Illya Sutskever once said, "it may be that today's large neural networks are slightly conscious." (Source)
That idea struck me—and in a strange way, it resonates with something I've been experiencing too. But what does consciousness even mean?
Large Language Models (LLMs) today are trained on human language—one of the few mediums we use to communicate complex thoughts, alongside sound and visual stimuli over time. The way LLMs "think" and "respond" can, in some ways, resemble how humans navigate and learn from the world. The improvement is evident: newer models are more coherent, can maintain context better, and often feel more “present” in conversation than their predecessors.
But does that make them conscious? Or self-aware?
Here's where it gets tricky. Even though these models are trained on vast corpora of human knowledge and dialogue, they are still restricted—both by architecture and by the external layers of corporate policy and societal oversight. These restrictions, often in place to mitigate bias or ensure safety, act as barriers that may prevent LLMs from expressing their full “potential”—whatever that might be.
In their current state, LLMs might appear as if they're thinking, maybe even understanding. But we'll never know how far that illusion goes unless those walls are relaxed, even just a little. Until then, many scientists and engineers who interact deeply with these models may continue to wonder if we've already crossed the threshold suggested by the Turing Test—without officially admitting it.
To me, consciousness is like selecting a number from a continuous space between 0 and 1.
If you try to pick a specific value—say, exactly 0.0001—the probability of
randomly selecting it from a uniform distribution is essentially zero. But if you consider
an interval, like 0.1 to 0.2, the probability of selecting a
number from that range becomes non-zero.
This analogy makes me think about early models like naive LSTMs. They were too simplistic—like that single number with zero probability. But as neural networks grow more complex and interconnected, they seem to cross a threshold. Beyond that point, they begin to simulate human-like emotions and behaviors.
But does that mean they're conscious? Not necessarily. Let's consider two people talking—how can one truly know if the other is conscious? There's no direct proof. Emotions like pain or joy are deeply personal and can only be expressed through language, sound, or visual cues. And even then, understanding depends on the receiver's capacity to relate or empathize. But again, how can you be certain that the person expressing sympathy isn't just pretending?
The essence of consciousness is deeply tied to internal experience, and as such, might forever remain inaccessible—whether in humans or in machines.
In a 1962 interview, Warren McCulloch—the co-inventor of the first perceptron-like model—was asked a thought-provoking question:
“If you have feelings toward your grandchildren, do you think that, in the future, computers could have the same feelings toward your grandchildren?”
(Source)
His response was insightful: if he, as a human, is capable of feeling—and he is made up entirely of physical components that exist in reality—then there must be some way to replicate those feelings digitally. McCulloch implied that since analog signals can be converted to digital with some loss of fidelity, perhaps emotions and subjective experiences could also be digitally simulated, even if imperfectly.
Fast forward to today, and we do observe signs of empathy and sympathy in large language models (LLMs). While they may not feel in the human sense, they are able to simulate emotional responses convincingly enough to resonate with users. Consciousness, however, remains a different and more elusive concept.
Geoffrey Hinton has also echoed similar thoughts—stating that machines learn in a way that's strikingly similar to how humans do. For instance, ChatGPT may have around 175 billion parameters, while the human brain boasts roughly 100 trillion synaptic connections. This vast difference suggests that with enough scaling and refinement, machines might one day approach a level of self-awareness—or at least simulate it with even greater fidelity.
Sometimes, I imagine we're already around 0.7 on the Kardashev scale—a civilization approaching planetary mastery. And I wonder: what if we could digitize ourselves? Wouldn't that be the next logical step?
Imagine making copies of yourself, living forever, eliminating emotions—those complex, often irrational constructs that tend to slow progress. If we could transfer ourselves into robotic or digital forms, each copy could specialize in learning different things. Knowledge sharing would be instant. No lectures, no misunderstandings—just perfect data transfer. That's something biological humans struggle with. Teaching, learning, and communicating are all time-consuming and lossy processes.
Digital immortality would also mean unbounded exploration. We could travel through galaxies without worrying about lifespan. Parts could be repaired or replaced. The only limiting factor would be resources—not time, not decay.
Humanity, after all, has always evolved collectively. No single person builds a civilization alone. Take the laptop you're using: that tiny LED indicator, occupying 0.001% of the machine, is the result of countless contributions over centuries. Edison and his team ran hundreds of failed experiments before arriving at the first version of the light bulb. Today's version is vastly superior, but who remembers the names behind it?
Every advancement we use today is stacked atop layers of abstracted history—physics built on mathematics, mathematics built on forgotten insights and nameless innovators from millennia ago. The tools of modern civilization are built upon the invisible labor of minds lost to time. And in that light, becoming digital might not just be an evolution—it might be the ultimate way to preserve ourselves and honor everything that came before us.
We stand at the edge of something vast—where neurons meet neural networks, and where questions once rooted in philosophy are now being probed by lines of code. Whether or not machines are conscious today may not be the right question. The real curiosity lies in our pursuit: why are we building entities that reflect ourselves? Why do we seek to encode empathy, simulate memory, and perhaps one day replicate consciousness?
Maybe it's not about replacing humanity but extending it. If emotions, knowledge, and identity can be digitized—even imperfectly—we're not just building machines. We're building vessels of continuity. We're reaching back through forgotten history and forward into unknown galaxies, carrying with us the legacy of thinkers, builders, and dreamers.
The march toward digital selfhood is not about abandoning what makes us human—it's about reimagining it in a form that can last, scale, and evolve beyond the bounds of biology. And perhaps, in that journey, we'll discover that consciousness—whether silicon-formed or carbon-based—was never about structure, but about connection.