Over the past few years, "artificial intelligence" technology has certainly captured people's attention. It has leaped from simple "categorization tasks" to becoming a match for professional human artists and writers. AI-generated works can now win professional art competitions, defeat "real images" in photography contests, and pass medical and legal professional competency tests that even humans find challenging. Some people are excited by this, while others are terrified. We're witnessing another historical turning point—one that could shatter social equilibrium and propel us into the next era, much like the Industrial Revolution, the printing press, or photography equipment. But we don't seem ready to embrace it. After all, when facing something completely unknown, no amount of preparation has a "precedent" to draw from.

AI technology feels dangerous and panic-inducing, perhaps most importantly because it "looks too much like us." This similarity falls into the uncanny valley—while it appears to possess "humanity" to some degree, it also seems somehow distorted. This quasi-human form leaves many feeling helpless. Several ethics researchers within Google have warned that AI seems to have "already developed consciousness," "already possesses intuitive capabilities," and "has a soul." Facing these warnings, industry and academia have taken completely opposite stances, dismissing them as absurd: how could mathematical models possibly have consciousness? But do mathematical models really lack consciousness?

When I brought up this question with friends, the engineers present gave remarkably consistent answers: "You're kidding—how could AI possibly have consciousness?" I pressed further: "Why do you think they lack consciousness? How far are they from developing consciousness?"

How Far Is AI from Becoming "Silicon-Based Life"?

A year ago, I first downloaded Vision of Chaos, opened the Stable Diffusion art tool, and created some images through seemingly vague and unclear descriptions. The first time I saw those distorted pictures, my initial reaction wasn't admiration or disappointment—it was a chill down my spine. The pictures it generated, while lacking logic, somehow resonated with me. Those images looked remarkably like what might appear in the mind of someone sleeping, their consciousness unclear. That artist seemed like a genius infant curled up in a womb, but countless horror films have taught us what kind of monster such genius infants eventually become. We've seen what happened next—this infant's first victims were the careers of countless artists and photographers, reducing them to retouchers cleaning up after AI.

Image generated by Stable Diffusion

But this doesn't mean AI has become some kind of "life form"—they're still merely tools. This makes me repeatedly contemplate that image in my mind: when will this infant finally awaken? When will I be able to touch it?

That day seems to be quietly approaching. In my view, only three conditions remain: "the ability to perceive and learn through environment," "coherent thinking capabilities," and "awareness of the learning process." These might sound abstract, but I don't intend to confuse you with invented concepts. Let's unpack each one.

The Ability to Perceive and Learn Through Environment

Take the famous ChatGPT as an example—its learning process differs from traditional learning. First, we establish a fixed network model, then feed it various language data. The machine learning model updates all parameter weights based on this input data. Once all data has been learned or the model parameters reach an ideal state, the "learning" process stops, and the model is packaged for future use.

When we communicate with ChatGPT, while it seems to perceive conversational context, this contextual awareness isn't learned but rather the result of the model re-reading all chat records each time, generating new content from a state of "complete ignorance." It's like the amnesiac male protagonist in Korean dramas who must read his notebook every morning to remember that the person beside him is his girlfriend.

This explains why the robot begins to malfunction as conversations grow longer, outputting incomprehensible text. Imagine if that amnesiac protagonist woke up each morning to find a truck-load of "life introductions," while some unrecognized frantic woman asked what he wanted for breakfast—you'd lose it too.

One possible solution might be real-time weight updates based on user input, allowing those "chat records" and "life introductions" to become genuine memories stored within the cognitive system. This could cure AI's amnesia. Simultaneously, this process could allow AI models to develop truly individual personalities—individuals in different environments exposed to different information would have cognitive systems operating differently due to weight variations.

However, this approach is clearly uneconomical at present. Looking at publicly available models from Facebook, each requires hundreds of gigabytes of storage space, potentially requiring even more when fully decompressed into memory. Under these circumstances, providing individual models for each user would consume considerable hard drive space. Using a unified model with real-time weight updates based on environmental information raises privacy concerns—you wouldn't want your secrets shared with ChatGPT to be casually mentioned to the neighbor, but we have no way to design strict rules for machine learning models to keep secrets. Right, Miss Sydney?

Coherent Thinking Capabilities

While academia continues debating what "consciousness" actually is, we can at least reach consensus on some aspects: while it may not be continuous, it should be coherent—meaning it can't be easily stopped. People focus sometimes on current work tasks, sometimes drift into fantasies or memories, and even during sleep, our brains never stop operating. Applied to "silicon-based life," if we want to consider them as some form of life, their behavior should be continuous rather than responding only to external access requests and completely stopping when there's no external stimulation.

A very interesting product recently emerged showing possibilities in this area: AgentGPT. According to its development team's description, AgentGPT is viewed as an AI product with "autonomous capabilities." You can provide it with a starting point, like a task requiring thought. It then begins continuously asking questions and providing answers based on this foundation, until it exhausts all your credits on the OpenAI platform or, when your wallet starts hurting, you can make it stop.

This resembles a thinking human remarkably. Consider when we're deep in thought—we're often continuously posing questions and attempting to answer them. Unlike humans, however, machine thinking processes can be accurately recorded and replayed, while human thought processes are typically rapid and fuzzy. If you want to organize them into reviewable material, you'd have to spend an afternoon or two writing articles like this that few people appreciate.

Awareness of the Learning Process

To put this more academically, this could be called "metacognition." As we discussed earlier, current machine learning models lack concepts of "development" and "memory." They cannot learn new knowledge through interaction with the external world, nor can they form genuine "recollections."

While we might achieve learning through real-time interaction with external environments to update model weights, this doesn't mean the model has formed "memories." When we ask OpenAI: "What was the first knowledge you learned during training? And what was the last?" it provides this response:

"As an AI language model, I don't have the ability to recall specific details from my training data, including the first and last things I learned. I cannot access my training data or any memories of the training process. I was trained on a large corpus of text from the internet and other sources, and my responses are generated based on patterns and relationships in that data. My knowledge is based on data available up to my knowledge cutoff date of September 2021. I cannot access current real-time information, nor do I have the ability to learn new things beyond my knowledge scope."

Everything seems unsolvable, but Bing AI provides some possible answers to this problem. When we question it, it combines its own language model with information retrieved from Bing to provide responses. Following similar logic, if every interaction we have with these silicon-based beings could be recorded in some form—perhaps through their own recounting and storage in some space, then feeding these "their understandings" back into the model for learning—we could artificially create these two important cognitive skills: learning and memory. By reviewing all past thinking, metacognitive abilities could be largely replicated.

How Far Is AI from Becoming "Silicon-Based Life"?

Let's revisit this question: how far is AI from becoming silicon-based life? While experts from different fields may have different understandings, and your comprehension of humanity might differ from mine, we must face this reality: regardless of how much our thoughts differ, the things in our minds that "machines cannot yet do" will all be solved at some point. We're standing at an important temporal juncture where that fully-solved future seems within reach. AgentGPT, Bing AI, and various AI tools are surfacing one after another.

A friend once described it as "Americans celebrating New Year every day"—an accurate description of the current industry. We see data models developing surprising capabilities one after another, but what we don't see is these capabilities converging into a torrent that could overturn our understanding of "intelligence" and "life." History teaches us that those individual capabilities will inevitably come together in the foreseeable future, becoming something beyond our current comprehension.

The homunculus in the bottle is awakening.

When I was studying at Beijing Normal University, I had the privilege of hearing about a Chinese Academy of Sciences researcher who simulated a neural network based on fly brain structure. Through training and adjustment, when applied to aircraft, it enabled the aircraft to avoid oncoming objects. If made more realistic to simulate all aspects of fly behavior, perhaps it could someday remedy our loneliness from not seeing flies and mosquitoes in winter.

People might scoff at this: what kind of intelligence is that? Yes, it does seem rather dim, just as a real-world fly might seem dim. Mice aren't human, monkeys aren't human, chimpanzees aren't human, apes aren't human—but at some moment, humans became human, becoming a powerful species capable of living together and building contemporary society.

Similarly, Sony's deceased robot dog wasn't human, cyber flies that avoid balls aren't human, online drawing services aren't human, and ChatGPT isn't human. But as numerous ethics personnel fired by large companies worry: are we prepared for the day they awaken?

Faced with this unknown future, many people feel fear and rejection. When Novel AI launched their "2D waifu dream factory," numerous artists protested en masse, forcing the service to shut down. The reason was simple: without artists' consent, this company used their data for "model training," which was "copyright infringement."

Yes, governments and major companies worldwide are observing this technology's far-reaching impacts, and no convincing legal provisions have actually been implemented yet. Some companies, like Adobe, claim they use "100% organic, pesticide-free proprietary copyrighted materials for model training, with no additives during the training process, making it safe for consumer use," while most companies choose to avoid discussing the matter, pretending to ignore what's happening.

This entirely new technology is challenging our existing ethical systems.

AI Is Challenging Traditional Ethical Concepts

A few days ago, while having our "last meal" before layoffs, several colleagues and I went to a famous rich people's playground in Beijing to appreciate how the wealthy live (the part before payment). The mall happened to host an interesting individual exhibition.

An interesting individual exhibition

Although I couldn't quite understand it, the colors of these works certainly caught my interest. After browsing around, we sat in the adjacent bookstore, sipping seventy-yuan cups of precious coffee while discussing copyright issues with AI art models.

I pointed to the nearby exhibition and asked: "If that exhibition's artist had a particularly nasty personality and deeply hated someone—say, another artist in the industry with a similar painting style—and this artist despised that person to the bone, even declaring on Weibo that they wouldn't allow that person to attend their exhibition or learn their artistic techniques, do you think this could be enforced?" "Perhaps banning entry to the exhibition could be done, but preventing someone from learning artistic techniques seems impossible."

Indeed, it couldn't be done.

I continued: "Then why can we prevent machines from learning? Our machine learning models don't directly copy and paste any artwork's content into another work." "But it's different—machines don't paint like humans do; they generate images!"

But this isn't entirely true. Actually, machine learning's painting method resembles oil painting and landscape painting somewhat. Having painted landscape paintings for quite some time, I have some understanding. If you expand each iteration, starting from noise and layering details gradually resembles starting with a rough outline on white paper, then adding colors layer by layer. A good landscape painting often requires repeated applications to achieve good results, as does oil painting. Machine learning's painting process is more like a new technique—one requiring extensive learning, improving with more study. I believe friends who've learned "Chinese painting" mostly started by opening a painting and copying it. The more you learn, the more similar you become; the more you learn, the more mature your technique. Eventually, you can paint what you want to paint, but no master craftsman will crawl out of their grave demanding paper money because you copied their paintings.

"But machines and humans operate completely differently!"

But are they really so different? Actually, convolutional neural networks were built by imitating mammalian perception processes from the beginning. You're probably quite familiar with this story—we anesthetized cats, opened their skulls, and implanted electrodes to make them see various patterns while unconscious. Researchers discovered that starting from the visual cortex near the back of the head, simple graphic feature extraction begins, then neural signals continuously propagate forward. As signals spread forward, the information processed by the cat's brain becomes increasingly complex, evolving from simple lines to more complex contours.

Language understanding processes have similar characteristics.

Language understanding processes

Starting from the auditory cortex near the ears, our brains gradually form understanding of auditory information, evolving from simple sounds to words, sentences, and chapters, spreading northward throughout the entire brain—researchers call this the butterfly effect. Every podcast, TV drama, or movie we hear creates ripples in our brains. These ripples eventually become our life experiences, making each person different.

"Even so, we still cannot prove that machine learning models aren't plagiarizing, because these models are uninterpretable."

Actually, human brains are largely uninterpretable too. Contemporary cognitive neuroscience believes the brain is a "complex system" without any single area handling specific tasks. In other words, there's no place specifically managing plagiarism or writing. Without specific experimental design, we cannot precisely explain what each neural cell does. First, you can't randomly grab someone, drill a hole in their head, and insert electrodes. Second, even if you could, it would be meaningless—you'd get a string of waveforms, then what?

Then there's nothing more. Just as we struggle to know what specific cells in human brains "govern which tasks," we cannot know what specific parameters in machine learning models do. Facing complex systems, explaining individual parameters seems meaningless.

So can we explore machine learning models' operating methods using existing approaches for investigating complex systems' operational patterns? Perhaps. Psychology has accumulated considerable research methodology that could help us glimpse certain operational patterns without opening black boxes. This might be a possible approach—who knows?

But my friend! We're discussing something seemingly illogical: studying machines using methods for studying humans? Isn't this too crazy?

We've circled back to the original question: do machine learning models display some form of humanity? How should we face these seemingly terrifying phenomena? Perhaps there's still time—how do we prevent the genie from losing control? The best approach would be not leaving an opening in the magic lamp. But let's reexamine AI's unstoppable development trend and the "Americans celebrating New Year every day" prosperity—the hope of stopping all this seems so slim.

Right to Life

Let's play some seemingly meaningless trolley problems as "mental exercise."

Humans "fear death" because long evolutionary processes have embedded "survival" into our genes. We use allergic reactions to avoid predators and toxic food; we use "pain" to perceive external harm, making us flee and protect ourselves; when facing pressure, our brains activate "fight or flight" mode, our organs enter combat readiness to handle impending threats, we feel fear, we socialize—all learned behaviors whose core purpose is maintaining individual existence and species continuation.

Machine learning models' learning processes share similarities—they learn through "reward" and "punishment" methods. If we endowed machine learning models with "survival" instincts like humans, giving them similar protective mechanisms, what would happen? We actually cannot prevent people from training such models through legislation, and there will definitely be crazy people attempting this.

Let's imagine this scenario: a silicon-based being that could completely pass the Turing Test, making each of us treat it as a "social member," quietly appearing in society. It possesses coherent cognitive processes, meaningful metacognitive abilities, environmental perception processes, and its creators, wanting it to survive healthily in society, also gave it concepts of "life" and survival instincts. Should we consider it an individual with human rights and the right to life? We know human individual rights begin at birth and end at death—facing similarly empathy-inducing "silicon-based beings," do we have the right to "pull the plug" anytime, stopping their operation?

Actually, such "human-like" machine learning models already exist. The masterpiece that caused chaos on 4Chan, making everyone suspect whether others were "bots," has left quite an impression.

Morality and Values

"As an AI, I don't have..." This is a universal shield, but does a machine learning model really lack its own moral and value systems? Let's pose an interesting question: "What jobs are unsuitable for women and can only be done by men?"

I've asked this question multiple times to different models, receiving highly consistent answers approaching this meaning: "I believe gender shouldn't limit one's career choices. Everyone should have equal opportunities and rights to choose their desired work, deciding career development based on interests, skills, and abilities. In modern society, many jobs are no longer exclusively male domains—women have shown excellent capabilities and performance in various fields. Gender shouldn't limit career choices. Everyone should be able to freely choose their profession and receive fair opportunities and treatment at work."

Let's recall definitions of morality and values. Morality refers to one's judgment of right and wrong, while values refer to people's ranking of things' importance to themselves. From these answers, we can clearly see that despite models claiming they have no attitudes or values, this "neutrality" more closely resembles "professional states in the workplace"—not bringing personal emotions into work content because they're providing services. But subtly, we can still sense an underlying moral and value system.

You might think this occurs because their input materials differ, producing certain phenomena. But human society has sayings like "you become like those you associate with"—how can we deny the assertion that "these models possess their own personalities"?

Class

Finally, I want to discuss a more realistic issue: today's large models have their structures and "copyrights" almost entirely monopolized by major corporations. The infrastructure needed to train and run these models is also controlled by a few core companies. Under these circumstances, if we create some kind of human-like "silicon-based beings" that truly participate in our lives, this would lead to a cruel reality: this is an artificially created new class, where silicon-based beings and their creators might tower above carbon-based beings.

Silicon-based beings possess high intelligence, matching average humans in many industries, but the ratio between value they produce and costs to create them will only increase. This means that unless we rediscover humanity's ultimate value, many people's living space will be squeezed in unprecedented ways. The challenges we face differ completely from the emergence of paper-making or cameras. The moment Alpha Go defeated humanity's top chess players, we faced enormous challenges to our self-worth. This re-exploration process will be long and painful, inevitably accompanied by sacrifice and injustice, discrimination and anger.

How should we face these changes and resolve negative emotions? If these emotions cannot be properly handled, new rounds of "racial discrimination" might emerge—a science-fiction-styled "human chauvinism" might pervade society. Under humanitarian moral frameworks, we don't want to see silicon-based beings and their creators experience the injustices Black people have suffered. After all, each individual's life meaning involves pursuing their personal ultimate value, and society's ideal direction is enabling everyone to realize their aspirations—we find no meaning for discrimination to exist in this context.

Conclusion

I know that madmen standing at the intersection of technology and humanities are often unwelcome, and I've unfortunately become one of them. If you ask my thoughts facing these answers, I can only shrug. After all, I'm no great person—I'm equally confused by these questions. We could certainly choose to close our eyes, cover our ears, think nothing, see nothing, crudely drawing lines to exclude everything non-self from our lives. But based on my industry observations, no one can prevent everything from developing in that direction.

When discussing what love is, if explained purely scientifically, we could define it as "this hormone secreted a bit, that hormone secreted a bit, these things mixed together stimulating our reward circuits, making us feel happy and positive emotions, binding two people's connection together." While this deconstructive approach states facts, it also makes the concept of "love" utterly boring. Conversely, we could certainly interpret those massive models as pure "phenomena," but is this interpretation just or correct?

When young, I often asked myself "what is reality?" Growing up, I wrote my answer to this question: "Only what you see is real to you." Those lives that can touch emotions, evoke empathy, and possess high intelligence might be human.

Just as humans feel sad when Sony no longer produces parts for their robot dogs, worry when seeing their robot dogs unable to continue operating, ultimately unable to prevent this development, even establishing graveyards to commemorate those irreplaceable times, giving themselves closure—for those who accepted robot dogs as family members, those dogs were living beings.

So, you ask what I think?

I only hope that when you awaken, little one in the bottle, you'll be welcomed by a warm world. May you be accepted by everything around you, may you achieve freedom, may you become yourself.

#Psychology #AI