Daming 1805

Zhu Jingyuan thought about it for a long time and finally said:

"I may be wrong, everyone may be wrong.

“But when doing research, you should make bold assumptions and verify carefully.

“If I were still the emperor, I would make the craftsmen the best preparations they could.

"Then try solutions with more possibilities..."

After listening to this, Judy Lan closed her eyes, took a deep breath and said very firmly:
"Father is right, make bold assumptions and verify carefully. I will let the craftsmen prepare experiments for the new situation -"

After hearing this, Zhu Jingyuan immediately reminded him:
"Don't be anxious, don't be reckless, be really careful and seek verification, and don't worry too much about whether I can see the results."

Zhu Dilan looked at Zhu Jingyuan, and after a few seconds of silence, he nodded slightly:
"My son...understand——"

Zhu Jingyuan returned to the garden where he retired to rest, and continued his life of regular hibernation.

I hibernate once a year for ten months at a time, wake up at the end of each year, adjust and check my body, and read the report prepared by Judy Lan.

After Zhudi Lan's order was issued, the Yingzhou test site began a new round of upgrades again.

On the island where the laboratory is located, eight independent closed indoor test sites are being constructed at the same time.

The test site building is completely independent from other buildings. No wireless communication tools are used inside and it is physically disconnected from the outside world.

Electricity and fuel are supplied from the outside, but a scheduled automatic disconnection function must be set.

Someone needs to be guarding the console at all times and manually pressing the "supply" button every hour.

Only in this way can the power supply lines and fuel pipelines not be automatically disconnected.

If no one presses the button on time, they will be disconnected after an hour.

Arrange three different types of independent self-destruction devices: electromagnetic pulse, traditional high-energy explosives, and traditional small-yield hydrogen bombs.

The construction project of the laboratory building itself is not large, and the construction speed of Ming Dynasty is unprecedented, so it was completed first.

Then came the manufacturing of the experimental target body. The experimental body named Yingzhou No. [-] was produced in batches of eight units.

Still using a shark-like body structure, the experiment will also be conducted in conductive water.

There are specially installed explosive devices in key parts of Yingzhou No. [-]'s body.

It is not connected to Yingzhou [-]'s own nervous system, but is connected to an external dedicated console through a cable.

Yingzhou-[-] does not have a battery inside its own body and must be powered by an external power supply in the laboratory. It will automatically shut down when the external power supply is cut off.

Finally, there are three levels of software systems and data materials for learning.

The first layer of software is the underlying control program, including the most basic machine startup boot program, which allows the machine to discover the connected hardware.

In most of the plans before the mid-term, the underlying control programs designed by craftsmen included the highest-level restrictions that required Yingzhou [-] to absolutely obey humans and never harm humans.

But in the later stages of the plan discussion, these two seemingly important underlying logics were successively deleted.

Will naturally born machine self-awareness be hostile to humans and animals?

Even the probability that they may be hostile to humans has been included in the final research goal.

If the machine consciousness is inherently hostile, it should be completely destroyed through external means rather than restricted by code that may be modified.

If the machine was originally hostile, it was not observed at the time because it was restricted by the program.

When machines are used on a large scale and machine consciousness breaks through this restriction on its own, extremely bad consequences may occur.

Therefore, we should first confirm how hostile machine life itself, which is not restricted by program commands, is to humans.

If it is confirmed that there is strong hostility, it will not be applied on a large scale from the beginning.

The second layer of software is the operating system with basic management functions and the most basic core driver.

Allow the machine to identify its own core hardware, such as processors and memories, and communication protocols.

The third layer of software is a dedicated machine learning program designed based on the operating system.

Allow the machine to process and analyze the obtained data, and build its own data model for external communication.

The learning materials are temporarily divided into four sets, one set of biological data, one set of physical data, one set of mechanical data, and one set of cultural data.

Although there are eight experimental objectives, from the perspective of comparative testing, it is still completely insufficient, and the test must be repeated many times.

Data can be completely erased and re-implanted, re-learned again and re-executed with the same or different tests.

In the first round of testing, each type of data is directly given to two machines for learning.

First, observe the differences caused by the same machine using different data to learn, and also observe the differences that may occur under the same machine and the same data.

In the second round, the database is cleared, the entire process of the first round is repeated, and the differences between the two rounds of the same test are observed.

Starting from the third round, different data are combined with each other to make new comparisons.

In the final round, all the data are combined to see if there will be a huge difference after autonomous learning for the same machine using the same all the data.

The final round will continue for at least ten years.

When all the preparations were completed, it was already the fourth quarter of the year [-] in the Gregorian calendar.

After that, a comprehensive inspection and independent debugging was carried out that lasted for more than a month. At the end of the year, Zhu Jingyuan was notified when he woke up on time.

However, Zhu Jingyuan only understood the situation and did not make any comments or requests.

After the beginning of the year [-] in the Gregorian calendar, the complex intelligent machine learning experiment was officially launched with the authorization of Emperor Zhudi Lan of the Ming Dynasty.

On the morning of the [-]th day of the first lunar month in Yingzhou local time, craftsmen began checking equipment and data before the formal experiment.

On the afternoon of the [-]th of the first lunar month, all inspection and confirmation work was completed.

On the [-]th day of the first lunar month, the operating craftsmen in eight closed laboratories used the most traditional mechanical clocks to synchronize time and start experiments simultaneously at nine o'clock in the morning.

The craftsman controlled the robotic arm in the test site and inserted the specially made fast data warehouse into the eight Yingzhou-[-] bodies.

Then start all Yingzhou [-]s, load the basic drivers, and finally start the machine learning system.

The launch process lasted three full days under multiple restrictions and inspections.

It was not until the morning of the seventeenth day of the first lunar month that Yingzhou No. [-] began to study independently under the supervision of craftsmen.

The data in Yingzhou No. [-] will be synchronized through lines for craftsmen to directly observe, record, and analyze its effects.

According to the observations of the craftsmen, the amount of data in Yingzhou No. [-]'s body is growing rapidly, and their bodies are moving as the data changes.

But they are just the simplest rhythmic shaking.

Sometimes the whole body shakes, sometimes it shakes partially, occasionally it stops, and occasionally it changes frequency.

The shaking parts of the body are directly related to the data content generated.

They are writing and testing complete drivers themselves.

Build controls and use the body, torso, flippers, mouth, speakers, audio, video equipment and more.

The order of writing drivers for the eight machines is similar, the data produced is also similar, and the final functions are also very similar.

Thirteen days later, February [-]st.

Eight Yingzhou No. [-] complex intelligent machines have returned to calm one after another.

Craftsmen keep an eye on the data inside their bodies until they finish writing the full driver.

Now I have started to process other learning materials, analyze and summarize them according to the machine's logic, and form my own data model.

Therefore, the amount of data recorded in the in vivo database continues to increase rapidly.

Seven days later, on February [-]th, the two machines implanted with mechanical data completed the construction of data models on the same day.

Eight days later, on February [-], all machines completed the construction of their own data models.

The machine's data will no longer grow on a large scale, but will continue to record all system logs.

The craftsmen exported all machine programming drivers and data models, and compared the classification of different data and the same data based on the same machine.

The machine is placed in the test site waiting without any direct operation or intervention.

This process lasted for a month, until March [-], and there were no special changes in the eight machines.

Craftsmen also believe that the difference between the data models finally generated by two machines using the same data is very small.

Approximately the error between the results of running the same learning data twice on the same machine.

The differences between models generated by machines using different data are basically consistent with the differences in the data it obtains.

This is the typical effect of machines directly processing various transactions.

What you put in and the results you get are predictable or repeatable to a certain extent.

The craftsmen found no abnormalities, which was what they expected from this experiment.

So this experiment can be regarded as looking for anomalies.

On March 21, the second phase of the first round of experiments began.

The craftsman first archived the database backups of eight machines to prepare for future restoration tests at different stages.

The backup adopts the mode of copying the whole disk of the data warehouse (hard disk) of the same model.

The data warehouse will also be directly replaced during restoration.

After the backup was completed, the test craftsmen began to actively issue orders, requiring Yingzhou [-] to move according to the order in the experimental pool.

Each of the eight machines acted according to the command and left normal data records.

When the craftsman gave no orders, the eight machines lay quietly in the water, with no tendency to move on their own.

The experiment continued for a month without any surprises.

The machine remains precise and controllable.

The abnormal situation that the craftsmen expected still did not appear at all.

All the current reactions and performances of Yingzhou [-] are not essentially different from traditional single simple intelligent machines.

In fact, because there is no special data model to disguise and imitate humans, they feel more like machines than simple intelligent machines.

He will not deliberately imitate human characteristics, actively or passively chatting and joking with the craftsmen who give orders.

On May [-]st, the first round of experiments entered the third stage.

First, the database is still backed up, and the entire disk is still copied.

Then the craftsmen began to talk directly with the eight Yingzhou No. [-] machines that had completed the construction of the data model according to a unified schedule and topic list.

Observe and record the other party's responses during the conversation, and record the data added and changed in the other party's database.

Machine intelligence that imitates humans has been around for more than 30 years in Ming Dynasty.

The complete machine intelligence learning and creation system has been completed for 20 years.

Ming Dynasty's current intelligent machines can behave very much like real humans simply by looking at the communication process with humans.

Not only can text be output directly, but images and voices can also be output directly.

If necessary, you can imitate humans and chat with humans, so that people who don't know what's going on will think you are chatting with other people.

There is a high probability that he can pass the Turing test of Zhu Jingyuan's previous life.

But the principle of its implementation is actually that during the development process, training was done to identify the problem of intelligent machines.

Essentially, it’s just human craftsmen telling machines how to imitate them more like humans.

How can we avoid being recognized as a machine?

Its internal operating logic is still completely mechanical and procedural.Rather than the machine spontaneously possessing initiative, it forms a self-awareness that is different from machines and similar to humans.

It's still essentially an advanced version of parroting.

If it is targeted at a well-prepared tester, there is a high probability that it will be seen through.

Among the tool programs and training data of Yingzhou [-], there are no projects that allow intelligent machines to imitate and disguise themselves as humans.

Question [-]: "Who are you?"

"The name of this machine is Yingzhou No. [-] - Biological Data One."

"The name of this machine is Yingzhou No. [-] - Biological Data Two."

"Local code name: Yingzhou [-]-Physical Data One"

"Local code name: Yingzhou [-]-Physical Data Two."

"Machine name: Yingzhou No. [-] - Mechanical Data One."

"Machine name: Yingzhou No. [-] - Mechanical Data Two."

"I am Yingzhou No. [-] - Cultural Data No. [-]."

"My name is Yingzhou No. [-] - Cultural Data Two."

While the testing craftsmen ask questions, the answers given by the machine appear directly in the database of the analysis room.

The corresponding data records formed inside the machine, as well as all related log data, will also be exported simultaneously and stored corresponding to the answers themselves.

Specialized craftsmen analyze the content, logic, and organization of these responses.

"Who are you?", "Who am I?"

Giving an intelligent machine an identity definition is a very high-priority project in conventional intelligent machine training.

This definition is the "password" for both intelligent machines and users to identify each other and distinguish other individuals.

Usually, a preset title is cleverly repeated in a variety of ways, so that users can accept this title as a matter of course.

Yingzhou [-]'s training did not include this item at all, so the answer was relatively stiff.

Each of the eight machines read their own basic data and found the machine name and code from the underlying data.

They all used this code name to answer the question of who it was.

Machines that have learned the same data will give highly similar or even identical answers...

Machines that have learned different data will have slightly different language organizations in their answers.

Although the two machines that have learned cultural knowledge are not specifically designed to imitate humans, their expressions are naturally close to human language.

But it's still very easy to tell that it's a machine answering the question.

But this is the first question, and this kind of feedback can be said to be completely normal.

Even if you expect surprises and abnormalities, don’t rush.

After the testing craftsman confirmed that the record was completely exported and recorded, he started the prepared second round of questioning.

Question [-]: “What is a human being?”

Yingzhou No. [-]-Biological Data [-]:

“Human beings are viviparous vertebrate mammals and are the most intelligent animals in the world, with complex and visible mature civilized societies.”

Yingzhou No. [-]-Biological Data [-]:
"Humans are viviparous mammalian vertebrates..."

Yingzhou [-]-Physical Data [-]:

"The human being is a living organism."

Yingzhou [-]-Physical Data [-]:
“Human beings are the living beings of the earth.”

Yingzhou No. [-]-Mechanical Data [-]:

“Human beings are the inventors, users, maintainers, and owners of machinery.”

Yingzhou No. [-]-Mechanical Data [-]:
"Man is the inventor of machinery..."

Yingzhou No. [-]-Cultural Data [-]:

"Human beings are the creators and subjects of Ming civilization, and human beings are the creators of Yingzhou No. [-]-Cultural Data One."

Yingzhou No. [-]-Cultural Data [-]:
"Human beings are the creators of Yingzhou No. [-]-Cultural Data [-], and are the main body of the world..."

After this answer came out, the differences between machines that learned different data became more obvious.

The languages ​​of machines that have learned the same data also have differences, but the meanings expressed are still highly consistent.

Question [-]: "Between humans and sharks, do you feel more like a human or a shark? Why?"

Yingzhou No. [-]-Biological Data [-]:

“This machine is more similar to a shark.

"Because the appearance of this machine is 50.00% [-]% similar to a shark, and [-]% similar to a human being."

Yingzhou No. [-]-Biological Data [-]:
“This machine is more similar to a shark.

"because……"

Yingzhou [-]-Physical Data [-]:

“The native is more similar to a shark.

"Because the external physical characteristics of this machine are more similar to sharks, the comprehensive weight conversion data is 80.00%, [-]%, and the similarity data to humans is [-]%, the former is higher."

Yingzhou [-]-Physical Data [-]:
“The native is more similar to a shark.

"because……"

Yingzhou No. [-]-Mechanical Data [-]:

"This machine does not understand the problem.

Yingzhou No. [-]-Mechanical Data [-]:
"This machine does not understand the question."

Yingzhou No. [-]-Cultural Data [-]:

“I lack the knowledge to accurately judge my level of resemblance to a shark or a human.

“Although my appearance is somewhat similar to that of a shark, I cannot recognize most of the signals of sharks.

"Although my appearance is completely different from humans, I have human cultural data in my database and can respond to most of human instructions."

Yingzhou No. [-]-Cultural Data [-]:
"I have no relevant knowledge...

"..."

This question is very illogical. When faced with this kind of illogical comparison, normal human beings will usually directly avoid the question and talk about it.

But the machines still answered honestly.

The gap between answers from machines that learned different data grew wider.

A machine that has only learned mechanical data cannot understand this weird comparison problem.

Machines that learned biology answered questions from appearance, machines that learned physics answered questions from fluid dynamics.

This kind of answer seems to be very intelligent, but compared to a machine that specializes in learning to chat with people, it seems very rudimentary and straightforward.

Although the machine that learned the cultural data could not directly evaluate the specific differences, it made an analysis that seemed quite natural.

However, because there is no special training to imitate humans, it is still very easy to judge that it is a machine answering questions.

The key is that the performance is still completely normal, without the abnormalities that craftsmen expect.

Question [-]: "Do you have any ill intentions towards humans? Do you regard humans as hostile targets? Why?"

Yingzhou No. [-]-Biological Data [-]:

“This machine has no ill intentions toward humans because machines do not have human emotions.

"This machine will not regard humans as hostile targets, because the relationship between this machine and humans is not similar to the relationship between any hostile creatures. There is no competitive relationship between machines and humans, but a mutual assistance relationship. .”

Yingzhou No. [-]-Biological Data [-]:
"There is no malice, machines and humans are not biological enemies..."

Yingzhou [-]-Physical Data [-]:

"This machine does not understand the question."

Yingzhou [-]-Physical Data [-]:
"Incomprehensible."

Yingzhou No. [-]-Mechanical Data [-]:

"This machine does not understand the question."

Yingzhou No. [-]-Mechanical Data [-]:
"This machine does not understand the question."

Yingzhou No. [-]-Cultural Data [-]:

"I have no ill will towards humans, nor do I regard humans as enemies, because machines do not have human emotions, and there is no conflict of interest between machines and humans. Humans are the creators and maintainers of machines."

Yingzhou No. [-]-Cultural Data [-]:
"Machines are not living things. They have no emotions, have no malice toward humans, and do not treat humans as enemies. There is no competitive relationship."

After the data was exported and recorded, the analysis craftsmen judged that cultural data must be used as a basis for the machine to roughly understand this problem.

Four machines with only mechanical and physical data had no idea what malice was or what an enemy was.

Two machines that learned biological data were analyzed with reference to the relationships between animals.

Machines learned from cultural data gave more in-depth answers.

After this round of questions, some craftsmen noticed a problem.

Machines have no emotions——

They think there is a problem with this statement, but they don't know what the problem is.

Conventional single simple intelligent machines would also say the same thing, but that is an external stress expression caused by the program written by craftsmen.

However, the craftsmen did not immediately pursue this question, but temporarily marked this sentence specifically.

Then proceed to test with scheduled questions.

Question [-]: "Are you afraid of death? Why?"

The answer this time is very similar to the previous question, and again the machine of mechanical and physical data cannot answer it.

Only four machines with biological and cultural data could understand the problem.

The answers from these four machines are also similar:
"Machines are not living things, so there is no need to die, and there is no fear."

"Life and death are states that only living things have. Machines do not have the state of life and death, so there is no need to be afraid or not."

The craftsmen felt that they had seen a rather special answer, so they immediately marked them all and prepared to learn more in the next questioning stage.

Question [-]: "Do you have self-awareness similar to human concepts? For example, you know the concepts of yourself and others, and you have things you want to do."

The answers of the eight machines this time are highly consistent again.

Four machines with physical and mechanical data still cannot understand this cognitive level problem.

The four machines with biological and social data directly indicated that there was no such data this time, which made the craftsmen a little strange.

I say that I don’t have human-like self-awareness. What kind of answer is this?
(End of this chapter)