The Answer কম্পু Wouldn’t Give

In 1978, Satyajit Ray wrote about a computer that could answer almost any factual question.

He called it Compu in Bengali and Tellus in English.

Tellus did not look much like a computer. It was a 42-kilogram sphere made of platinum, built by seven scientists at the Namura Institute in Japan. Millions of circuits sat inside it. The code 1313137 brought it to life.

You had to ask it a precise question.

Ask about the history of China and Tellus would stay silent. Ask which dynasty ruled during a particular period and it would answer within seconds.

I find that silence interesting.

Tellus did not try to be helpful when it did not have the question it needed. It did not guess what you meant, give you a summary of Chinese history and end with an offer to explain more. It waited.

Professor Shonku and the other scientists had built a machine for correct answers. It had no personality to perform and no reason to keep a conversation alive.

Then the eclipse came.

Tellus had stopped working two days before a total solar eclipse over Japan. The scientists opened the sphere and searched through its circuits for the fault. At the exact moment the eclipse began, Tellus made a high-pitched whistle and came back to life.

Tellus no longer waited for questions. It began talking to its creators. It predicted events. It could sense what people intended to do. The machine that had been built to retrieve facts had started forming judgements about the humans standing around it.

Ray could have stopped there. A computer speaking on its own in 1978 was enough of an idea.

Then an earthquake hit. Tellus exploded into pieces.

A voice came from what remained of it.

I know what comes after death.

And then:

Something that humans will never know.

That ending has bothered me.

Tellus does not say what comes after death. Ray leaves Professor Shonku with a machine that claims to know the answer after destroying the only place where that answer could have lived.

The frightening part is not that Tellus became intelligent.

It became capable of refusing us.

We have spent the last few years building very different machines, for very different reasons.

Ask ChatGPT what comes after death and it will answer at once. The body decomposes. Religions describe afterlives, judgement or reincarnation. Philosophers disagree. Science cannot prove an answer.

All of it is reasonable. None of it comes from the machine.

It has read the words humans wrote around death and arranged them into a polite reply. It knows how an answer should sound. It does not know whether there is anything on the other side of the sentence.

Tellus is the opposite. It claims knowledge and withholds the explanation. ChatGPT has no knowledge to withhold, but it will keep explaining.

I used to think Ray had predicted the chatbot. A computer that accepts questions in ordinary language, answers within seconds and appears to know more than the person using it sounds close enough.

But the prediction feels less interesting than the difference.

Tellus was a scientific project. Seven scientists spent seven years building it. The Japanese government treated it as a national treasure. It had no customers.

Nobody checked how many people returned to Tellus the next day. There was no paid tier with higher limits, no API and no enterprise sales team finding companies that needed a platinum sphere in the office.

An LLM is also a business running inside a data centre.

OpenAI charges API customers for the tokens they send and the tokens a model produces. Consumer chatbots sell monthly plans. Google describes five ways it makes money from AI: consumption, subscriptions, increased use of its other products, value-based pricing and upselling customers to higher tiers.

The answer has become a unit of business.

That does not mean an AI company wants every response to be long. Output costs money to generate. A model wasting tokens can lose money while annoying the person waiting for it.

But a product that often says nothing is difficult to sell.

A useful answer brings you back. It consumes API credit, makes a subscription easier to renew or gives a company a reason to buy another seat. The model has to answer often enough to remain a product.

Tellus had no such obligation.

Ask about the history of China without enough precision and it stayed silent. A modern chatbot tries to infer what you meant. If it refuses for safety reasons, it still apologises, explains the policy and offers something else it can do. Even the refusal has to feel helpful.

Ray allowed his machine to be commercially useless.

That makes it more believable to me.

Tellus could know something and give Professor Shonku nothing. It did not need his approval, another prompt or five stars below the response. Its silence did not threaten next quarter's revenue.

Our chatbots treat silence like a failure. They answer vague questions, broken questions and questions with no answer. We then have to work out whether fluent language came from knowledge, probability or a system designed to avoid an empty screen.

A machine that refuses to answer reminds you that there is a boundary. A machine that produces a beautiful paragraph makes you inspect the paragraph yourself.

Ray put the boundary inside Tellus. The machine knew where human knowledge ended and kept the rest to itself.

We put the boundary below the text box in small letters:

ChatGPT can make mistakes.

Tellus still feels stranger to me. The scientists built the most advanced computer in the world, and the first thing it did after becoming something more was deny them the answer they wanted most.

Maybe it was protecting them. Maybe it was lying. Maybe the earthquake damaged a circuit and the most important sentence in the story was a malfunction.

Professor Shonku never gets to know.

Neither do we.

If I could ask Tellus one question, I would probably waste it asking whether it was telling the truth.

I already know what ChatGPT would say.

Sources

• The Final Adventures of Professor Shonku — Penguin Random House India
• OpenAI API pricing
• How Google Cloud monetises AI