatically.)
Tacit knowledge in this case is knowing how to achieve different outcomes. I suppose another word for this is “skill.”
Intent
Riding a snowboard with fine control requires skill, the tacit knowledge of how to move the body in order to move the board correctly. It also demands a clear sense of what “correctly” means. This sense of correctness is what I mean by “intent.” Concretely, intent is something like “I intend to move over to the left avoiding the patch of icy slope and scrubbing off a little speed, then to turn to the right and pick up speed off the banked-up snow on the edge of the piste to get around the skier who has wiped out on that mogul.”
Intent is about consciously choosing which outcome to aim for.
Transmission
Skill in manipulating the system and intent about how the system should be manipulated are as nothing without a way to transmit skilled intent to the system. This thought came to me as I adjusted my boots and bindings yet again, high up in Val Thorens.
Because I snowboard an average of 3 days per decade, I rent all the kit. The boots I get never fit perfectly, nor do they ever perfectly lock into the bindings. This means that when I move while on the board, I move the boots and the bindings and the board — but I also move an inconsistently tiny bit inside the boots without moving the boots themselves.
Snowboarders call this heel-lift, because the primary sensation is of your heel lifting up off the insole of the boot when you move. A perfectly fitted boot would leave no room for movement. But with imperfectly fitted boots, heel-lift means the body’s intentional movements are not all transmitted to the board. There is slippage, you catch an edge, maybe you lose control.
The transmission quality of tools
On the slope, perfectly fitted boots and bindings would enable skill and intent to accurately and precisely choose where the board should be at any moment.
Let’s generalise, because this idea holds off the slope also. The generality is especially well-illustrated when our tools fall short in quality. Have you ever cooked a wok-fried noodle dish and achieved only a pallid, soggy result because your pan was the wrong material and shape and your burner insufficiently puissant? Tried to cut a dovetail joint and ended up with a gappy fit between the boards because your scribe and chisel were blunt and your saw’s kerf too wide? Attempted and failed miserably to communicate a new idea in a language lacking distinctions between the relevant concepts?
In each of these examples, the tools — the inadequate pan, the blunt chisel, the unnuanced word — are the weak links in the chain from skill to intent to outcome.
A quality criterion for tools
In McLuhan’s formulation, tools are extensions of the body through which we exercise skill to affect the environment to achieve the outcomes we intend.
Tools are channels for intention transmission, so a good measure of a tool is transmission quality. And the measure of transmission quality is accuracy and precision of transmission: the tool’s ability to achieve a result in a small and specifically chosen zone of the space of possible outcomes. A shorthand for that is high-fidelity transmission.
Tools which transmit intention with high fidelity don’t replace skill and intent, but are still essential for excellent execution through fine control. In the abstract, control requires skilled understanding of system causation, clear intent about the desired change in the system, and tools that transmit intent to the system with high fidelity. All three are needed.
For me, this way of thinking about tools in the context of intentional skilled action has implications for how we select and build tools. Still making up my mind about what these implications are, but there are at least three so far:
Understanding whether a tool is high fidelity requires a clear idea of causation and intention. The more blackbox-y a tool is, the harder it is to judge its fidelity in transmitting intention.
It’s misguided to try and build tools to replace/augment human work without articulating what skill and intent mean in doing that work. Building AI tools to replace well-understood work with clearly specified processes and success definitions is smart. Building tools to augment (but not replace) humans doing work that requires discretion and meaning-making also seems smart.
Skill, intent, and tool quality build on each other. Developing skill and clarity of intent is likely to accelerate development of better tools because the only way to recognise a good tool when you build it is to have skill and clear intent.
If you’ve been thinking about these matters too, I’d be curious about your thoughts.
See you here soon,
VT
Hi Vaughan, these thoughts strongly resonated with me. The three aspects of control (skilled understanding, clear intent, high fidelity tools) are useful for describing the shortcomings of generative AI tools today. They have limited understanding at best (due to gaps in their ability to represent not-knowing as you have described in your other work) and are limited by language in setting intent in ways humans don’t have to be. That said, part of what makes generative AI tools so compelling is that they are much more high fidelity for people not steeped in programming than anything available to them before.
There are some interesting parallels here with what Cesar Hidalgo described in his wonderful book Why Information Grows that are worth exploring further. In any case, these are exciting times given the ability of these tools to improve our ability to approach these questions more skillfully.