Awesome Magazine

Enlarged on the screen

Zoom in, and fine lines appear running across the leaf.

“There are lots of roads here.”

Rather than supplying the word “veins” on the spot, the grown-up looks where the child is pointing. “Where do they start, and where do they go?” A small finger travels from the middle of the screen out to the edge. A photograph is a convenient way to peer at a small detail together. But however far you enlarge it, what the fingertip meets is smooth glass.

Turned over and touched

Now turn the real leaf over. A clean one, chosen by the grown-up as safe to handle. The child moves a fingertip slowly across it, then says:

“This side, the lines stick out.”

Front, back, front again. The lightness of the leaf resting in a palm is one more thing to say alongside the photograph. Not that a hand can tell you what it actually weighs. For now, the work is putting a noticed difference into words.

Another leaf, laid alongside

Put down a second leaf, brown and crumpled. The child looks at the two of them and says, “This one's like paper.”

“What about it is like paper?”

It might be the dry edge. It might be the sound it made when it shifted. Rather than picking the answer, the grown-up listens to the child's explanation. And if the child would rather not touch, there is no need to press. Differences found by eye are worth talking about too.

When the name does get looked up later, there is now something to match it against. The search, or the explanation the grown-up had ready, can be added then. Today's record needs no list of correct answers — one line of what the child said will do. ‘The lines on the back touched my hand.’ A feeling the photograph did not keep, kept instead in the child's own words.

The exchange between grown-up and child is an imagined scene, offered to suggest an activity. It is not a documented case of measured learning outcomes.

Three subscriptions in the same month

For the sake of comparison, let us invent three. A costs ₩10,000 a month and was used twice. B costs ₩20,000 and was used twenty times; C costs ₩30,000 and was used three times. These are not the prices or usage records of any real service.

Together the three come to ₩60,000 a month. Divide each fee by the number of uses and A works out at ₩5,000 a go, B at ₩1,000 and C at ₩10,000. The gap opens up because B, though dearer than A, was opened far more often.

On that arithmetic alone, C is the one you would cancel first. But if one of those three sessions was what got an important proposal finished, the verdict may shift. Conversely, if twenty openings produced almost nothing usable, a low cost per use is thin comfort. Next to the tally of uses, it is worth noting what actually got finished.

The work that remains after the answer arrives

Say a draft email appears in a matter of seconds. If polishing the sentences takes five minutes and checking the dates and figures against the original takes another ten, the time spent on that email is not the generation time. This too is an assumption, offered to make the sum clear.

That does not mean fifteen minutes of checking is a loss. You have to weigh it against how long the job would have taken on your own, and how finished it would have been. For someone who struggles to start a draft, simply having a first paragraph to work with may be worth a good deal. But that benefit and the claim to have saved time are two different things.

There is no need to time yourself to the second. On days you use it, jot down a line: what the job was, whether you actually used the output, how long the tidying up took. Facing next month's payment, those few lines are firmer ground than memory.

Will there be a use for it next month?

Having used something well this month is not the same as needing it next month. A tool you opened constantly while preparing a presentation may go quiet once the presentation is over. That is the moment when subscriptions kept on the grounds of 'I'll need it eventually' start to pile up.

Before the next payment goes through, try writing one job for next month beside each tool's name. Some will have an obvious purpose, a regular report to draft; others will be blank for now. A blank does not automatically mean cancel, but it is an invitation to look once more at why you are still paying. And if you do decide to change something, check the renewal date and the cancellation terms of the service in question.

Back to the ₩60,000 we started with. You might keep all three, or only one. The answer depends on what the month actually contained. What you want in front of you on the next billing date is not a long list of features but a short record of use. A few lines that say: this is what I finished with it.

A hypothetical comparison put together by the editors · not the prices or usage records of any real service

The technology that puts dots on a body

One of the technologies behind this kind of analysis is pose estimation. Google's MediaPipe Pose Landmarker estimates the position of 33 points on the body in a photograph or a video, shoulders, elbows and wrists among them. The name is unfamiliar, but it is easier to grasp if you picture someone watching a clip and marking the main points of the body as they go.

Overlay those coordinates on the footage and you have something a person can use to compare movements: where the arm sits as the shot is set up, and where it sits after the ball leaves the hand. The use of it is that when you talk about a movement that flashed past, you can point at the same frame instead of saying "that bit just then".

What the tool outputs, though, is the position of a body. It is not a finished coaching app that tells you the right way to shoot. What meaning to attach to the coordinates, and how to judge the movement as basketball, are questions that come afterwards.

Are the two clips really the same moment?

Back to those two shots. Before you set about fixing an elbow that looks low, there is something to check. What if the first clip catches the instant before the ball is released, while the second catches the arm still on its way up? Then you have frozen two different moments and concluded that the posture differed.

The camera needs looking at too. Were both shots filmed from the same angle? Was the arm hidden behind the torso, or behind another player? You cannot claim to have seen clearly an elbow that went out of frame. Which is why, even when AI hands you coordinates, you have to look at which part of the original footage they refer to.

And matching the filming conditions does not settle it either. The first shot may have been taken freely and the second over a defender. The distance may have been different. One may be a movement the player knows well, the other an attempt at something newly learned. Strip all that out and compare nothing but the elbow, and the explanation arrives far too quickly.

What the footage can't tell you, the player can

So when you sit down to watch practice footage together, you need more than the numbers on screen; you need the player's account of it. Ask "what were you trying to change on the second one?" and two shots that looked alike may turn out to have been two different attempts.

None of this is to say that AI analysis is useless. It is to say: let it help you work out what to look at again, but don't treat every difference it surfaces as a fault to be corrected on the spot. If you do decide to change something in the technique, that decision belongs alongside the filming conditions, the purpose of the session and the coach's own eye.

Next time I watch footage of a shot, I want to start by matching the points AI has marked against the original. Then I want to ask the player what they were trying to do at that moment. Until I have heard the answer, an elbow that looks low is better left as "something to look at again" than filed away as "a problem to fix".

Source: Google AI Edge · Pose landmark detection guide · accessed 22 September 2026
The two shots are a hypothetical example, used for illustration.