There is no so-called AI search engine, and there is no explosive GPT-5, but OpenAI still makes everyone feel what it means to be "Only OpenAI Can Do".
In the early hours of this morning, OpenAI's spring conference finally came.
They didn't say a few more nonsense, the whole press conference lasted about 20 minutes, and the most important of them was the release of the new GPT-4o model.
Regarding the reason for the name of GPT-4o, OpenAI officially explained it like this: The full name of "o" is "omni", which means "omni".
The reason why it is said to be all-powerful is because it can connect text, visuals, and audio. In other words, you're talking to GPT-4o as if you're a real person.
On top of that, GPT-4o is free and open to users, and you can use it without a Plus membership, but the Plus membership has to be used more times.
Anyway, the colleagues who squatted on the live broadcast with Shichao watched it, and they all shouted that they wanted to...
Even Ultraman has not shied away from tweeting about linking GPT-4o to the sci-fi movie "Her". (This movie tells the story of a person and an AI voice assistant who fall in love)
Netizens in the comment area also said that "witnessing history" and "the movie is finally going to come to life", and even used Ultraman's photos to make a "him".
Jokes are jokes, but Shichao feels that OpenAI's live demonstration effect of GPT-4o, especially the ability to interact with real-time voice and video, has indeed thrown off several taillights of other large models in service.
However, just by saying that, everyone may not have a deep feeling about GPT-4o, so let's go directly to the demo.
For example, the presenter asks GPT-4o to count from one to ten, and thinks it is too slow to speak halfway, so let it say faster, GPT-4o can react immediately, and it can get it instantly at any time when it interjects in the future.
In terms of the strength of continuous dialogue one after another, to be honest, Shichao has not experienced it on AI.
Give GPT-4o a microphone and camera permission, and it can become an interview instructor in minutes, such as the guy in the video, who says that he is going to OpenAI for an interview, and asks GPT-4o to give some dress advice.
It quickly "looked" at the old man, and politely and humorously gave the pertinent advice of "grab your hair". But the elder brother planned to go to a disobedient persona, so he took out a hat and asked if he could wear it to the interview.
GPT-4o reacted quickly, saw through the old brother's knowingly prank, and followed his words, it is not impossible, at least it can make you stand out from the crowd.
If I hadn't shown Shichao the demo screen, I might have thought that this was a joke between two friends.
Of course, with GPT-4o's knowledge and multimodal capabilities, it also has quite serious application scenarios.
If you want to know what is in front of you, GPT-4o can also depict it vividly, and even guide the visually impaired to take a taxi step by step.
GPT-4o can also become a math teacher, open the screen to it, it can guide you step by step to do the problem, and you can see what you write and draw on the problem GPT-4o can also see clearly.
One thing to say, this is no different from the tutor that Shichao's parents hired when he was in middle school...
After watching the official case video, I believe that everyone, like Shichao, has been shocked.
In order to do this, OpenAI has put a lot of hard work into the underlying model.
In the past, we used ChatGPT to have voice conversations, all of which came round by round. It has to convert what we say into text and input it to GPT-4, and then convert it into speech output after GPT-4 generates text answers.
In other words, in the past, ChatGPT had to use three models, one to convert audio to text, one GPT-4 to text, and one to text.
After a whole set of processes, patience is almost worn out, and unless we pause manually, there is no chance to interject.
But GPT-4o is different, it is a native multimodal model, whether it is text, audio, video, etc., all inputs and outputs are processed in the same neural network.
This kind of mother-beating has its own multimodal characteristics, so that GPT-4o can understand good words, express his emotions, and talk to it as silky as if he had eaten Dove.
However, the above videos are still "PPT" after all, and before they have been measured, Shichao does not dare to easily conclude whether this thing is really so strong. After all, some time ago, the short film "Balloon Man" generated by Sora was rumored to add a large number of post-human films.
In practice, OpenAI is still the same dog as always, saying that GPT-4o can be used for free, but it also has to be divided into first-come, first-served, and the functions are slowly opening up.
Now only the Plus account can use GPT-4o, and the heaviest real-time video and voice functions are not visible to Mao.
But there is better than nothing, and Shichao also happens to have a Plus account at hand, and the official also said that GPT-4o will generate twice as fast as GPT-4.
So in this wave, let's try the strength of GPT-4o for poor friends in advance.
Let it tell us a fairy tale first, and then draw a storyboard of the whole story by the way. GPT-4o didn't hesitate for a few seconds, and Ku Eat Ku Eat began to write.
After finishing writing, there was no pause in the middle, and I started to draw the storyboard.
And the same problem is thrown to GPT-4, it has to grind for a long time in the first place...
After clarifying the requirements, I was willing to continue to write, and I didn't continue to draw storyboards after writing, and I had to wait for Shichao to emphasize it in person.
However, although GPT-4 has a lot of ink than GPT-4o, judging from the content, I still think GPT-4 will be a little richer.
Then borrow the photos of our old employees' hot pots, and let GPT-4 and GPT-4o generate a pixel-style picture respectively.
GPT-4o wins in terms of generation speed, but if you want to pick out the details, it's half a pound.
GPT-4 recognized that the hot pot was holding a bag in its mouth, GPT-4o recognized one more slipper, if you have to compare, in fact, GPT-4's golden retriever is more like a hot pot...
（ 左边 GPT-4 ，右边 GPT-4o ）
Then, I asked GPT-4o to replace the LOGO of our bad review with a 3D picture, but it turned out the words, and I searched through the entire Xinhua dictionary, and I guess I couldn't find it.
But not knowing Chinese characters is also an old problem of ChatGPT over the years.,Change to English and then ask for it.,Don't you say.,The overall effect of this time is okay.,It's just that it can't stand up to scrutiny.,A lot of English spelling is wrong.。。
The biggest impression I get to know about GPT-4o is that it's fast, very fast.
But to be honest, apart from being fast, Shichao really doesn't think there's anything special about it, not so much that GPT-4o is an evolved version of GPT-4's IQ, but rather that GPT-4's interaction ability has been singled out and strengthened.
As Altman mentioned in his blog, "reaching the level of human response time and expression" is a huge change in AI. The ultimate form of the large model should also be to return the communication between people and AI to the most primitive and simple form.
In the future, we will face a computer and a mobile phone, and direct communication is the most important way to interact.
However, this time, GPT-4o's most powerful killer feature: the video communication function, has not yet been released, and it is not too late for us to experience it when the time comes.
In addition, netizens didn't sit still, and they were also studying GPT-4o in various ways, but as a result, they picked up some extra-fanned materials...
I don't know where GPT-4o's Chinese corpus came from, and there are a lot of prohibited words on phishing websites, such as "Japanese x-film free video" and the like.
Foreigners are speechless.
There may be a lot of fun like this in the future. But to get back to the point, the advent of GPT-4o does give AI a new copy.
According to the previous ChatGPT came out, other friends and businessmen were approaching step by step. After GPT-4o, similar "Claude-o" and "Gemini-o" should appear soon.
It just so happens that tomorrow is Google's I/O, so let's look forward to it.