Hengyu is from the Au Fei Temple
Quantum Position | 公众号 QbitAI
Okay, it's worthy of OpenAI's latest flagship, open various social software, and GPT-4o's hands-on tests are pushed to my homepage.
Please! See!
This is to use GPT-4o, it takes less than 30 seconds to generate complete charts and statistical analysis of the content in a spreadsheet through a single prompt.
In the past, it didn't take us a long time to do this stuff in Excel?
The picture below is an STL file of a 3D model of a four-legged table created by netizens with GPT-4o in less than 20s.
Cow wow cow wow!
After all, GPT-4o's capabilities span listening, speaking, and watching, and the main thing is that it's free!
As netizens have concluded, now, every user can generate something very amazing through AI and simple Prompts.
However, there is still a lot of exploration on how to generate something with a complex structure.
Let's take a look at how netizens are selflessly playing GPT-4o——
狂玩GPT-4o
24 hours before the annual Google I/O developer conference, OpenAI raided the release of GPT-4o.
"o" is an abbreviation for Omni, which means "all-round".
The reason for this name is that GPT-4o accepts any combination of text, audio, and images as input, and generates text, audio, and image outputs.
In the official OpenAI demo on May 14, it was very silky to use, and even the speed of responding to audio input caught up with humans.
With the attitude of "I don't believe unless I try", netizens have gone crazy.
Einstein puzzles
First of all, this so-called "Einstein puzzle" is very similar to the Olympiad problems we did when we were children, and it is used to test the logical ability of large models.
The background of the question is as follows:
On one street, there were five houses, sprayed with five colors. Each house is inhabited by people of different nationalities. Everyone drinks different drinks, smokes different brands of cigarettes, and has different pets.
Prompt:
(1) The British live in red houses.
(2) Swedes have dogs.
(3) Danes drink tea.
(4) The green house is next door to the left of the white house.
(5) The owner of the green house drinks coffee.
（6）抽Pall Mall香烟的人养鸟。
(7)黄色房子主人抽Dun Hill香烟。
(8) People who live in the middle house drink milk.
(9) Norwegians live in the first room.
(10) The person who smokes Blends cigarettes lives next door to the person who owns a cat.
(11) The horse breeder lives next door to the man who smokes Dun Hill cigarettes.
(12) Smoke Blue Master and drink beer.
(13) Germans smoke Prince cigarettes.
(14) Norwegians live next door to the blue house.
(15) The person who smokes Blends cigarettes has a neighbor who drinks water.
The question arises, who raises fish? Who lives in the Blue House?
A few days ago, when netizens tested i-am-gpt2-bot (the mysterious GPT-2 that kills in the big model arena) in lmsys, they still couldn't solve the Einstein puzzle - and there was no other AI that could solve this problem.
But when I tried it, GPT-4o answered right at the speed of light.
You can do it yourself (manual dog head).
Automatic stock picker
I just saw OpenAI say that GPT can't be used for stock selection, so it has no reference significance.
On the back foot, some netizens posted the GPT-4o automatic stock picker on Twitter, with the text: It's terrifyingly strong!
Specifically, he used GPT-4o to automatically rewrite more than 200 lines of stock selection indicators into an automatic stock picker, output charts, and data archiving.
△AI-generated content does not represent any investment advice
And it only takes 1 round of interaction to complete a more satisfactory effect, and the efficiency beats GPT-4 (oh? I beat myself before I did)
According to him, using GPT-4 to do this requires repeated revisions and is very inefficient to process more than 100 lines of code.
In this regard, netizens' comments are very incisive:
If you can predict it 100%, it is really perfect! But if the prediction is not right, it is better not to predict......
Paper prototype transcription initial HTML
Some netizens also tried to use GPT-4o to transcribe the prototype written on paper into the initial HTML in the computer.
Here's what he wrote in black and white:
然后把这张图喂给了GPT-4o。
Then GPT-4o says:
And then you get:
The netizen himself was very satisfied with the result of spitting, and he wrote excitedly on Twitter:
It's like we're having a conversation that transcends the world, and it's really Soooooo Cool~
He is not the only one, but some netizens said on Hacker News that they can also use various versions of GPT to dynamically convert the original dynamic data into beautiful HTML layouts.
This saves a lot of development time by making low-traffic pages, such as change/audit logs, and also keeps the HTML updated when the data structure changes.
However, the attempt doesn't work back, as GPT-4-Turbo sometimes ignores context and instructions almost entirely.
OCR
However, some netizens said that they were also a little angry with GPT-4o's OCR ability
Here's the thing, he threw GPT-4o a copy of this picture.
How to say it, it is indeed dense, and the company logo has images and words, which is a bit difficult for human eyes to see.
The results of GPT-4o shocked the tester himself, saying, "It keeps spitting out the contents of the graph that is difficult for even humans to recognize." ”
Onlookers have left messages on the post, which are nothing more than "In the future, it seems that there will be a big difference in the work between those who use it and those who don't" and "If you can't keep up with advanced technology, you will fall behind" and so on.
Punch Google, but get kicked by Musk?
OpenAI surprised the world again, and Google really couldn't sit still.
At the press conference in the early hours of this morning, Google brought Project Astra, its latest large-scale product.
Like GPT-4o, Project Astra can write, listen, read, and speak, and communicate with humans with little to no delay.
However, Nvidia scientist Jim Fan was the first to come out and comment:
First of all, Google looks like it's a multimodal input, but not a multimodal output.
Google's image generation model, Imagen-3, and the music generation model are still separate components and are not integrated into them.
He mentioned his view that the integration of all modalities is an inevitable future trend, and of course, there are some details that he believes are indispensable, as follows.
- Enable task options such as "Use a more mechanical voice", "Speak 2x faster", "Iteratively edit this image", and "Generate consistent comics";
- No loss of cross-modal information, such as emotions and background sounds.
- Breaking new contextual capabilities, you can teach models to combine different senses in novel ways with a handful of examples.
In contrast, GPT-4o isn't perfect, but it's largely correct.
And what about Google?
Jim Fan, a veteran surfer, said that one of the things Google did right was that "they finally started working on integrating AI into the search box."
Some netizens really got started with Google's freshly released Project Astra and posted a horizontal evaluation video:
We listened to the content, and in general, he personally felt that the demo display link at the Google conference was not very good, and he and three other people went to the booth to try out Project Astra, and they could only play for about 2 minutes.
The feeling after playing is that it is capitalized "Google is fighting an unprepared battle".
The beta player ahead of him asked Project Astra to tell a story about something, and Astra swore yes, and then there was no more......
But let Astra recognize the drawn sailboat and smiling face, and it can still do the job.
Comparatively speaking, he thinks GPT-4o is smoother, but because he hasn't gotten his hands on GPT-4o himself, he won't make too many comments.
There is also a drama in everyone's trial carnival.
That's Grok of Musk's large-scale model company xAI, who correctly answered the question of Ilya's departure from the company.
And OpenAI's own large model failed to provide the correct response.
Netizens deliberately solemnly announced:
突发事件！xAI的Grok 1.0击败OpenAI新推出的GPT-4o。
Of course, this has to be thanks to the fact that xAI has real-time data/posts/news on Twitter (X) behind it, and nothing is faster, richer, and more realistic.
Interestingly, there was a fantastic discussion on Hacker News.
Someone showed a link, jumping to a limerick dataset Needle in a Needlestack released in 2021 (it can only be said that the world is so big), saying that he firmly believes that when GPT-4o was trained, he used the data in it,
Here's how it started -
Needle in a Needlestack, a measure of how much attention a large model pays attention to the information in the context window, contains thousands of limerick prompts that ask questions about a limerick in a specific location.
To put it simply, it's a somewhat interesting version of looking for a needle in a haystack.
But so far, no one has a large model that has performed impressively in this test.
However, GPT-4o has made a breakthrough!
It's a near-perfect performance.
So netizens began to discuss passionately what OpenAI had done to make GPT-4o's performance jump significantly from GPT-4-Turbo.
The main point is as the poster said, it is definitely OpenAI using Needle in a Needlestack to train itself, otherwise the team behind the dataset came out to explain how to check and ensure that the dataset is not used by any large model for training?
Of course, there are also voices that have already been heard before (we heard similar statements at the Dark Side of the Moon conference in March and interviews with Lin Dahua, a leading scientist at the Shanghai Artificial Intelligence Laboratory):
The needle-in-a-haystack test has very limited knowledge of the model's actual long-context capabilities.
It is widely used because early models perform poorly in this area and are easy to test.
In fact, most of the latest models are now doing a pretty good job at this task.
However, this time there is a little more information increment, and many people believe that the ability of large models to perform any complex operations will be greatly reduced when executing long contexts of more than 32k tokens.
In the end, is OpenAI really a human being?
Before the Google I/O developer conference, he put his face on the output of GPT's new features, and when the Google conference ended, he immediately brought heavy news, and Ilya, the chief scientist of OpenAI, who had been silent for half a year, really announced his resignation as everyone speculated.
Good news:
Ilya is still alive.
Bad news:
Google, you can't afford any traffic......
— END —
量子位 QbitAI 头条号签约
Follow us and be the first to know about cutting-edge technology trends