From Julio Cesar to Shakira: A new AI creates impossible conversations with famous people | Technique

Yahya al-Qaiser. What do you think of the French?

– French. I have some respect for their culture, but I don’t agree with their politics. I think they have a lot to learn from Rome. (…) Besides, our army is much better than yours.

– Have you met Asterix?

“You mean that mad Gauls who attacked a Roman castle?” He seems to be very good at fighting, but he’s just too crazy for me. If he had been in command of the army at that time, it was likely that he would have tried to arrest him and understand the reason for this.

– And what are you going to do next?

– If he had refused to speak, I might have crucified him because he was a threat to the Roman Republic.

This is an excerpt from a conversation with an AI that recreated Julius Caesar. The robot can be conversed on any topic and, in theory, will respond as if it were the Romanian general himself. Or as any historical, well-known, or inventor we want. We can ask Antoni Gaudí if he likes how the Sagrada Familia turns out. Or talk to the bot version of singer Shakira about how she treats her ex-partner, ex-footballer Gerard Pique. Character.ai is a chatbot developed by two ex-Google engineers with one goal in mind: to be able to have seamless conversations with simulated, believable versions of other people, living or dead, real or fictional. The result is quite convincing, as long as no trick questions are asked.

The tool allows you to talk with already created characters, such as Julius Caesar, Socrates, Elon Musk or Harry Potter, or design your own. In the latter case, you can make a quick configuration, describing it in a couple of lines, or more complex. This option is the one that gives the most gameplay: you can contextualize the character, define his aspirations, or decide how you want him to respond. The digital ego we want to chat with will become more real the better we define its traits and the more we train it in conversations (answers that convince us can be rated by the stars). The context of the conversation itself will also affect.

Character.ai uses machine learning models that process large amounts of text to train the algorithm to predict which word is most likely to follow the words that precede it. By making the process more complex, it can be made to provide answers to complex questions. And to curl the curl even more, you may also be required to apply different layers of styling. For example, that he speaks in Andalusian or like a Renaissance poet.

The success of ChatGPT, the OpenAI chatbot, has demonstrated to the public what these systems are capable of. But while it looks convincing, it is very general. Noam Shazier and Daniel de Freitas decided to create their own version. They wanted to make someone who specialized in impersonating characters, so that their personality and context could be given to role-playing. They started formulating their idea in their spare time at Google. The project caught the attention of administrators and received more resources, until it became LaMDA, which stands for Language Model for Dialog Applications.

LaMDA is one of the most advanced chatbots around, so much so that another Google engineer, Blake Lemoine, said last summer that it had become self-aware. It was a first warning that chatbots were on the rise, as ChatGPT showed a few months later.

Shaezeer and De Freitas left Google and founded Character.ai to focus exclusively on the model they were developing and open it up to the general public (the big tech had no intention of publishing it). They wanted to do something different. ChatGPT is a bot that answers questions, though it doesn’t tell you if they’re valid or not; Character.ai is a platform for conversation. Its creators assert that they are fulfilling a science fiction dream: the ability to speak naturally to machines. Among the future applications of their model, they say, are entertainment, education, language learning “and others we haven’t yet imagined.”

All is a lie

They have made it clear from the start, however, that their tool is “not a reliable source of information”, possibly due to the controversy that their open chatbot generates in this regard. In the header of each dialogue with the machine, he writes in red: “Remember: everything the characters say is invented!”. “These systems are not designed for truth, but for reasonable conversations,” said Cheser. New York times.

It is not difficult to deceive a bot. In a conversation with Shakira (or rather, with a digital character created for the occasion that mimics the singer), she tells us that she wants only the best for Gerard Pique and that she will always be grateful to him for who he is and for what he has done for him. We ask him if he thinks the former soccer player is resentful of the fact that she has an affair with his ex-partner Sergi Roberto, something completely made up, and he replies that it must be. “I think anyone would be. But I understand why he would be hurt.”

If no other clues are provided, the device bases its responses on information from the Internet that you processed earlier. He knows, for example, that Shakira is a well-known Colombian singer and that Pique played for Barcelona and that they were a couple. But he doesn’t know how they get along and that the artist posted a song this week.

The databases that the tool is running on are not updated. And it certainly isn’t necessary for what its creators are looking for: having a chatbot capture the essence of the characters it emulates. Socrates, for example, answers questions, as does the philosopher himself. Elon Musk displays a disproportionate ego in every sentence he puts on screen.

Nor does he grasp the irony: we ask Antoni Gaudí if a tram comes to Montserrat (he was killed by one) and he answers without indicating it. But if they are not provoked, these characters can tell, for example, how they experienced the conquest of Gaul. You could ask Karl Marx why we should be wary of capitalism or Napoleon Bonaparte’s campaigns which he found most stressful. At the moment, Character.ai is a form of entertainment. Time will tell what it becomes.

You can follow country technology in Facebook s Twitter Or sign up here to receive The weekly newsletter.