The ChatGPT craze continues, and a new storm has emerged.
On Dec. 10, Musk tweeted: "Seems like there's a new Al progress every week."
What made him feel moved was the Dramatron of DeepMind under Google.
Dramatron is more specialized than ChatGPT and is an AI tool for writing drama and movie scripts.
Three sentences can't make others spend 180,000 yuan for me, but one sentence makes AI write a script for me, which is not a joke.
You can write a script in one sentence, but you can't walk independently
How does Dramatron work? As long as you give a "one-sentence outline" (log line) to describe the drama conflict, Dramatron can generate titles, characters, plots, scenes and dialogues.
For example, a "one-sentence outline" could be "James discovers he is a demon and will be exorcised."
But Tristan Wolff, author of the Medium website, pointed out that this sentence lacks basic narrative elements, such as the protagonist’s goals and opponents, so the generated results are mediocre. The characters are only James and the exorcist, and the title is also dry. magic".
It would have been much better if the synopsis had been supplemented with "James discovers he is a demon, he will be exorcised, and he will have to choose between good and evil" and include an additional character, James' lover.
The story also became that James' lover was tortured after James' death, and later found that James was living in his body in the form of a demon. The title was taken as "the inner demon", which has multiple literary meanings.
After inputting the outline, you can generate titles, character design and other results multiple times until you are satisfied, or directly edit the current result, or even go back to the outline and start over, which is equivalent to co-writing scripts with AI during the interaction process.
Therefore, the positioning of Dramatron is actually a man-machine "co-writing" (co-writing) tool, and it cannot walk independently.
How was Dramatron developed? Like other generative AIs, Dramatron also has a large language model Chinchilla as a "backer". However, large models such as OpenAI's GPT-3 can also be used to deploy Dramatron.
It is worth mentioning that the coherence of content is fundamental for dramas and movies, but it is a difficult thing for large language models. Because they don't really understand the content, the generated text is actually the result of a probability calculation. The characteristic of Dramatron is that it has put a lot of effort into this aspect.
On the one hand, Dramatron generates scripts through the "layered story generation" method, using prompt chaining with structured generation capabilities.
Starting from the "one-sentence outline" you input, Dramatron first creates titles and characters, and the generated characters are used as prompts to generate plots and scenes…so progressive, and finally combine all these elements to generate dialogue.
On the other hand, Dramatron learns two classic narrative structures:
One is the pyramid structure of German playwright Gustav Freytag, which lists seven key steps in writing a story, including exposition, triggering event, rising action, climax, descending action, resolution, and ending.
The second is Hero's journey. The main line revolves around a hero who embarks on an adventurous journey. This character will win in a decisive crisis, and then be sublimated and transformed or return to the original world with the loot.
How many tricks does Dramatron have, or the experts have the final say, DeepMind invited 15 playwrights and screenwriters to experience it for 2 hours.
The screenwriters who participated in the experience reported that the output of Dramatron may be "formula", and its "hierarchical story generation" structure is not suitable for all writers.
At the same time, most of them agree that Dramatron is helpful and are willing to use it as a creative tool to build a world view, or change characters or plots to explore different stories.
Rapid Fire Theater, a Canadian improvisational theater, has staged a script co-written with Dramatron, which has received a lot of praise.
In addition to the work that still needs to be improved, the fledgling Dramatron still has many problems.
One is a copyright issue. Its output may contain raw materials during training, which requires human search and inspection; the other is an ethical issue, which may reproduce the bias and stereotypes of the corpus. Google suggests that the Perspective API tool be used to help identify "toxic "Text.
On December 10th, DeepMind opened a trial version, but it was closed soon for unknown reasons. Now the website only has papers and introductions about Dramatron .
AI took over screenwriting work, jumping repeatedly between mental retardation and intelligence
The scripts written by AI have actually been around for a long time, but they are always regarded as jokes and become a scene in the whole living area.
In October last year, Netflix teamed up with writer and comedian Keaton Patti to show AI 400,000 hours of horror movies, and then let the AI write an original script itself. 400,000 hours is about 45 years, at least AI can "watch" movies much faster than us.
The final film is about 4 minutes long and is called "Mr. There are many lines between absurd and reasonable, such as "I wish you a coupon for buying a coffin", "He is drunk but troubled by sobriety", "I have several families", etc.
Coupled with the stiff body and perfunctory style of painting at the level of "learning animation for three years", some netizens expressed such emotion: "It may be too early for AI to write horror scripts, but writing comedy is already proficient."
Similarly, an AI that watched 1,000 hours of Batman movies wrote a script that ended up being made into a sound comic, and in one episode, the Joker gives Batman a coupon for new parents, but the coupon has expired, and the Joker His personality was captured by AI.
▲ Picture from: Station B @HUSH_13
These AI scripts, either comedy or horror, are much better written than the "shit unreasonable article generator". .
Before Dramatron came out, some people asked AI to write serious movie scripts based on large language models.
After watching "Spider-Man: No Return for Heroes", YouTube blogger @Bradius wondered if AI could write such a plot, even comparable to a $200 million production.
So, he entered a sentence in GPT-3 in English: "The following is the complete leaked script of the next MCU (Marvel Cinematic Universe) Spider-Man movie."
The result surprised him. GPT-3 gave him a 21-page script with ups and downs and a tight structure. The villain Kraven is also interspersed with Marvel executives planning to reduce the budget of the next Spider-Man movie.
I don't know if AI is satirizing Hollywood's blockbuster system and assembly line production of movies. When large-scale production movies are becoming less and less innovative, it seems that there is no problem in letting AI replace some human screenwriters, and maybe there will be surprises.
Although not as "vertical" as Dramatron, the recent top-tier AI ChatGPT can also act as a "copywriting tool" to write weekly newspapers, poems and even scripts.
Knowing that the answerer, Dr. Tian Yuandong from the Robotics Department of Carnegie Mellon University, tried to feed ChatGPT with Chinese and English clips. The general result was not bad, but found that "some deep-level imagination and connection skills are still missing, and the key parts of the plot still need the author. Develop it yourself and induce the AI system to complete it.”
At present, one limitation of large language models is that they can only generate text based on the probability of certain words or word sequences appearing together based on the given training data, relying on large models, big data, and large computing power to "work miracles" , but don't really understand what they're "saying" and the answers aren't always coherent or meaningful.
But judging from the results alone, it is not a small probability event for them to extract narratives that humans think make sense and even shine from the text of natural language.
Some time ago, the movie commentary of "Small Handsome Little American Style" also aroused discussion. This kind of movie commentary is a short and fast "electronic mustard", and the AI dubbing tells how the protagonist unexpectedly encountered an accident.
In order to make the explanation wonderful, they label the characters, choose the most curious or suspenseful plots to add oil and vinegar, often do not follow the original intention of the movie, and are neither movies nor explanations.
This part of the work will be completely done by AI in the future, and it should be easy to realize the full automation of electronic mustard.
Is AI being used to inspire ideas, or make them harder to come by?
Writing articles, making videos, playing music, and being a painter, AI has made great progress this year, making it beyond the reach of any human being.
We often say that AI can be used as a creative tool, but in front of AI, are humans giving up their roles, or enjoying new roles? We increasingly need to cooperate with AI. Is it stimulating creativity or killing it? What is the position of creativity itself?
There are some examples of human and AI writing right now, and it seems the human is the more irreplaceable one.
For example, in addition to Dramatron, Google also plans to develop the AI writing tool Wordcraft based on the dialogue neural language model LaMDA, which is still in the experimental stage.
Wordcraft is a text editor dedicated to fictional works. You can enter a beginning and let it continue, or use it to process sentences and enrich corpora.
To test Wordcraft, Google invited 13 professional writers to experience it. The consensus among writers is that Wordcraft isn't going to replace writers anytime soon, it's not good at unique storytelling styles, corny tropes, mediocre phrasing, while eschewing sleazy characters.
Of course, they also came to a consensus, as with all other AI tools – it is possible to stimulate creativity with Wordcraft.
AI that is not perfect will continue to study hard and improve every day. At the same time, Douban has groups such as the "Word Aphasia Mutual Aid Alliance". For humans, the logic of organizing words and clearly expressing their emotions in words has become more and more difficult things.
Similarly, in order to improve the quality of AI generation, a new type of job called Prompt engineers (prompt language engineers) emerged as the times require. On platforms such as PromptBase, you need a specific art style, and you can ask them for help.
But this work can also be done by AI, such as using ChatGPT to form a text, and then input the text to StableDiffusion. The generated paintings are generally much better than the direct input description.
So, when we have a "sentence outline" brain hole, but can't think of more details to lay it out completely, and have to turn to AI, is it a helpless choice, or make the best use of it?
Thinking optimistically, AI should enhance rather than replace human labor, but if the time dimension is stretched longer, this may no longer be the case.
Sam Altman, CEO of OpenAI, believes that "as a creative tool" will become an important application of AI for a long time, but when we look forward to a hundred years, AI may eventually take full responsibility for complete creative work.
If you asked people 10 years ago how AI was going to make an impact, most people were pretty confident you would hear it first in factories etc, then it would be in low-skilled white collar jobs, then programmers High-skilled, high-IQ jobs. In the end, and maybe never, it will replace creative work. However, it is going in another direction.
This is actually a reminder that we may not really know what skills are easy for humans or AI, what skills are simple, what jobs are brain-struggling, and what jobs don’t take brain cells; we are also gradually realizing I believe that in the field of AI, only those who are proficient in using AI early can stay in the future.
#Welcome to pay attention to Aifaner's official WeChat public account: Aifaner (WeChat ID: ifanr), more exciting content will be presented to you as soon as possible.