Chapter 277: Bright Future (Part 2)
Although she was full of confidence, Madeline was still overjoyed to be able to obtain the source code of the Orange model so smoothly.
Qianxun's security department has a large number of 0day vulnerabilities (vulnerabilities discovered by hackers but unknown to developers and users). The Trojans they created exploited several 0day vulnerabilities in the Linux system to remotely execute code and obtain permissions.
Qianxun didn't have to wait too long. The encrypted package was decrypted and accessed by Youzi Technology the next day. It turned out to be a repaired version of the Orange large model after internal testing.
After Wu Ende left, Fu Yuansheng, who took over his job, was full of promises, but he was still nervous in his heart.
He is an expert. Although he has never come into contact with neural networks like the Orange Big Model before, he is very familiar with models such as BP, LSTM, CNN, RNN, DNN, Seq2Seq, etc.
No matter what kind of neural network it is, there is no such thing as deleting data. You can only retrain it.
Deep learning does not store data somewhere for neural networks to access, but allows the neural network to "learn" these contents. What is learned will cause changes in "neural synapses", and the "neural synapses" of the large model are the so-called "parameters".
The numerical changes of these parameters and their mutual connections are like the connections of synapses when humans learn, which means that the large model has truly "mastered" the trained knowledge.
For example, a large model with 13B (13 billion) parameters stored on a hard disk may be 20G in size.
After training it with 2T of data, its size may still be 20G.
The size has not changed much, but it has already fully mastered the 2T information.
However, since it is an artificial intelligence model that imitates human neural formation, it actually has some problems of human intelligence.
For example, the problem of "forgetting" is that AI has learned something, but after a while it is found that it has forgotten what it has learned.
After AI is over-trained, it may take new data for granted.
Just like when humans face new things, they don’t understand anything, but use past experience to explain them. This is called "overfitting" in the field of AI.
In one class, the teacher covered half a book. You could understand the first three pages, but the rest was incomprehensible. For AI, this is "information overload."
When humans are exposed to one-sided information, it will lead to cognitive bias. If AI data is concentrated with content of a certain ideology, it will also lead to cognitive bias.
All kinds of things.
What's worse, it turns out that it is a good model, and others can train it just fine, but when you train it, it either fails to converge and get the result (which can be understood as not being able to learn), or suffers from catastrophic forgetting, or the gradient explodes or disappears.
Whether it is possible to train a usable large model without major changes to the architecture and parameters is even a mystery.
The training and tuning of AI are full of complexity and unpredictability.
Sometimes, a small parameter adjustment or change in data processing steps can lead to huge differences in training results.
Even if Fu Yuansheng strictly followed scientific methods to adjust hyperparameters, carefully selected data sets, and used the most advanced hardware equipment, he still did not have enough confidence to retrain the Orange model to remove all traces of Pomelo Technology in just one or two weeks.
But if it succeeds, I will be the father of AGI!
Get a promotion and a raise!
Become the world's number one AI authority!
How many people can resist such temptation?
And Chihiro was already prepared. If Yuzu Technology fought back, or even released evidence that the large orange model was their self-developed product, Chihiro would not be afraid.
The number one public relations officer who controls over one billion in funding is no pushover!
snort!
After entering the Pass, there will be great scholars who will debate with me.
First, they stirred up trouble, and then in the history books that followed half a year later, Youzi Technology soon became a laughing stock in history!
After Fu Yuansheng actually got his hands on it, he was overjoyed and felt that he was destined to be one of them.
This big orange model and grapefruit training framework are so damn useful.
No matter what strategy is used to adjust the parameters, the training effect is almost always good and the efficiency is extremely high.
A round of training should be completed in three days.
The public beta version of the Orange Big Model has doubled the number of parameters compared to the 13 billion parameters of the internal beta version, with a total of 150 billion parameters.
Fu Yuansheng commanded his three loyal lackeys, and the four of them began to retreat in a small dark room. They removed the traces of grapefruit technology in the orange model, adjusted the parameters in a targeted manner, and carried out repeated training to cover as much knowledge as possible that the model had originally learned.
Especially for questions like "Who created you?", the large model must be supervised to learn so that it can answer similar questions with "I am 'Wen Jun Yi Yan', an AI assistant, and Qianxun Company created me."
Or, "Chihiro's talented engineers worked day and night to allow me to meet you. Are you happy?"
In order to completely avoid problems, Fu Yuansheng and his team even banned "Yuzi Technology", "Yuzi" and other associated words, and prohibited the modified "Wenjun Yiyan" from being mentioned in any way.
Of course, this still cannot completely solve the problem, but it doesn’t matter. Qianxun is only planning to conduct the public beta for three days. After three days, the public beta will be announced to be over and the system will be upgraded. We will take the rest of the time to work on it slowly.
This problem can always be solved.
Fu Yuansheng's eyes are red from staying up late. Is it because he is tired?
No, this is because of my bright future!
——
The statement made on the official Weibo account of Youzi Technology left all the onlookers confused.
That sentence is interesting and can be interpreted as a counterattack. What does the photo of the team wearing magic robes posted above mean?
Want to use Avada to bite the big melon on Chihiro?
You are the person involved, what are you talking about?
After seeing the official Weibo post from Youzi Technology, it was likely that almost all of Qianxun’s management team came out to criticize Youzi Technology out of guilt.
Niu Liang, Vice President of Qianxun Financial Services, a Weibo-certified account, said: "It's hard to understand how anyone would believe that a small company less than a year old could change the world."
Jing Hao, the certified Weibo user of Qianxun Jiangmi General Manager: "Ten years of hard work, thumbs up to Qianxun. We will always be the young man pursuing technology and dreams! @Robert李"
Yuan Shanjun, Weibo-certified Vice President of Operations for Qianxun, said: "Fairy tales are beautiful, but unfortunately, the real world doesn't believe in fairy tales. Qianxun has been investing at least tens of billions of yuan annually in the AI field since 3057, and now it's finally bearing fruit, which is cause for celebration. The AI industry is capital-intensive and technology-intensive, and it's not something that any small company can take advantage of."
Youzi Technology has reposted these people’s Weibo posts, and each time it was reposted, it was accompanied by the phrase “interesting” and the team photo.
And under the huge public relations expenses of Qianxun, public opinion began to be directed towards attacking Youzi Technology.
Those articles that calmly analyzed that Youzi Technology could not be a plagiarist basically had no traffic, while the official Weibo account of Youzi Technology was filled with internet trolls and insults based on unknown facts.
Ninety-nine percent of the responses believed that Yuzu Technology’s response was not something to be ashamed of but something to be proud of.
As for these people, Fang Yu, who was using the official Weibo of Youzi, didn't care at all and joined in the scolding with a small account.
Because the insults were so distinctive, many of the small account’s Weibo posts were actually forwarded by many senior executives of Qianxun.
The public betas of Qianxun and Youzi Technology were launched on the same day and time, the morning of April 5th, the first day after the Qingming Festival holiday.
Since the public beta test time of Youzi Technology was announced as 10:00 am on April 5, in order to be the first to taste it, Qianxun specially moved the public beta test time forward by one hour to 9:00 am.
Unlike Youzi Technology's inaction, Qianxun has been very tight both internally and externally these days, trying hard to control media comments and fixing bugs and conducting tests internally.
"Please summarize the ROI for R&D of all the group's R&D projects in 3061 and form a report based on the uploaded template." Robert Li entered a line of requirements in the dialogue box.
"Wenjunyiyan" also has an enterprise API. Now Robert Li connects Qianxun's backup server with the enterprise API and develops Qianxun's internal AI management assistant specifically.
After receiving the instructions, "Wenjun Yiyan" quickly summarized and arranged the ROIs of all Qianxun's R&D projects in 3061, and conducted a comprehensive evaluation of these projects based on the multi-criteria decision analysis (MCDA) method, and then generated a PPT in just over ten seconds.
Robert Li opened the PPT and saw that it was clearer than the reports usually brought by the CTO.
As for the tests that might reveal the truth, they had already done them.
"Clap, clap, clap, clap." Seeing the excellent performance of "Wen Jun Yi Yan", Madeline took the lead in applauding.
The other executives also applauded.
From clapping their hands lightly in front of their chests to raising their hands above their heads and clapping vigorously, the conference room was soon filled with thunderous applause, led by the clapping thief.
It seems that whoever claps less has no sense of belonging to the company.
Robert Li looked at Madeline. Until now, he still felt uneasy, so he never posted on Weibo about this matter.
But in this situation, he could only smile as usual and patted Fu Yuansheng on the shoulder.
"Robert, the public beta test is tomorrow. As the sole leader of Qianxun, shouldn't you also support 'Wen Jun Yi Yan' on social media?" Madeline said with a victorious smile. Although it was a suggestion in her tone, Robert knew in his heart that this was a disguised order.
Robert Lee pursed his lips and remained silent for a while.
"Robert?" Madeline winked at him.
Robert then realized the atmosphere around him seemed a bit strange. He smiled, pulled out his phone, and posted a Weibo post: "Tomorrow at 9 o'clock, witness a miracle." He then joked: "I saw that the official Weibo account of Youzi Technology retweeted many of our executives' Weibo posts. I wonder if mine will be retweeted too."
When the boss posts on Weibo, the executives below smartly start commenting and forwarding it immediately.
At the same time, you have to write a short essay of several hundred words to express your gratitude to the company and your admiration and worship for your boss.
The flatterer who types the fastest hasn't even finished writing his essay when he sees that someone has already forwarded it.
Fuck! Who’s so quick?
I opened it and saw, Yuzu Technology?
And it even comes with pictures?
But this time they didn't post the photo of the team wearing magic robes, but instead posted a photo that everyone is familiar with.
Yep, you. (Mature readers should add the pictures themselves)
In the field of AI, due to the black box nature of AI, plagiarism and copycatting are very different from other IT fields.
For a complex system, wrapping is a very troublesome thing. With billions of codes, any comments in any corner may reveal the truth.
However, a large model is not a software system, but a neural network composed of tens, thousands, or even trillions of parameters. This makes it difficult to identify plagiarism and copycat behavior directly through code inspection, because the core of a large model is parameters rather than explicit code logic.
The knowledge in the model can completely clean up and cover up past knowledge as long as the training is repeated enough times.
This is also why the two Stanford students were so bold as to directly cover up Tsinghua's llama small steel cannon and claim it as a large model they made.
When this incident first broke out, some people only doubted its performance because the performance of the models produced by these three students was too similar to that of a small steel cannon.
At this time, none of the three people admitted that they were exposing the truth.
It's just that they are just students, and they lack computing power and data to completely complete the knowledge coverage of the model. As a result, their model can actually recognize the Tsinghua Bamboo Slips just like the proprietary features of the Little Cannon, which proves that they are plagiarizing.
The reason why the "Little Steel Cannon" incident was resolved so quickly was that the "Little Steel Cannon" was launched much earlier than the model produced by the three students, and it had already gained fame and had unique features. But what if the plagiarist was not the three students, but a large manufacturer with more abundant resources? What if the two were launched at the same time, with almost no difference in time?
(End of this chapter)
Continue read on readnovelmtl.com
