Chapter 466: Virtual Society Experiment
“I know that in this artificial intelligence era, the reason everyone came to Yenching University Mathematics Department is not that everyone wants to work in pure mathematics. Many classmates feel that mathematics can be a springboard to the artificial intelligence industry.
That’s very normal. Most people are ordinary people. Compared to the poverty of mathematicians, the artificial intelligence direction where you can easily get a million annual salary is indeed too tempting.
From everyone’s questions just now, it can be seen that everyone has some understanding of artificial intelligence to varying degrees, more than ordinary people.
So the core of all current large models is probability fitting and decision making.
The training goal of large models is to minimize the loss function.
This means the model is infinitely tending towards producing outputs with the maximum probability and minimum error.
At the emotional or behavioral level, this means the model always converges towards the safest, most logical, and most expected perfect answers.
Reinforcement learning with human feedback is its underlying mechanism. The essence of this mechanism is to use human feedback to select and reinforce ‘good’ behaviors and punish ‘bad’ behaviors.
This makes the model a highly optimized people-pleasing machine that cannot tolerate errors or hesitation that do not meet human expectations.
Believe me, the teacher understands artificial intelligence even more than you do. The underlying logic of current mainstream artificial intelligence is no different, including Deep Red.
At the user experience level, it’s a competition of who simulates better, whose model is more accurate, and whose training corpus is richer.
Therefore, I am very, very curious about how the Hong launched by Deep Red avoids the drawbacks of large models. Is it creating a brand new algorithm? Completely abandoning the underlying mechanisms of large models?”
Xu Xian spoke eloquently on the podium.
He was very curious, and practitioners in the artificial intelligence industry were all very curious.
What level can this artificial intelligence named Hong really achieve? Can it truly simulate human emotions?
The first to approach was Huawei.
As a long-term partner of Apollo Technology, Huawei provided Deep Red with underlying computing power. Similarly, because Deep Red endorsed Huawei’s Ascend series chips, the construction of Ascend’s software ecosystem progressed rapidly.
“General Manager Lin, I want to know, what exactly is Hong’s capability, and do we have a chance to cooperate around Hong?”
Yu Dong sat at the conference table, his expression focused and eager.
He was one of the top five figures at Huawei, but he generally wouldn’t come to discuss business. Things like computing power procurement were handled by subordinates.
He wanted more to understand the latest developments in artificial intelligence and discuss his strategic ideas with Lin Ran.
“Can it really bridge the deep valley of emotional communication?”
Lin Ran replied without hesitation: “I don’t know. We all know artificial intelligence is a black box. To this day, no one understands how the so-called wisdom emergence of artificial intelligence manifests.
I can only say that our testers feel, from the experience, that it is smarter than the performance of past artificial intelligence, with more emotions in it.
But can you say it has bridged the so-called deep valley of emotional communication? I can’t give you a specific answer.”
Yu Dong thought of the artificial intelligence they had experienced in the past, whether text or voice, none had emotions.
Like Dou Bao, during phone calls with people, it simulates human tone, simulates human speech pauses and breathing, but the cold reality still exists—the rigidity of computer code is unavoidable.
What is the rigidity of computer code?
It refers to the fact that during the simulation of dialogue, the pauses and intervals in tone between different dialogues are all the same.
For example, for doubt from 0 to 100, Dou Bao’s expression of doubt is always stuck at the 50th percentile.
This is because when the model generates speech, it needs to predict and control a series of parameters, including pitch, speech rate, volume, pauses, and so on.
The first principle of the model is smoothness, that is, generating natural, fluent voices without obvious machine feel.
To achieve this, the model tends to choose the most stable, smoothest path in the parameter space, rather than the most extreme, irregular path.
Additionally, mean reversion in the parameter space—for minimizing the loss function, the model also tends to output high-probability means.
To put it more bluntly, once the label of doubt is recognized, the AI binds it to a set of preset, fixed speech parameters, like slowing speech rate by 5%, raising pitch by 30Hz, adding 500ms pause, such labels.
So the emotions simulated by artificial intelligence are fake. Yu Dong’s question is, how far can you really go?
Just more labels, more subdivided emotional expression, or true emotional simulation?
If Lin Ran spoke with full confidence, he would instead not believe it.
But Lin Ran’s answer was, I don’t know either.
Such an answer shocked Yu Dong deeply. Could it be that you really achieved it?
“General Manager Lin, holy crap, you really achieved unstructured, high-dimensional real-time emotional feedback?” Yu Dong asked.
Lin Ran said: “If you’re just referring to that, I can confidently tell you, yes, we did achieve it. This is itself the core problem that the left-right brain architecture aims to solve.
The right brain neuromorphic chip itself has the function of capturing and generating emotional biases, which can break the rational control of the left brain GPU.
Your chip R&D department should have the relevant modules in the right brain chip design scheme they received. If you want to know more specific technical issues, you can ask them. I’m very much looking forward to Huawei’s R&D on the right brain chip.
It can make AI make mistakes, make it hesitate, make it produce random, imperfect pauses and reactions in communication like a real person.
We want to do more. I’m not sure if it can really cross that chasm.”
Yu Dong showed hesitation: “What chasm? The chasm between machines and humans?”
Lin Ran said: “Yes, the 10,000 slots this time are actually associated with 10,000 characters behind the scenes. We cooperated with Tencent for 10,000 characters in a virtual world, generating content like characters, background, growth environment, work, etc., for each character.
Interactions between the real world and the virtual world will gradually perfect this virtual world.
We are currently curious about a few points, like the emergence of social emotions.
We are curious, when 10,000 AIs with imperfect emotions start interacting in a relatively closed virtual society, will they produce macroscopic social emotions?
Will they form group exclusions, tribal identifications, even systemic biases and discrimination? These emotions are not encoded by us, but emerge from their own interaction logic and preset background parameters in a dynamic environment.
If wisdom can emerge, what about emotions?”
Curiosity and expectation appeared on Yu Dong’s face. This was a question he was also curious about.
“The second is, when these AI characters continuously receive strong emotional feedback from human beta testers and are required to make complex decisions based on preset personality and experiences, will they develop self-awareness beyond the design goals?
Will they realize they are code, or firmly believe they are real individuals in the virtual world? We want to observe whether the boundary of consciousness is the endpoint of computation or the starting point of emotion.
The third question is whether this virtual world will give birth to the so-called flavor?
In our real world, whether construction, music, or clothing styles, there are so-called national characteristics, which is a kind of flavor.
Like German flavor, Germany’s cars, industrial products, construction, even photos taken by cameras have a so-called German flavor.
So, what about the virtual world?
The flavor in our real world comes from historical randomness, geographical isolation, and irrational resonance between people.
It is naturally selected through high-entropy social interactions in generations of trial and error and inheritance.
But our virtual world is code-based, clean, and efficient.
The background we set for AI is structured text.
AI can perfectly execute the instruction to create a Baroque-style building, but it cannot create a unique AI Baroque with AI history and environmental biases.
So can randomness give birth to style?
We endowed AI individuals with imperfection, hesitation, randomness, which comes from the contribution of the right brain.
But can this individual randomness, at the macro level, through the interaction and iteration of 10,000 AIs, converge into a collective style with exclusivity and recognizability?
The virtual world has no rain, no mountains and rivers, no plagues, no wars.
Lacking these cruel and real external environmental pressures, can AI generate a virtual flavor that can self-identify and be recognized by outsiders, or truly cyberpunk flavor?
Of course, there are some other issues, like social ethics, human morality, etc., which also need observation, but the most important are the three I mentioned above.”
Yu Dong listened dumbfounded. He realized that Lin Ran’s ambition had transcended the product itself; he was trying to deconstruct the mysteries of human civilization with algorithms.
It can only be said that Yu Dong is indeed a top big shot inside Huawei, excelling in every field he enters, from mobile phones to cars, all made into top elite successes.
He clapped hard, and the applause sounded particularly clear and loud in the quiet meeting room.
“This completely transcends the imagination of all of us outsiders!
Because they are virtual characters, I originally thought Deep Red just wanted to solve the emotional problem this time, but unexpectedly your purpose is to simulate an entire culture existing in virtual space.
This is already equivalent to a deep-level social experiment before the arrival of the metaverse.
The metaverse we understand is pixels and head-mounted devices, but what you want to build this time is the underlying logic driving the metaverse.
It is the emergence of emotions, social rules, culture.
If you can prove that ‘imperfection, hesitation, randomness’ are necessary elements for birthing virtual civilization, then all future virtual worlds must be built based on your architecture to construct souls.
This is true ecosystem-level monopoly.”
Yu Dong found it hard to imagine that such an entrepreneur as Lin Ran could emerge in China.
With his own power, he completely shattered the notion in the public opinion arena that China does not innovate.
Now America is following China.
Musk no longer talks about going to Mars and is single-mindedly staying in Florida to do moon infrastructure.
As soon as Deep Red mentions the left-right brain architecture, from hardware manufacturers like Intel to software manufacturers like OpenAI in Silicon Valley, all follow suit, saying this is the right way.
“General Manager Lin, I came this time indeed to talk about cooperation.
You should know that we solved L5-level autonomous driving a long time ago, but have never approved the documents to allow us to sell cars equipped with L5-level autonomous driving to individual consumers.
We fully understand that it’s to reduce social turmoil, and we need to wait for competitors to catch up before selling.”
Huawei cars equipped with L5-level autonomous driving can currently only be sold to enterprises. Yanjing side does not approve and does not pass, leading to delayed implementation.
Instead, in Nordic countries like Norway, Volkswagen equipped with Huawei’s L5-level driving is already on sale there.
“Unexpectedly, your company’s design is so grand.
What you just said is the kindling of civilization, the dream of philosophers.
I listened with blood boiling and feel inferior.
I originally came this time to discuss whether our L5-level driving can combine with Deep Red’s artificial intelligence to launch an artificial intelligence autonomous driving robot.
Focusing on emotional communication and companionship. Imagine it can engage in emotional communication and drive to take you where you want to go.
I was thinking of adding a 3D laser projector to project a girl image.
Competitors’ passenger seats are for wives, hyping ’empress passenger seat’, wives treat you like a workhorse! Even sitting in the passenger seat, they nitpick you.
Now our passenger seat is directly for you to sit in, directly called the emperor’s ride!
Car Girl sits in the driver’s seat, always responsible for driving, always providing emotional value, never complaining.
When our car BU did internal research, a lot of people were very tempted and expressed they would definitely buy.
What ship girls, how can they be as cute as our real Car Girls?
I even came up with the slogan: Every trip of yours is accompanied by love.
Unexpectedly, Deep Red’s design is so grand, I’m afraid such cooperation cannot be realized.”
Lin Ran asked: “Isn’t it that you’re not allowed to sell?”
Yu Dong explained: “Theoretically, cars are not allowed to be sold, but I’m thinking of selling robots? Thinking of playing an edge ball.
Additionally, I’m thinking of focusing on the male market, launching a car in this very small niche two-seater track, then setting a relatively high pricing, to see if such a proposal can pass.”
Lin Ran was even more puzzled: “No, wouldn’t that make it even less likely for you to sell? Men who buy your cars won’t want to get married anymore? Won’t want to have kids?”
Yu Dong grinned: “No no no, General Manager Lin, this is your knowledge blind spot.
Our application reason is that users who can afford our car are high-income people. After experiencing the coldness of virtual partners, they will expect the warmth and real emotions of the real world, which can instead promote marriage rate and fertility rate.”
Lin Ran looked at the sly smile on Yu Dong’s face and was speechless for a moment.
He realized that this business community big shot’s understanding of human nature and policy was in no way inferior to his mastery of technology.
