How did a human disguise as Sora become a traffic password?

Will Smith once again contributed a meme, last time it was Oscar's slap, this time it was spaghetti mukbang.

In March 2023, a Reddit netizen used AI to create a funny video of Will Smith eating noodles. His face was hideous and his facial features were deformed, which made people lose their appetite.

In February 2024, two days after OpenAI released Sora, Will Smith posted a comparison video of eating noodles on Instagram. He became more handsome and his eating movements became more decent.

Do you think the second half of the video was generated by Sora? No, the noodles were eaten by Will Smith himself, and he did the whole job.

The front foot was shocked by Sora's official demonstration, and the back foot saw this dynamic, and people fell into doubts about who I am, where I come from, and where I am going.

Although Sora is not yet open to the public, the Internet has become more and more ridiculous: AI videos are getting closer and closer to reality, and more and more real-life videos are fake AI.

I don’t know if it can kill Hollywood. Sora burned my brain stem first.

We have more or less seen the videos of Sora's stable performance. At the same time, OpenAI has also released the rollover videos generously. Some of them are more popular on social media, especially the following prompt: "Archaeologists found a handful of cars in the desert. Regular plastic chair, carefully dug up and dusted" video.

The chair seems to be a breathing, thinking creature from another world that is not bound by gravity or human power. It appears and transforms out of thin air, and even secretly takes a human being away. This may not be the fault of the AI. After all, the prompt word does not indicate whether they are on earth.

If the video were distributed to social media, it might be tagged with: #streamofconsciousness, #surrealism, #preciousvideoofearlyhumankindtamingthewildchair…

OpenAI seriously explained why the bug occurred: Sora failed to model the chair as a rigid object, that is, an object that does not change shape and volume under the action of external forces, so the physical interaction was inaccurate.

Sora's video of breaking a cup seems to be an art that can only be seen after eating poisonous mushrooms. The cup actually moves automatically without any wind. The liquid flashes out of the glass like jelly before breaking. Every step is unexpected.

In addition, there are athletes using treadmills backwards, birthday candles that cannot be blown out, wolves that appear and disappear out of thin air… AI has inadvertently done things that neither the human brain nor special effects can imagine.

Since the official announcement of Sora, AI tycoons have been debating whether it can understand the physical world. OpenAI’s technical report did not make a clear position, but only mentioned that Sora can simulate some aspects of people, animals and the environment in the real world, but it may be difficult to Accurately simulate the physical principles of complex scenes, such as confusing left and right, and leaving no bite marks after biting a cookie.

While AI has not yet evolved to a complete state, you might as well use your imagination. AI videos that are not yet stable can be used as a material library for supernatural, suspense, horror, and fantasy themes, providing new inspiration for creation. The more you try to lift Newton's coffin, the more artistic your performance becomes.

Just like "digging chairs" and "breaking cups", they can already be used in MVs, advertisements, movies, and games. It still looks like a big production. Even though the effect is weird, it is surprisingly smooth. It beats Station B. "Learned Five Years Animation" series.

Even though the content was so magical, some netizens still hesitated for a second: "If the bug wasn't revealed, I would doubt it was filmed." "Is this a magic trick?"

This is because even Sora's rollover video can still beat "predecessors" such as Pika and Runway in terms of duration, image quality, and stability.

While other tools generate videos of no more than 3 to 4 seconds and try to keep a single shot as stable as possible, Sora raises the upper limit to 1 minute to achieve seamless switching of multiple shots and relatively accurately maintain the consistency of the main subject and visual style of the picture, as if it knows how to use it. Lens language and narrative rhythm tell stories.

At first glance, it seems that there are real cameras following the scene. People, objects and scenes are all moving in three-dimensional space, and the same character can have multiple perspectives.

▲ Sora can achieve multiple perspectives.

Some people joke that this is a leap from "animated pictures" to "videos" generated by AI.

The video of Sora overturning the car can certainly be counted among them. It's like writing a bunch of code in a mess, but it happens to run.

Zhou Hongyi, chairman of 360, talked about an interesting point on Weibo: Sora works like human dreaming. We will dream of strange things, but it is based on what we see with our eyes and accumulated experience in life. , there is no need to 3D model and render frame by frame like in the film and television industry.

From a certain perspective, generative AI is indeed like a dream machine, getting closer and closer to the way humans think. It uses various prompt words to create reasonable or unreasonable answers with probability. Regardless of whether it is right or wrong, it will definitely give you an answer.

In fact, since the beginning of ChatGPT, people have wanted to use AI to bring dreams to reality. There are many posts on Xiaohongshu that visualize dreams. Although they cannot truly restore the effect in the mind, they also bring some phantoms of the spiritual world into reality.

In the future, Sora may make dreams or human creativity more vivid. By inputting text, pictures or videos, you can add light and shadow changes, adjust the picture angle and even add sound effects with "one click"…

I don’t know if Hollywood’s elite will lose their jobs. Netizens are already eager to try it. In addition to their instinct to engage in pornography, they also have a place to put their imagination.

Human beings who laugh at their own failure use imitating Sora as a traffic password

In addition to the car rollover video, Sora's other videos are amazing at first glance, but they cannot withstand the scrutiny of a magnifying glass. The garbled characters on the shop sign, the cat’s fifth leg, the model’s wrong steps…

Bugs may be the last wall between AI and reality, but humans themselves want to knock it down.

Although Sora is not yet open to the public, those who sell courses and chips are making money, and those who have fun playing tricks have also found ways to fish in troubled waters and invented a new way to attract traffic.

When they posted the video on social media, they added "generated by Sora" and decent prompts to the copy, pretending to be made by AI, just like Will Smith imitating AI imitating himself.

Short video bloggers from various circuits are participating. Those who show off cute pets, play games, Amway idol singles, advertise products…are they really not providing employment inspiration to Sora?

Among them, the most difficult to distinguish between genuine and fake are undoubtedly Bollywood movies. The plots and special effects of Indian police movies are particularly amazing. They are too advanced for humans and make physics non-existent. Even AI has to learn from them. .

If it weren't for the enthusiastic people who added comments and attached a YouTube link uploaded 11 years ago to prove the inhuman pioneering nature of Indian films, no one would have dared to vouch for it.

Even the official account of Netflix India joined in the fun. It intercepted a clip from the Indian anti-British movie "RRR" and pretended to be generated by Sora. According to the usual exaggeration of Indian blockbusters, the first few seconds may indeed deceive the audience.

Although these fake AI videos are created by humans, and the AI ​​mixed in them seems ordinary, AI may not be unable to imitate them.

Confusing the public opinion and bringing it upon yourself, the harm is still caused by human beings, and a chain of suspicion has emerged.

Netizens began to worry that when AI-generated and real-life photos are really unclear, they may have "conclusive evidence" even though they have not done anything.

This is the plot of the suspenseful British drama "Truth Capture": the intelligence agency forged surveillance videos in order to convict criminals.

They believe that forging surveillance videos is just "re-enacting" illegal evidence such as wiretapping materials into legal evidence. Things did happen and the criminals did commit crimes.

But no one knows whether people who use technology will cross the line and go from grafting and grafting to creating something out of nothing.

Identifying AI's response methods may also be used from another angle as a trick to escape from the golden cicada's shell.

Because AI is not good at drawing hands, there was a meme in the past where criminals wore fake fingers to commit crimes. It looked like they had six fingers. This could slander the surveillance video as being generated by AI and could not be used as evidence in court.

AI is close to reality, while humans are pretending to be AI. AI can also be used to falsify evidence presented in court and invalidate evidence. A magical world where magic defeats magic may really be coming. AI can't actively cause chaos yet, and it's humans themselves who can break trust.

AI may not be able to simulate the world yet, but it has already affected the way we see the world.

Regarding the issue of AI making it more difficult to distinguish between real and fake, when we talk about Deepfake face-changing, we have already discussed it too much. How could Sora be any different?

Recently, OpenAI made an unexpected move – opening a TikTok account and uploading Sora's works.

OpenAI not only carefully labels each video generated by AI to prevent people from confusing reality, but also includes prompt words.

"A realistic video shows a butterfly that can swim underwater across a beautiful coral reef" "A macro shot of a leaf shows tiny trains moving through its veins" "A camera low to the ground closely tracks ants in the jungle"… …

These videos are mixed in with the real short video information flow and are accompanied by background music. If you are not careful and slide your finger too fast, it may be considered real.

Netizens half-jokingly and half-seriously urged in the OpenAI comment area: "You'd better make something like a watermark, otherwise humanity is doomed."

Sora has not yet been launched. In fact, everyone is guessing that video content practitioners may be the calmest group of people because they are more professional and understand business better. Sora’s Demo has a lot of problems with the consistency and accuracy of the content. If it is not enough and fails to meet Party A’s precise requirements, it will be difficult for the technology to be introduced into the industrial process.

But I am not very optimistic. AI has already affected not only truth and falsehood, but the way we see the world. It does not directly replace a certain editing software, a certain script, or a certain director.

AI practitioner @Kwebbelkop speculates that OpenAI currently only releases AI short videos on TikTok, possibly to collect relevant user data such as the number of views, fine-tune the model from human feedback, and even create an AI version of TikTok in the future.

The short video algorithm is already very good at guessing what you like. What will happen if generative AI is added to make the video content more customized?

In the future, our movie-watching experience may also undergo qualitative changes. Instead of relying on cinemas and streaming media, large language models and video generation models can be used to determine story content and starring lineups.

However, OpenAI’s expectations for AI products are definitely not just to chat with you and make videos. The deeper ambition is to let AI learn human natural language and the physical laws of the world.

Even if it is limited to the field of video, if Sora can make rough concept films, it is already very impressive. The way of making videos, the style of content, and even our preferences for content may all be changed because of AI.

Putting aside AI, the mobile phone owned by almost everyone has actually affected the way videos are shot and produced, allowing everyone to become the director of their own lives.

Beauty videos where you put on makeup while talking to the camera, Vlogs that show fragments of your personal life, etc. are the products that many bloggers can shoot in their bedrooms with their mobile phones under this influence. The explosion of short videos has also made us more and more accustomed to swiping up and down on our mobile phones, leaving us less patient and more distracted.

Although compared to Sora, video generation tools such as Runway and Pika have limited capabilities, some people have combined them with image generation tools such as Midjourney and used them to make movie trailers.

Because the stability is poor, the video style has also been chosen. It is mainly fast editing, with narration, focusing on rhythm and visual impact, but it lacks character dialogue and more complex scenes.

▲ AI science fiction short film "Borrowing Time".

In contrast, Sora can support more complex scenes, character actions, and interactions between characters and the surrounding world. Someone used Sora's sample film, then used the AI ​​voice tool ElevenLabs to dub it, and used iMovie to edit it, to create a smoother "movie trailer" that seems to be "one shot to the end".

The two types of videos Sora displays to the outside world can be roughly divided into two types. One is creative imagination, such as astronauts standing on a cold planet, two pirate ships dueling in coffee, and cartoon characters dancing disco. The other is close to reality and gold mining. California during the heat, reflections in train windows, Nigeria’s outdoors in 2056…

The reflection on the train window is very similar to what a Vlog would shoot. The outdoor scene in Nigeria, with the camera panning from the open-air market to the urban landscape, is also very similar to the empty shot of a news video. Someone has already planned to combine digital people with this video.

Photography may even seem unnecessary at times. Sora can directly create a bird's-eye view of a tourist attraction through prompt words, which is not far different from the footage of a human flying a drone. Our eyes and brains know what Santorini looks like, and AI also “knows” it, so it can be left to AI to do the job.

There was a very interesting metaphor before, this world is a huge "Earth Online", a massively multiplayer online role-playing game developed by Solar System, with the best 3D naked eyes and VR system.

The concept of "world simulator" proposed by OpenAI treats everything as information input to a certain extent, absorbing text, images, and videos, and then outputs information, turning text into paintings and pictures into videos, as if " "The Lord of the Earth Online", but we may not understand the principle. We may just give the request and get the result.

Such a future should be far away. At least for now, watching too many Sora videos will make people feel bored. Women walking on the streets of Tokyo have no emotion in their eyes and cannot feel the human atmosphere. But when we watch short videos Sometimes this is also the case, the world needs subtle movies and accepts huge redundancy, just like the resistance of the protagonist of "The Matrix" is also designed.

A future in which AI becomes more and more involved is coming. Next time we see a suspected bug, we may not directly judge it to be fake. We may think about whether AI has any purpose here, just like doing reading comprehension. We may not like it, but we have to accept it.

It is as sharp as autumn frost and can ward off evil disasters. Work email: [email protected]

# Welcome to follow the official WeChat public account of Aifaner: Aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo