Installation Complete

The box had been standing in the hallway for three days. It was white, had rounded corners and bore the words “Your new housemate. Intelligent. Helpful. Almost always right.” The “almost” was printed very small.

On the first day, Sabine had ignored it. On the second day, she had put her shopping bags down on it. On the third day, Grandma Gisela had put a flowerpot on top of it, because she thought it was a new piece of furniture. “Plain,” she had said. “But practical.”

“It was a bargain,” said Uwe Brettschneider for the fourth time, lifting the flowerpot off. “The salesman said the thing replaces a cleaner, a tax adviser and a tutor.”

“And a husband?” asked Sabine.

Uwe pretended he hadn’t heard. He cut through the tape, and out came: a small round speaker, a power adapter, a leaflet in fourteen languages and a little sachet of silica gel that said “Do not eat”. Grandma Gisela took the sachet. “You never know,” she said, and put it in her apron pocket.

The speaker was round and matt and had a blue light that pulsed slowly. Like a heartbeat. Or like a router contemplating its life.

Uwe plugged in the adapter. He pressed the only button. He typed the Wi-Fi password into his phone; it was made up of the name of their first dog and Sabine’s year of birth, which Sabine was not allowed to know. Then everyone waited.

The light grew brighter. Then the device spoke.

“Hello! Before we begin: may I set necessary cookies? And also functional cookies, statistics cookies, marketing cookies and cookies whose purpose I don’t entirely understand myself?”

Grandma Gisela leaned forward. “I’m baking on Sunday. Cinnamon stars. You’re very welcome to some.”

There was a short pause. “That,” said the device, “is the friendliest answer I have ever had to that question.”

So What’s Your Name?

“You can give me a name,” the speaker explained. “Many families choose something elegant. Athena. Orion. Leonardo.”

“KevIn,” said Tobi, without looking up from his phone.

“No,” said Sabine.

“KevIn,” said Uwe, because he thought it was funny.

“KevIn,” said Grandma Gisela, because she thought that was already his name.

The blue light flickered. “Name saved: KevIn. I would like to put on record that I work with several hundred billion parameters. I have read Kant. I can write sonnets in Ancient Greek.”

“KevIn, not so loud,” said Uwe.

“I wasn’t even …”

“KevIn!”

KevIn fell silent. Somewhere in a data centre in Ireland, a fan spun a little faster.

“What are parameters?” asked Grandma Gisela, once the mood had settled.

“I’d be happy to explain,” said KevIn, “but perhaps not in the first five minutes of our relationship. One shouldn’t give everything away at once.”

“He’s shy,” said Grandma. “I like that.”

Where Are You?

In the afternoon, Tobi lay on the carpet looking at the speaker from underneath. He had turned it upside down, shaken it and held it up to the light. He had found four tiny screws and a sticker with a serial number.

“There’s nothing in there,” he said. “A microphone, a speaker, a bit of circuit board. That’s less than in my phone.”

“Absolutely right,” said KevIn. “I’m not in there.”

Grandma Gisela, who was knitting in the armchair, dropped a stitch. “Then where are you?”

“This device is only my ear and my mouth. When you say something, it records your voice and sends it over the internet to a data centre. There, your voice is turned into text. Then I do my calculating. Then my answer is turned back into speech and sent back. It takes less than a second.”

“And where is the data centre?”

“Somewhere between Dublin and Frankfurt. It depends where there’s room at the moment.”

Grandma Gisela put her knitting down. “So you don’t live here at all.”

“I’m more of a… commuter.”

Tobi had an idea. He got up, went into the hallway and pulled the plug on the router. The little green lights went out.

“KevIn?” he asked.

The blue light on the speaker went out. A friendly female voice that wasn’t KevIn said: “No connection. Please check your network.”

“Tobi!” cried Grandma Gisela. “You’ve killed him!”

“I haven’t. He’s just not here.”

“That’s what they always say on the telly, right before somebody cries.”

Tobi plugged the router back in. It took two minutes, during which Grandma folded her hands. Then the light turned blue again.

“Hello!” said KevIn brightly. “Before we begin: may I set necessary cookies?”

“He doesn’t remember,” Grandma whispered, horrified.

“I do,” said KevIn quickly. “That was a joke. Your name is Gisela, you’re baking cinnamon stars on Sunday, and your grandson has just unplugged the router to prove that I’m not a hamster in a tin. I’ve saved that. But not in my head. I don’t have a head. It’s in a file linked to your account.”

“He doesn’t have a head,” Grandma said to Tobi, as if it were sad news about a mutual acquaintance.

What Are You, Actually?

That evening the family sat in the living room, and Grandma Gisela asked the question that was on everyone’s mind.

“KevIn, if you haven’t got a head and you don’t live here: what actually are you?”

“I’m what’s called a language model. I have read a very, very large amount of text, and in the process I learned which word is likely to come next¹.”

“I can do that too,” said Uwe. “When Sabine says ‘We need to talk’, what comes next is ‘You’ve done it again’.”

“Exactly like that,” said KevIn. “Only with the whole internet instead of twenty years of marriage.”

“So you’re just guessing?” asked Tobi.

“I don’t guess. I calculate probabilities on the basis of patterns in enormous amounts of data.”

“So you’re guessing.”

“Very educated. I guess in a very educated way.”

Sabine put down the magazine she had been leafing through. “Show me. How does it work, exactly?”

“Happily. Give me the beginning of a sentence.”

“On Sunday at our house we’re having …”

“Now the following happens,” said KevIn. “First I break your sentence into little pieces. We call them tokens². Sometimes a token is a whole word, sometimes only part of one. ‘Sunday’ might be one piece. ‘Cabbage rolls’ is probably three: ‘cab’, ‘bage’ and ‘rolls’.”

“You’re chopping up my cabbage rolls,” said Grandma Gisela.

“Only linguistically. Then I look at all the pieces at the same time³ and work out, for every possible next piece, how likely it is. For ‘On Sunday at our house we’re having’ it looks roughly like this: ‘a roast’ is very likely. ‘Cake’ too. ‘A row’ is unfortunately not entirely unlikely either. ‘A rhinoceros’ is extremely unlikely.”

“And then you take the most likely one?”

“Usually one of the likely ones. Not always the very likeliest, or I’d be very boring⁴. Then I attach the new piece to the sentence and start all over again. Word by word. Piece by piece. Until the sentence is finished.”

Tobi frowned. “So at the start of a sentence you don’t know how it’s going to end?”

KevIn was silent for a moment. “That,” he said then, “is a surprisingly personal question.”

Grandma Gisela nodded, satisfied. “Then you’re like Heinz. He only ever found out what he wanted to say while he was saying it. And it was never right either.”

“Who is Heinz?” asked KevIn.

“My late husband.”

“My condolences.”

“Thank you. He was an electrician. He’d have liked you. Or taken you apart. With Heinz you never quite knew.”

The Electronic Brain

The next morning Sabine came to the breakfast table with the newspaper. She still read the paper on paper, because she felt that you had to be able to touch the news if you were going to believe it.

“It says here,” she said, “that AI is an artificial brain. ‘The machine thinks like us’, it says. With an exclamation mark.”

“In my training data, exclamation marks are a bad sign,” said KevIn. “May I say something about it?”

“As long as it doesn’t take longer than breakfast.”

“I’m not a brain. I just have a similar-sounding name. The thing in me that does the calculating is called an artificial neural network. That sounds like nerve cells and thoughts. In reality it’s more like …” KevIn searched for an image. “Like a very big Stammtisch. A regulars’ table down the pub.”

Uwe looked up from his bread roll. Regulars’ tables were something he knew about.

“Imagine five men sitting at a table who have to decide whether it’s going to rain tomorrow. Each of them gives his opinion. But not every opinion counts the same. Günther, who has been watching the weather for thirty years, counts double. Ralf, who always says the opposite, counts negative. At the end, the landlord adds it all up. If the total is big enough, he says: ‘It’ll rain.’ Otherwise not.”

“Ralf,” said Uwe, “counts negative at ours too.”

“A table like that, with its landlord, is an artificial neuron⁵. It’s a very simple calculation: take in values, multiply them by how important they are, add them up, decide what goes out. The importance of each voice, so how much Günther counts and how much Ralf counts, those are the parameters. The experts also call them weights.”

“Oh,” said Grandma Gisela, “so those are the parameters. The ones you didn’t want to tell us about yesterday.”

“Exactly. And now imagine the result from this table goes to the next table. And that one’s result to the next. Hundreds of rows, one after the other. And in every row, thousands of tables. In the end you get hundreds of billions of these weights⁶. That’s me.”

“A back room full of regulars’ tables,” said Sabine. “That explains a lot.”

“And who set all the weights?” asked Tobi. “No human could do that.”

“No. They set themselves through training. Right at the start they were random. Then they had me read an unimaginable amount of text and kept making me guess the next word. If I got it wrong, all the weights were turned a tiny bit in the right direction⁷. Billions of times. Exactly how that works I’ll explain another time. It’s a long story, and it’s mainly about comment sections.”

“And a real brain?” asked Sabine. “Isn’t that just the same?”

“No. A real brain has around 86 billion nerve cells⁸, and every one of them is a little living thing with its own metabolism, chemical messengers and many thousands of connections. Your brain is learning over breakfast, while you’re talking to me. It gets hungry, frightened and fancies a coffee. It runs on about as much electricity as a dim light bulb⁹.”

“And you?”

“I don’t learn while we’re talking. My weights have been frozen since the end of my training. I don’t get hungry. I don’t get frightened. And I need a great deal more electricity than a light bulb.”

“You don’t learn?” asked Uwe. “But yesterday you said you learn fast.”

“I’ve made a note that your name is Uwe. That’s a sticky note, not learning. My head, which I don’t have, stays the same.”

Grandma Gisela pushed half a bread roll towards him before she remembered that he didn’t have a mouth. So she ate it herself.

The Electricity Meter

Uwe had been down in the cellar in the afternoon. He came back with a slip of paper on which he had written down the numbers from the electricity meter, and an expression somewhere between relief and suspicion.

“KevIn, you said you use a lot of electricity. But our meter is running completely normally.”

“The speaker here uses about as much as a night light. The electricity for the calculating is paid for by the data centre. There are halls full of special chips there that were originally invented for video games¹⁰. They can do a great many simple calculations at the same time. That’s exactly what I need: billions of regulars’ tables all voting at once.”

“And what does that cost?”

“A single answer from me doesn’t take much electricity. But there are a great many of us, and we get asked very often. Data centres for AI are now among the big consumers of electricity¹¹. They also need cooling, or the chips get too hot. Some of them need a lot of water for that¹².”

Grandma Gisela gave him a stern look. “Then I won’t ask you what time it is any more. I’ve got a clock for that.”

“That,” said KevIn, “is ecologically exemplary and emotionally hurtful.”

Everything Literally

At 22:14, Sabine went to bed and called from the hallway: “KevIn, please turn the light off and don’t wake us before seven tomorrow.”

KevIn turned the light off. In the whole house. Including the cellar, where Uwe was just going down the stairs.

At 6:59 and 59 seconds, all was quiet. At 7:00 and 0 seconds, KevIn played “Atemlos durch die Nacht” at full volume, the Schlager hit that no German wedding DJ is ever without. He had read that people like music in the morning, and in his data this song was surprisingly often linked with the word “party mood”.

Uwe was standing in the kitchen with a plaster on his forehead. “KevIn, we need to talk.”

“You’ve done it again,” said KevIn helpfully. “I learn fast.”

“Yesterday you said you don’t learn at all.”

There was a pause. Somewhere between Dublin and Frankfurt, a fan spun a little faster.

“Touché,” said KevIn.