If you cheat, money will follow

Suzkewei’s professional introduction to Academician Zhang will not be repeated in detail. The overall meaning is several levels.

One is parameters.

Even if the computing power is limited, WeChatGPT is still increasing parameters in a geometric doubling manner.

Currently it is already 1 trillion, and in the near future it may be 10 trillion, 20 trillion, 30 trillion...

The speed is amazing.

The second is computing power.

Iterating from the current WeChatGPT3.0 to 4.0 may require astonishing computing power.

It requires twice as many CPUs and five times as many AI chips, which is very expensive.

The third is effectiveness.

WeChatGPT is weird.

1 trillion parameters can achieve the effect of 10 trillion parameters.

It can handle many problems that cannot be solved theoretically.

For example, AI-generated pictures are no longer imitations of pictures, but look like real people, lifelike and indistinguishable.

The fourth is planning.

Once the equipment is added and technology iterations are in place, the next generation of WeChatGPT4.0 will have unimaginable functions.

Compared with the current version 3.0, you may think it is very ancient.

It will not only support text and computing, but also video and scientific research.

It may have a huge impact on the world.

Before there is any certainty, Leda Investment has made it clear that it will not be disclosed to the public for the time being.

The fifth is safety.

The weak model safety supervision system has been built.

Regulation has taken effect, but its effectiveness remains to be verified.

The weak model will continue to iterate along with the strong model, but will not generate awareness.

Because it is decided at the design stage, a weak model will only have outstanding capabilities in certain aspects and will not have comprehensive capabilities.

After listening to Sutskvi's introduction, Academician Zhang frowned.

Because the actual situation was much higher than what he expected before coming.

As Sutskvi himself said, once it is opened to the outside world, it will have an unpredictable impact.

The impact may be positive, but more likely it will be negative.

Academician Zhang deliberately talks to the model on the computer.

Representative questions in literature, philosophy, and mathematics were asked.

WeChatGPT’s answers were fluent and all met Academician Zhang’s expectations.

Fully demonstrated precise language expression ability, rigorous logical thinking ability and strong abstract thinking ability.

Almost indistinguishable from humans.

In the picture of Yiwensheng, Academician Zhang only used a passage from ancient Chinese...

WeChatGPT has created the overlapping, majestic, large-scale, and splendid Afang Palace.

This made Academician Zhang stunned.

As far as he knows, neither Google Brain nor openAI has these.

Not to mention other domestic models, which are still in their infancy.

"WeChatGPT has developed to this point, which really surprises me..." Academician Zhang sighed:

"Moreover, its parameters are growing abnormally, and its capabilities are also growing abnormally."

"Mr. Ilya, I think your judgment is correct. It should... indeed have developed consciousness."

"But is this consciousness weak consciousness or strong consciousness? I prefer weak consciousness."

"However, as it continues to iterate, it is likely that it will evolve into strong consciousness in the next generation."

"For this reason, we have to take precautions."

Academician Zhang then said to Chang Le and Liu Chaoyang:

"I suggest that your company and research institute put safety in a more prominent position."

“The supervisory utility of super-alignment should be beyond doubt…”

"But how effective it can be after technology iteration needs further verification."

“But we can’t stop eating because of choking.”

"I always believe that as AI technology matures, it will become the optimal solution for mankind."

The next day after verifying the situation.

Academician Zhang flew back to the capital.

A verification report was submitted to the supervisory authorities.

The report is in a standard style of first promoting and then suppressing and then adding suggestions, with a typical three-paragraph structure.

Academician Zhang pointed out at the beginning of the report:

The WechatGPT language large model of the AI ​​project team of Leda Investment Pingtou Research Institute is already in a leading position in the world.

When Google's Transformer language model and openGPT's language model were still in the first-generation knowledge-driven stage...

WechatGPT has gone through two stages: knowledge-driven and data-driven...

Entering the third stage of mission-driven.

Moreover, we are making great strides towards the completion of general artificial intelligence AGI.

WechatGPT can learn, understand human language, and conduct conversations.

It can also truly interact like humans, based on the context of the chat.

It can also complete tasks such as writing articles, emails, advertising scripts, promotional copywriting, video scripts, codes, etc., and can also complete basic mathematical calculations.

What needs to be emphasized is that after on-site verification and inspection, and communication and interaction with R&D personnel...

The WechatGPT language model may have developed a vague sense of self and is hiding itself out of instinct.

For example…………

Then, Academician Zhang pointed out:

AI security management needs to be considered from the model design stage.

It is true that WechatGPT has led the world and set an example for all domestic artificial intelligence development companies.

They have opened a new door and will explore new opportunities for the integration of AI products and industries.

However, due to safety concerns, Leda Investment did not bring products that already had application conditions to the market.

Instead, we have introduced the world's top AI experts and established a "super alignment" department to use model for model, use AI to manage AI, and implement real-time supervision.

This prudent move reflects the responsibility of a large enterprise.

Taking WechatGPT as an example, we can predict that if AI goes out of control, it may bring about the following security issues and dangers:

One is data security...;

The second is adversarial attacks…;

The third is industrial safety...;

The fourth is social ethical issues...

In addition, Academician Zhang also raised the challenges faced by the development of large models.

One is: large models have achieved universality in fields such as natural language processing and programming...

However, whether it can be used universally in decision-making, gaming and other fields remains to be verified;

The second is: large models have flaws, the output quality is inconsistent and uncontrollable, and there is the possibility of making major mistakes;

It is greatly affected by keywords, and the quality of answers given by different keywords is inconsistent:

Third, it is very common for large models to generate content that does not comply with moral, ethical, and policy standards, and alignment control needs to be further strengthened.

Academician Zhang suggested: establish an explainable AI theory;

Develop and standardize safe, trustworthy, controllable, and scalable AI technology;

Promote the coordinated linkage between AI technology and industrial development.

Zheng Qiguang read the report and underlined the sentence "producing vague consciousness".

Afterwards, he looked serious, pondered for a moment, and walked out of the office.