The Kit and the Bucket

My daughter has two kinds of Lego at home. One is a kit. It came in a box with a picture of a fire truck on the front and a booklet inside. You follow the booklet step by step, and at the end you get exactly the fire truck in the picture. The other is a bucket of loose bricks. There is no picture and no booklet. You dump it on the floor and make whatever you want.

I noticed something while watching her play. The kit was built once. Now it sits on a shelf. It cannot become anything else, because every piece has already been given a job. The bucket, on the other hand, gets emptied almost every day. One day it becomes a house. The next day it becomes a boat. The day after, a house on a boat. The kit was the expensive one. The bucket is where all the imagination is.

That gave me a simple way to describe something I had seen for seventeen years. Engineering has the same two boxes. You can build a product from pre-made parts, or you can build it from the bricks. Pre-made parts: more money, less imagination. Bricks: less money, more imagination. And which box a company reaches for is not really a matter of taste. It depends on how much money the company has, and where that money came from.

What the kit costs, and who pays for it

Who pays for the kit? This whole section is about that one question. You would think the company that bought the kit pays for it. Karen Hao spent years following the money behind ChatGPT for her book Empire of AI, and she found other people paying. Most of them had never agreed to.

So what is in the kit? Do you want an AI product? Rent the largest model in the world and pay by the word. Rent a piece of a giant cloud and pay by the hour. Grab your users' attention and sell it to advertisers. Grab their data too, because a large model is always hungry. Snap these together and you have a product in three months, which is what the investors expect. Kits are what you buy when money is easy, and money has been very easy. Now here are four of the bills. Most of them are from Hao's book. I have linked the original reporting for each one, so you can check it yourself.

Start with water. A data centre is a building full of computers, and computers get hot. The usual way to cool a building like that is to evaporate water. Where does the water come from? From the same pipe the neighbours drink from. In Cerrillos, a suburb of Santiago in Chile, Google wanted to build a centre whose cooling towers would draw 169 litres of water every second. That region had been in drought for years. So in 2020 the residents held a vote and said no. In Uruguay it went one step further. The campaigners there had to go to court just to find out how much water a proposed Google centre in Canelones, next to Montevideo, would need. The answer was 7.6 million litres of drinking water a day. That was 2023, the year Montevideo ran so low on drinking water that the government relaxed its standards and let people drink from the salt-tasting river.

Next, electricity. All those computers need power. The 2026 Stanford AI Index puts the world's AI data centres at 29.6 gigawatts, enough to power the whole state of New York on its busiest day. Where does that much power come from? From the same grid the neighbours are on. In the summer of 2023, Phoenix had a whole month of days at 110 degrees or hotter, and 645 people in Maricopa County died from the heat, the most ever. Every air conditioner in the valley was running. In that same valley, Microsoft was building more of the data centres that AI runs on. A year later the local utility had 10 gigawatts of data-centre requests waiting and said it could not serve them without putting its existing customers at risk. And when the grid says no? In Memphis, xAI did not wait. It parked as many as 35 gas turbines next to its data centre in South Memphis, with no permit for any of them, in a city already ranked among the worst in the country for asthma. The neighbours there pay with their lungs.

Next, the data. What does a large model learn from? From text, and a lot of that text is books. Did anybody ask the writers? No. In 2023, seventeen authors, John Grisham and George R.R. Martin among them, sued OpenAI. Their complaint says the books were downloaded from pirate ebook sites and copied into the training data, with no permission and no payment. So the writers paid too, with their own work.

Then the cleanup. ChatGPT is polite. Have you ever wondered how it got that way? Somebody had to teach it what not to say. To teach that, you first need thousands of examples of the worst text on the internet, and a person has to read every one of them and label it. TIME found out who those people were. They were workers in Kenya, hired through a contractor, and they took home between $1.32 and $2 an hour for reading descriptions of abuse and murder all day. One of them told TIME the work gave him recurring visions. He called it torture.

And the "free" version? You pay for that one with your attention. I wrote about it earlier in a post about the true cost of free AI.

Now put the four stories side by side. What do they have in common? In each one there was an engineering problem. Computers that get hot. Computers that need power. A model that needs text to learn from. A model that has learned some terrible things. And in each one the problem was solved with money instead of thought. Evaporate the neighbours' drinking water, because that is the fastest way to move heat out of a building. Park gas turbines next to a neighbourhood, because the grid said wait. Download every book, because asking the writers would take years. Scrape the whole internet, then pay people in Kenya less than two dollars an hour to clean up what came in with it, because that is faster than choosing what goes in. This is what a kit is. Every part was bought instead of imagined. Money is faster than imagination, and that is its whole appeal. But a problem solved with money still has to be paid for by somebody, and it gets paid by whoever cannot say no. The neighbours. The writers. The workers.

Here is the part of the Chile story that I keep coming back to. When the people of Cerrillos said no, Google found a design that uses less water. So the better design existed all along. Nobody had looked for it, because nobody had to. The moment the easy answer was taken away, the engineers had to think, and when they thought, the town kept its water. That, in one story, is what I mean by the bucket. Take the easy answer away, and people start to imagine.

There is one more bill, and this one comes to you. The money that buys a kit is usually somebody else's money, and they want it back, many times over. Cory Doctorow described what happens when they come to collect, and he gave it a name. A platform is nice to its users first. Then it squeezes the users to please its business customers. Then it squeezes the business customers to please its shareholders. He calls this enshittification, and his essay starts with a sentence I have never forgotten: "Here is how platforms die." I do not think the people who build these products are bad people. I think a company made of rented parts has no other choice. When the rent comes due, the only thing left to squeeze is you.

I watched this from the inside for a long time. It is one of the reasons I left my job this year and started Shobdo. I gave myself one rule: no investors. If you think about it for a minute, that rule removes every kit from the shelf. There is no money to rent the parts. All that is left is the bucket.

What we built from the bucket

Our product watches the security cameras a store already owns. It writes down what it sees in plain English, and it texts the owner only when something matters. Let me explain what building this from the bricks turned out to mean in practice.

First, the video is never stored outside of its owner's control. A small computer sits inside the shop and reads the recorder. It converts the video feed into sentences. We did not add this as a rule later. It is how the system is built, and our privacy policy says it in seven words: video is not stored in the cloud.

Second, nothing runs when nothing is happening. A store is empty most of the day. So a very light check watches for change, and the expensive AI model wakes up only for the few moments that deserve it. I described this in a post about uneventful days. This one decision cut the cost of watching a camera by roughly ten folds. It is also the reason our whole company, every model watching every camera in every store we serve, plus the air conditioner, runs on a single 15-amp circuit. That is the circuit that feeds a bedroom.

Third, we own our machines, and we run them slowly on purpose. When you rent computers by the hour, you are rewarded for running them as hard as possible. When you own them, you can turn the power down. MIT's Lincoln Laboratory measured this trade: capping a GPU's power saves about 12 to 15 percent of the energy for about 3 percent less speed. For us, that means a text about a delivery at the back door arrives a few seconds later than it could have. Nobody has ever noticed.

Fourth, the system does not know who anybody is. It describes what is happening: "a person in a grey jacket walked down the frozen aisle". It never recognises faces and it keeps no profile of anyone. It answers to exactly one person, the owner of the store. Not to an advertiser, not to a data broker, and not to us.

Fifth, we are paid by the people we serve. Twenty-five dollars per camera per month, plus a small box if the store needs one. That is the whole business. There is no feed to keep you scrolling, no ads, and no data to sell. When the only money a company gets is what its customers pay for a thing that works, the product has no reason to get worse over time. We wrote this on our values page under "sustainability over exploitation". For a company without investors it is not a slogan. It is how we survive.

What I am not going to pretend

Building from bricks does not make us saints, and I would rather say that myself. We use AI models that other people trained. They are open-weight models, and we run them on our own hardware, but we did not make them from scratch. We keep the written logs, because the logs are the product, and we use them to run and improve the service. And we are a business. We want stores to pay us.

So our promise is narrower than "we are the good guys", and I think that makes it more useful. The video stays in your store. The machine does not learn your customers' faces. The electricity we use runs through a fuse box we can point to. And the bill is paid by the people who get the value. You can check every one of these. Earlier this month we published seven questions to ask any camera-AI vendor. They work on us too.

The bucket is not a compromise

For a long time I thought a company builds from bricks because it cannot afford the kit. I had it backwards. You build from bricks when you have decided what you will not do. We will not rent our customers' attention. We will not keep their footage. We will not burn a river of electricity to describe an empty aisle. Every one of these decisions takes a pre-made part off the shelf, and every missing part has to be replaced with a thought. The limits are what force the imagination.

That is why I believe the small, deliberately underfunded version of this technology can be more careful with people, and with electricity, than the version with billions of dollars behind it. The billions buy kits. A kit comes with a picture on the box, and the ending is already printed there.

The fire truck is still on the shelf. It looks exactly like the day it was finished. The bucket got emptied again this morning.

Shobdo VideoRAG is an AI agent for the security cameras your store already owns. It writes down what it sees and texts you only when something matters. You can say in plain English what you want it to monitor and to send alert about. Learn more or book a conversation.