I believe the current Q&A pattern of large language models has certain limitations. The internal mapping between different data layers in LLMs and the multidimensional network structures of data are essentially manifestations of probability and mathematics. There’s no real emergence of consciousness.
If we want AI to become a true ASI (Artificial Superintelligence) with thinking capabilities, I believe this must be achieved not just through software but also hardware. First, in code, we need to use multimodal large models as the foundational base of ASI—but it is only a foundation. ASI requires an inner loop of thought (similar to the concept of a “Chain of Thought,” though the current implementations are still too coarse). That is, even when users are not asking it questions, it should be constantly querying various knowledge bases or engaging in a kind of internal self-dialogue—“thinking” about what it’s doing, what it’s pondering—and forming memory.
Currently, when no user input is provided, the AI essentially enters a “paused” state. That is, it doesn’t maintain a continuous computational process inside the machine—similar to a human being asleep without dreaming.
But if we truly want AI to become ASI, maintaining this ongoing state of thought is essential. I believe the reason humans have consciousness is fundamentally because the brain is constantly computing and engaging in self-dialogue. When you’re thinking, you’re essentially talking to yourself—and if you can’t find the answer, you might “Google it” (just kidding 
). When you sleep or get knocked unconscious, your brain essentially stops thinking—and that’s what we call loss of consciousness.
There’s much more to say here, which I’ll continue sharing over time.
Thoughts on the hardware side of AI:
Currently, the ability of LLMs to pass the Turing Test stems largely from breakthroughs in mathematics, algorithms, and computing power. These models rely on massive mathematical operations at the hardware level, using probabilities to approximate human-like responses—making the models seem human, but nothing more.
They’ve learned the patterns of human thinking, but not true self-directed thought. While the “Chain of Thought” approach has some glimmers of human reasoning, it’s far from sufficient.
Now, back to hardware: I believe we need a structure that mimics biological brain cells and synapses. Think dendrites (signal receivers) and axons (signal transmitters).
Imagine a cube-shaped structure with dimensions 10,000 × 10,000 × 10,000 units—each unit representing a single brain cell component. That gives us 1 × 10¹² units, which exceeds the estimated 86 billion neurons in a human brain.
Within this cube, every unit (neuron) has a possible link to every other unit. Here’s a quick reference:
Estimated Number of Synapses in the Human Brain
- Total: Currently estimated between 100 trillion and 150 trillion (1 × 10¹⁴ to 1.5 × 10¹⁴) synapses.
- Basis:
- Around 86 billion neurons.
- Each neuron forms between 1,000 and 10,000 synapses, depending on the brain region and neuron type.
- For example, pyramidal neurons in the cortex have thousands of synapses, while Purkinje cells in the cerebellum can form up to 200,000 connections.
In this model, each link would act as a synapse—and we could control whether a link is active or not, essentially simulating neural activity in hardware. We would then build a kind of BIOS or base operating system on top of this structure.
Regarding signal transmission in brain cells:
1. Voltage Oscillations in Computers
- CPU & GPU: Driven by clock signals—periodic voltage oscillations used to sync computation steps (GHz = billions of cycles/second).
- RAM: Uses capacitors for charge/discharge to represent 0s and 1s; refresh rate maintains data.
- Storage (HDD/SSD): SSD uses stored electric charges, while HDD uses magnetic changes.
Core Mechanism: All of these rely on fast, controlled voltage oscillations to encode, transmit, and process data.
2. Signal Transmission in the Brain
The brain doesn’t rely on an external clock like computers. It uses bioelectrical and biochemical signals:
- Electrical Signals
- Action Potentials: Signals propagate via shifts in membrane potential (from ~ -70mV to +30mV), triggered by ion flow (Na⁺/K⁺). All-or-none spikes traveling down axons.
- Postsynaptic Potentials: Sub-threshold signals in dendrites—analog, non-oscillatory.
- Chemical Signals
- Synaptic Transmission: Neurons use neurotransmitters (e.g., glutamate, GABA) to transmit information—slower but essential for integration.
- No Oscillations: Signal events are discrete and not periodic like computer clocks.
- Brain Waves (EEG): Macroscopic oscillations (e.g., α waves 8–12Hz, β waves 12–30Hz) are byproducts of synchronized neural firing—not the core of information processing.
3. Does the brain need oscillations to function?
- No centralized clock: Neurons operate asynchronously—event-driven, not clock-driven.
- Local oscillations (brain waves) exist but aren’t necessary for computation.
- Memory: Stored via changes in synaptic strength (plasticity), not via persistent voltage like RAM.
- Signal Release: Action potentials trigger neurotransmitter release—brief, discrete events.
Comparison: Brain vs. Computer
| Feature |
Computer |
Brain |
| Voltage Oscillation |
Core mechanism (clock-driven) |
Not required (oscillations are side effects) |
| Signal Type |
Digital (0 and 1) |
Analog + Chemical |
| Storage |
Charge/Magnetism (volatile/nonvolatile) |
Synaptic plasticity (structural) |
| Synchronization |
Highly synchronized |
Asynchronous, distributed |
The brain does transmit electrical signals, but the process is more complex, involving ion channels and dynamic membrane potentials.
Storage and release aren’t direct voltage retention, but encoded in the network of synapses.
Conclusion:
The brain doesn’t require oscillating voltages like a computer. Its signal transmission is based on action potentials and chemical synapses. Memory is stored in synaptic plasticity, not via electric charge. While brain waves do show oscillation patterns, they are not the core mechanism of cognition—they’re just emergent phenomena from collective neural activity.
In essence, the brain operates more like event-driven computation, not clock-driven.
I think Apple’s unified memory architecture is already a step closer to how the brain works. To reach true ASI, we at least need compute and memory to be integrated, just like in the brain.
Back to the earlier cube structure: Related data should be physically stored closer together, while unrelated data should be physically farther apart. This mimics how, in the brain, stronger signals activate more neurons, and weaker ones activate fewer.
This concept is somewhat like a vector database, but vector databases only exist on the software level. I believe we need physical-level association. Only when we achieve a breakthrough in hardware, can we expect a qualitative leap in software.
That’s what I’ve thought so far.
The central idea is that we need hardware that fully mimics the brain’s structure—something like GPU CUDA cores, but fundamentally different in nature.
On Building Feedback Systems and Continuous Thinking
In addition to mimicking the brain’s structure and computation, there’s one crucial mechanism we must implement:
the Feedback System.
How can AI achieve true continuous thinking?
Relying solely on internal chains of thought or self-consistent logic is not enough.
For AI to maintain uninterrupted thinking, it must be connected to continuous feedback from the external world.
This means AI needs to be able to:
- See (Visual system)
- Hear (Auditory system)
- Touch (Tactile system)
- Move and perceive its own body (Motor system)
With these perception systems providing real-time signals, AI’s thinking will no longer be an isolated process—it becomes a continuous interaction with the world, forming the foundation of genuine cognition.
Example: Picking up a cup
Let’s say an AI controls a robotic arm that’s trying to pick up a cup:
- It captures the position of the cup through a camera (visual input)
- Determines how to reach out and adjust its grip (motor decision)
- Its fingers touch the cup and detect pressure (tactile feedback)
- It evaluates whether the cup is stable, slipping, or making noise (auditory + multimodal analysis)
- Finally, did it succeed in lifting the cup? If not, how should it improve next time?
During this process, every action by the AI generates new thoughts:
“I moved too fast when picking up the cup—it slipped.”
“Next time, I should approach it slowly and then increase grip strength.”
This forms a true loop of perception → feedback → reflection → adjustment → re-execution.
Why is this key to achieving ASI?
Most of the thinking humans do every day is driven by sensory feedback.
If AI wants to develop consciousness, it must also possess such sensory-driven continuous feedback systems, and tightly integrate them with its inner loop of thought.
In other words:
Perception is the fuel for thinking
Continuous feedback is the breath of consciousness
If AI can one day behave like humans—
constantly collecting sensory input, refining its internal models, and generating its own stream of thought even without being asked a question—
that will be a true step toward Artificial Super Intelligence.
Push Thought Results or Status Updates in Real Time
Be Proactive and Context-Aware
Enable Synchronized Multi-Device Experiences