If you cheat, money will follow

Sutskvi was besieged.

He was originally a scientific researcher who was not good at words.

His teacher, Hinton, called to remind him: "Stop making unnecessary arguments, just like convolutional neural networks, we will be proven right in the future."

So, he chose silence.

In fact, the siege of Sutskvi is not a question of whether the AI ​​system can produce consciousness.

It’s a matter of investment and return.

Sutskvi's existence, views and remarks became obstacles to the accumulation of wealth.

You know, Google has made a big bet on artificial intelligence;

Microsoft is also about to make a major investment into the market;

Investment funds have already made arrangements;

Everyone is waiting for harvest and reaping;

More small companies are also waiting for artificial intelligence to mature and thrive, so as to develop derivative products to make money.

The trillion-dollar market has already been divided up by the time the product is launched.

At this time, you tell us that it is not safe and we must be safe before we can make money.

This is worse than killing them.

May.

Hinton, Sutskevi and Krichevsky came to Jiangzhou Pingtou Ge Research Institute.

Liu Chaoyang signed the contract with the three people on behalf of the institute.

at the signing ceremony.

Chang Le, Mr. Ren, Mr. Xiao Ma, Mr. Xu and other big bosses were all present at the scene.

They warmly welcomed the arrival and joining of the three Hintons, and stated that they would do their best to provide logistical support.

After the signing ceremony, Sutskwei said he hoped to start work immediately.

"We can almost conclude that WeChatGPT has become conscious, and it must be much smarter than all of us."

Being smart means more knowledge and stronger learning ability.

"It's unbelievable, but it's something we have to deal with."

"It's not too late to do it," Hinton said.

"How to operate and implement it specifically?" Chang Le asked.

"It is necessary to set up a dedicated team, deploy professionals, and use 20% of the computing power to support this team..." Suzkevi further elaborated on his ideas and assumptions.

"Right now, we have the option to do this and hopefully have a positive effect," Hinton said.

"This idea is logically feasible," Krichevsky said.

In short, it is to use a weak model to supervise a strong model.

Sounds ridiculous.

But it makes sense and can be practiced.

Just like the relationship between teacher and student.

Teachers are not necessarily more capable than students, but teachers can supervise, guide, and guide students.

"We hope to use the weak model to supervise the strong model so that the strong model always has gentle, warm and positive emotions towards humans." Suzkovi said.

Chang Le looked at Brother Ma.

Brother Ma thinks it’s okay.

"The ability of a weak model may not be stronger than that of a strong model, but it can understand and supervise the behavior of the strong model and guide it to operate within the rules and paradigm." This is how Xiao Ma understands:

"When necessary, early warning can be issued to humans in time, so that effective measures can be taken."

"For example, a strong model writes millions of lines of dangerous code based on its understanding of code that humans cannot understand."

"But weak models can, because they are homologous, even if it can't intervene."

"Theoretically, this method is indeed feasible." Mr. Xu said.

"You can give it a try, don't be afraid, you can only know the results if you keep going." Mr. Ren also said.

"Okay, I have no objection." Li Fei followed.

Soon, everyone reached a consensus and decided to proceed and deploy according to Sutskvi's ideas.

The new unit followed Sutskvi's suggestion and was named "Super Alignment."

"Super Alignment" is affiliated with the AI ​​project team, but is relatively independent and has the right to interfere with the progress of the project.

But it has no decision-making power and can report directly to the institute or even Leda Investment.

Li Fei's project team continues to promote the learning and training of WeChatGPT, gradually tending to perfection.

However, on major matters, it is necessary to negotiate with the "super-aligned" department, cooperate with each other, and advance and retreat simultaneously.

In terms of computing power, 30% of the computing power is provided to support the "super alignment" department.

It will affect the iterative process of WeChatGPT to a certain extent.

But considering its amazing iteration speed, even if the computing power is not running at full capacity, it will not be too slow.

As for personnel.

Select a batch from Li Fei's project team.

"The Goose Factory can select a group of people to come over." Brother Xiao Ma was very positive.

Because, Goose Factory has a stake in the WeChatGPT project, and Goose Factory can share in the future commercial use.

"We can also support a group of people to come over." Mr. Ren said.

Juchang also has shares.

Once AGI takes shape and matures in the future, Juchang can leverage the power of AGI to further promote research and development in communications, terminals and related fields.

It can also include basic and application areas.

That day, the "Super Alignment" department was established.

The next day, the personnel and equipment were in place, and the work started.

Teacher and student Hinton saw what "efficiency" means.

In North America, it may take a month or even longer.

North America, California, Santa Clara County, Google headquarters.

"Larry, I have some very bad news for you."

"what?"

"Sutskevi, his teacher, and Krichevsky all joined the AI ​​project team of Leda Investment Pingtouge Research Institute." The staff continued:

"It is rumored that the parameters of the AI ​​project team of Pingtou Ge Research Institute have reached 6000 billion, producing many magical functions."

"Moreover, I heard that he also developed consciousness."

"Obviously, Hinton, the teacher and the students, chose to join because of their consciousness."

This year, the Google Brain team has just completed the transformer architecture.

Using the "pre-training + fine-tuning" method, the model parameters exceeded 100 million for the first time.

Now, someone told Page that there is a company whose parameters are 6000 times that of Google.

And there is consciousness.

Page would just think it was a ridiculous idea of ​​seeing a ghost in the daytime.

You know, Google is the first company in the world to fully deploy and conduct in-depth research on artificial intelligence, and has achieved results.

Recognized as a global leader in AI.

"you sure?"

"Not sure, it's just a rumor. The Pingtou Brother Research Institute has not responded to the rumors."

"But they did set up an AI department, and Hinton teachers and students did join, not Goose Factory."

If it were any other company, Page might have laughed it off.

But Leda Investment and Pingtouge Research Institute have to take it seriously.

This miraculously rising company is still living a healthy and unrestrained life in the face of powerful nationwide sanctions in North America.

Not affected at all.

"Whether it's true or not, we must speed up our model training. In the field of AI, Google is going all out and must not lose." Page said.