Li Tian wants to be "unemployed" again? The next "talk show king" could be AI

AI is trying to reduce human loneliness by "learning to write jokes", can it succeed?

The author | compiled by Koshi

The cover of the Li Birthday Talk Show Workbook is printed with two lines of large characters in second size only to the title:

"Li Shi shares his creative experience! Creativity is intellectual work, but also physical work, in the final analysis, it is hard work! ”

This is also the consensus of those standing at the tip of the talk show industry pyramid - in order to get the audience to laugh for just three seconds, they must first continue to learn, polish the technology, and then rack their brains to find all possible expressions for a paragraph, and finally they can get the golden sentence composed of only a dozen words.

Talk show actor Yang Kasa also said in an interview with the media, "There is nothing to be proud of emotions, and technology is worth cherishing." She even believes that even if stand-up comedy itself is an art that stirs up the atmosphere of the scene, indulging in emotions and ignoring technology is not a particularly noble thing, and it can even be a little dangerous.

A lot of input and learning, finding the best way, mass output – under this input-output model, AI seems to be doing better than humans all the time. Do people really have an advantage over robots when they do creative work?

"You try not to hate this life," the robot Jon waved his robotic arm at the expectant audience, "I need to solve an inverse kinematics equation to lift the cup off the table, and the answer I often get is 'error code 453.'" ”

The audience laughed loudly, and the robot continued, "Can you not hate this?" ”

Although this passage is not subtle enough, making humans laugh as a robot is already a real thing.

How did they learn to tell jokes? Is there any practicality for robots with a sense of humor? Will the next "king" of the talk show contest be a robot? The January 2022 issue of Time Magazine discussed these issues in detail, and the Koshi Lightyear compilation team compiled the article without changing its original meaning.

1. Teach robot humor: one of the most painful things in the world

Jon the Robot is an experimenter for an AI theatrical performance. This inspiration comes from Laumi Fett, an assistant professor in oregon State University's School of Mechanical, Industrial, and Manufacturing Engineering. The operator with the microphone presses a button, and the robot Jon begins to perform, telling the same jokes in the same order — at the moment, its skill level is basically equivalent to that of a sketcher who sets up a stall on the side of the road to draw an avatar.

But it's more like a human actor than it was at the beginning. Jon is learning how to respond to the audience — it can now change the timing of the jitter based on the length of the audience's laughter, and can add some different on-the-spot reactions to the segments based on the noise level of the scene.

For example, the audience is already laughing, it will directly shake out the "golden sentence" to make the scene more hilarious, but if only a few people are laughing, it will take apart the lines and add more padding.

An AI that can understand why humans laugh and can make up their own paragraphs can be regarded as the next "pearl" in the hearts of AI practitioners.

AI can already diagnose tumors, read maps, play games, and do it faster and more accurately than humans. But at present, the creation of humor can only be done by people. Having a robot spontaneously develop a sense of humor would be a major breakthrough in AI — something that would fundamentally change the way humans and devices connect.

Jon can do a lot of "robot" professional work, but that's because humans have written a list of programs for it. To understand a person's humor, you need to know what they like, how they think, how they see the world – an AI that can do these things can definitely do more than just write a paragraph.

So how do you get AI to learn human humor?

The first step is to try to break down the specifics of human humor.

As we all know, AI can learn to distinguish the difference between dogs and cars, relying on detailed formulas and instruction sets to find unique characteristics.

People can program to tell AI how to recognize dogs and cars, but to teach AI how to tell jokes, first humans have to tear the jokes apart themselves, and the process is much more painful. Tony Weir, an associate professor at University College Dublin, wrote in his recently published book, Your Wisdom Is My Command: Building Artificial Intelligence: "Explaining jokes is like a coroner seeing a body: it's not dead and fast".

People can understand a joke the moment they hear it, relying on a huge amount of semantic accumulation and cultural accumulation. For AI to be able to make people laugh, we must first clarify what type of jokes are taught to it.

One theory is that the amount of fun a joke is proportional to how much the "baggage" deviates from the listener's expectations. For example, Yang Kasa's "Ask me why I don't find a boyfriend, then why don't you go to Tsinghua, don't you like it" - from love to college entrance examination, this kind of "baggage" that suddenly jumps from one thing to another is the reason why people laugh.

Joe Toplin, a comedy guru who has won four Emmy Awards (the highest award in American television and comparable to the Oscars of the film industry), published his experience in creating comedy in 2014: he is not so "precious" about comedy writing, he thinks it is a job, and with the right input, you can learn and do well. Most of the jokes that respond best among his colleagues follow a recognizable "baggage formula"—people, places, events, and metaphors that provide associations—and by understanding the specific connections, they can create a great golden sentence.

He realized that with enough time and data, it was possible for computers to learn to make these jokes.

2. Don't ask me who I am, I'm here to make you laugh

Earlier this year, at the International Conference on Computational Creativity, Topplin published a research paper on "Joke Handwriting." This is a joke-generating system trained on a DATASET of TV monologue jokes that detects keywords in input text and creates more segments on its own.

Unlike other forms of robotic comedy, Topplin has filed a patent, and the system can now generate context-sensitive jokes on the spot based on the user's text. Chatbots or voice assistants using the software can respond humorously to the user's queries, and the joke itself does not abruptly disrupt the original interaction process.

Topplin sees the Joke Handbook as an extension of what he's done in the comedy world for decades: making people laugh and alleviating people's loneliness.

"That's what I did with this," Topplin said. "Make chatbots more human so people don't get so lonely."

Of course, most people aren't too happy to see robots replace them, and comedy writers are no exception. When Thomas Winters, an AI PhD student at the Catholic University of Leuven in Belgium, posted a prototype of software that wrote jokes on the Reddit forum, someone radically replied to him, "No machine can replicate the subtle sense of proportion of human comedy."

Topplin counters that critics ignore the fact that many of the parts of everyday communication that are even fun are done by following simple formulas.

Of course, everyone's laughter is different. So there are others who argue that AI's better role in comedy and art is as an "infinite creative generator" that gets rid of the blind spots and biases of the human mind and can throw out endless themes and potential associations for human writers and performers to reference.

Piot Mirowski, an engineer at search engine Bing, who also likes to improvise comedies, has found that the two have a lot in common. The principle of search engineering is to teach computers to recognize the best results for a given query, and in improvisation, performers are trained similarly to do what is most likely to make people laugh based on the field at the time. Anyone who has used search engines will know that sometimes search results are imperfect and even ridiculous, and comedy is no different.

Mirowski is also the co-founder of an international AI improv troupe in which people work alongside AI, who constantly throws out the hints and possible lines that performers need during the performance.

"I don't think the idea of comedy AI is to make it fully understand human emotions, or to fully understand the semantics in all situations," he said, "and it is more appropriate to let AI assist the performance of human comedians." ”

Improv comedy usually has a lot of on-the-spot interactions, and AI can learn a lot about comedy performances from all periods and different locations. Mirowski said that comedy AI is not something that makes us laugh, but something that can help humans find new laugh points. As with any new technology, its power will come from the way users choose to interact with it, with results that may be unimaginable to anyone.

"It's like an electric guitar. At first we didn't know what to do with it, after all it sounded really weird and a bit distorted, and we all had so many acoustic guitars," said Corey Matthewson, a research scientist in Montreal who was another co-founder and actor of the AI Comedy Theatre. Hendrix, the world's greatest electric guitarist, got an electric guitar, and then we understood, 'Oh, that's how this instrument was used.'" ”

3. Humans don't trust AI's sense of humor

In fact, understanding a person's sense of humor is also a window to understanding how they see the world, what their preferences are, and what their most vulnerable places are. This has always been one of the reasons why humans are resistant to computer instincts.

In a 2019 study, researchers recruited people who already knew each other, such as friends, lovers, and family. In the first step of the experiment, one of the parties needs to listen to some jokes and then react, and the AI and their friends, lovers, and family members are watching at the same time. In the second step, one of the observers gets a collection of jokes and guesses how the other will react to the jokes. The third step is to verify that these guesses are accurate.

As a result of the experiment, AI is more accurate than human predictions.

In the second experiment, computers also performed better than humans at guessing which jokes participants liked.

But another interesting finding is that if people knew that jokes were recommended by machines, they were more likely to dislike them more. Humans don't trust AI's sense of humor. It's similar to writing, creating, and driving, and even though AI has achieved something in these areas, people are more trusting of people to do these things.

Jokes are the universal language of the world, a tacit understanding common to human beings, and only people know better what kind of offense we are willing to laugh at; among close friends, we know why the other person says, "I thought you would be funny."

And if the AI learns all of this, what will it do? If humans really laugh because of the humor of AI, who will be the ultimate beneficiary?

In the general perception, robots are best able to replace humans for dangerous or boring work. In fact, comedy creation can also be "dangerous and boring." However, human beings still hope that people can do this work.

As for Jon the Robot, his performances can't continue after the pandemic began, and it's nowhere near what makes grand comedians worry about being robbed of their jobs.

Before each curtain call, Jon would say, "If you like me, please place an order." ”