I am the only one who practices magic: I practice magic in the city

Chapter 258 300 Internal Test Qualifications (Second Update)
01:00 California time, 04:00 New York time, 17:00 Singapore time. Seeing that the cuddly Black Widow had fallen asleep, Fang Yu started the teleportation and returned to Haixi.

And at the same time.

San Francisco, Bay Area, Mountain View, Google Brain Research Center.

Jeff Dean nervously urged Li Yueguo, who also looked solemn: "Li, did you get it? Quick, refresh it."

"It's already refreshing. The speed is too slow. Damn, what kind of crappy bandwidth is this at Youzi Technology?" Li Yueguo kept clicking his fingers on F5.

After a while, the page finally refreshed with a big congratulations on it.

"I got it!" Li Yueguo jumped up from his chair excitedly, "Twenty-nine thousand applicants, three hundred internal test qualifications, and I actually got it!"

Jeff Dean couldn't hide his excitement.

If it was a year ago, no, it should be said that two months ago, he would not believe that as the first person in charge of Google Brain, he would be so excited about grabbing an internal testing qualification from another company.

This is the internal testing qualification for the Orange large model!
The Orange model claims to have initially achieved AGI (artificial general intelligence)!
Since the Orange model defeated AlphaGo two months ago, there has been no movement, and even the updates on arxiv and github are no longer as frequent as before.

The only progress is the madly updated yuzu architecture.

Currently, this architecture has been updated to version 0.41, with abundant examples, and has become the most popular training architecture.

But the big orange model that had been so popular before seemed to have disappeared.

There are already many people who are spreading conspiracy theories saying that the orange model is a hoax.

But who would have thought that the Orange Big Model, which had been silent for two months, announced three days ago that it would conduct a small-scale internal test and open 300 internal test permissions to users and experts in related fields around the world.

To apply for internal testing permissions, you need to fill out the corresponding industry certification, and only those who meet the requirements can apply.

When each person applies, a random question will be given, and only those who can answer it can continue to apply.

Most of these questions focus on autoregressive models, normalization processing, matrix calculations, and beam searches, and each person encounters different problems.

In this way, in just three days, nearly 30,000 applicants still completed the application test.

There is no doubt that almost all the respectable researchers in the field of deep learning have applied!

"Yuzu Technology is very confident." Li Yueguo said with emotion, "In the research background, I filled in the ancient dog brain, and I also obtained it. This shows that they are not afraid of us knowing their progress at all."

"This JUZI.YOUZI.COM must be the login page. Let's take a look."
-
"Youzi Technology is very confident that it actually opened up the internal testing qualification to us." In the Pioneer Building in San Francisco, a middle-aged man with a muscular build and the inner corners of his eyes seemed to be glued together also sighed.

Almost everyone in the world who pays attention to electric vehicles will be familiar with this face.

Elon Musk, the founder of Tesla and SpaceX, is a man who claims to change the world and migrate to Mars, and is known as the "real-life Iron Man."

"Maybe they just don't see us as competitors. Elon, let me interview you. What do you think about this?" A man in a blue casual shirt was half lying on the sofa, holding a magazine in his hand and handing it to Musk's mouth.

"OpenAI Gym, which was originally scheduled to be released this month, has lost its significance after the launch of the yuzu architecture. From this point of view, they are indeed ahead of openAI."

“Sam, I’m more concerned about whether they can guarantee safety if they do achieve AGI.”

Although Musk looked normal, there was a hint of worry between his brows.

The man in front of him named Sam was his partner in openAI, Sam Altman, partner and president of Y Combinator, who also promised to invest but did not pay a penny.

Musk gets angry every time he sees Ultraman Sam.

Since they decided to jointly establish this non-profit AI research institute at the end of last year, he has donated 50 million Malaysian dollars in cash.

Sam Altman and Peter Thiel promised to contribute $1 billion, but they haven’t seen a dime of it yet. Even the office in the Pioneer Building that OpenAI is using now is half of what their company Neuralink, which is still keeping secret, gave to OpenAI.

Ultraman Sam shrugged, "Well, let's see what this big orange model is."

"Ilya, are you logged in?"

Ultraman Sam raised his voice to a slovenly young man in a gray T-shirt.

Sitting in front of the computer, the young man named Ilya didn't say anything, just nodded.

Elon Musk, Sam Altman, and a group of people gathered around.

The login interface of the Orange Big Model is very simple, it just requires a username and password.

Elijah entered the username and password he had previously registered on the Yuzu Technology homepage and logged into the operation interface smoothly.

After entering the operation page, a PDF version of the internal test instructions for download will automatically pop up on the page.

After downloading, Ilya clicked on the internal test instructions.

With a total length of more than 40 pages, I believe it explains the current functions of the Orange model, the content that needs to be tested, the testing tools and methods, and the feedback mechanism.

"Natural language processing? Is that true? Isn't that a bit exaggerated? It would be more believable if they said they solved the problem of natural language generation." After reading the beginning, Sam Ultraman couldn't help but frowned.

Natural language generation (NLG) is a subfield of natural language processing (NLP), which refers to the ability of computers to generate text that conforms to human language habits.

The latter means that in addition to generating texts that conform to human language habits, it also has many sub-fields such as text classification, sentiment analysis, machine translation, information extraction, and speech recognition.

It can be said that natural language processing means that if it is just a chat, it will be difficult for humans to tell whether the other party is a human or AI.

"Question-and-answer system, natural language understanding, information retrieval, answer generation, context understanding... How is this possible?" Sam Ultraman was more and more surprised the more he looked.

Just looking at the documentation, this is already a fairly complete model system.

"It's not difficult to generate pictures from text. We can use adversarial networks, but the examples in the document are a little surprising." Ilya pointed to a picture of a child holding a puppy in the document and spoke for the first time.

He has a distinct accent and speaks with a slight stutter.

"Especially here, when Yuzu Technology generated this picture, it described it as follows: The child and the puppy are looking into the distance, perhaps seeing their father in the distance. The child is filled with joy, while the puppy wants to break free from the arms of its little master and run towards him."

Ilya's speech gradually became fluent, except for some stuttering at the beginning of each sentence, it was much better at other times.

"Look at the dog's expression in this picture. Its mouth is open and its tongue is hanging out. Although it is struggling, its paws are deliberately avoiding its little master's face."

"This shows that, at least when making this picture, the orange model understood the relationship between the little master and the master, and understood the relationship between the dog, the master and the little master, and was able to guess the dog's emotions and how the dog would express similar emotions."

"If it is true as they say, this image can be generated after only one additional explanation and debugging, then it means that the Orange model does have the ability of NLP."

Elijah picked up a tissue and wiped the oil off his face from staying up all night.

"According to the documentation, the Orange model currently has 190 billion parameters, which is indeed an astonishing number, but it is still a long way from the trillion-level synaptic connections of humans."

“I still have to reserve my opinion on whether parameters of this magnitude can achieve the functions they claim.”

“They didn’t open source, or even disclose the parameter structure of the large model.” Sam Altman shrugged, “I don’t think they can really get to that point.”

"We'll know after testing it." Musk's eyes flickered. "Ilya, test it now. I'll design a test plan."

(End of this chapter)