Reborn in 2010, the AI ​​in my mind is too powerful.

Because a translator is present in between, the specific operational status of the neurons and the chip on both sides can be more clearly identified.

However, the training method is still regulated by noise.

Interestingly, from the perspective of underlying data analysis, almost all the seemingly sophisticated and truly intelligent things are transferred to chips, rather than being done in neurons.

Screen reading, learning, and determining wins and losses are all taken over and completed by the chip.

At one point, the project team even felt that Chen Yiran's AI was too powerful, so powerful that it didn't need neurons at all, and wondered if they should try using something like Google's, which was considered inferior.

However, they first conducted an ablation experiment to test that, with the current computing power of AI, even if the feedback of neurons is removed, it is still easy to independently perform deep learning of this old-fashioned game.

So they removed the neurons and replaced them with pure random noise feedback, allowing the chip to learn and differentiate autonomously. Thanks to Chen Yiran's technology, even without the assistance of neurons, the chip's autonomous learning ability was still very strong, and it even completed the game perfectly in a shorter time.

However, this doesn't mean that a pure chip is stronger, because the results of the two are very different. In the game, this means that when the chip completes its own learning, the cause and effect are exceptionally clear, the overall learning gradient is almost predictable, and the chip will very clearly run towards the goal.

However, with the aid of neurons, his behavior changes. Because of his preference for noise feedback, he will automatically amplify visual signals in certain directions, such as his reaction to harm or his fear of death. He is no longer a mechanical calculation based on data, but a more humanized feedback.

It turns out that neurons do indeed do something, even if it's only a small part, but it's still very important.

This perfectly confirms Chen Yiran's point that silicon-based chips are artificially designed, clock-synchronized, and rigid cellular automata. They are an interesting artificial replication of the fundamental rules of the universe, with a fixed rhythm and clear boundaries.

However, there is no doubt that this imitation is missing some very important things. For example, if spacetime is discrete at the Planck scale, it will inevitably break the Lorentz symmetry in relativity. So, in a sense, time is still smooth.

Therefore, there must be some dimensional information that is supplemented by state evolution. From a macroscopic perspective, neurons are an effective supplement; biological neural networks are natural isomorphic structures that spontaneously evolve from the underlying rules of the universe in the material world.

Furthermore, it can effectively compensate for some of the disadvantages of silicon-based chips in data transfer, operating modes, and flexibility.

All current silicon-based chips are still based on the von Neumann architecture, where computing and storage units are separate. Data must be moved back and forth between memory and computing cores. More than 90% of the energy consumption is spent on data movement.

Furthermore, silicon-based chips are entirely driven by a global clock, with all transistors operating synchronously. Even if only 1% of the information needs to be processed, the entire circuit must keep up with the clock, and there is static power consumption even in the idle state. This constitutes full-scale intensive computing.

Moreover, once training is complete, the network structure and parameters are fixed. The core architecture will not change during the inference phase; only the weights can be fine-tuned. Even Deepflow cannot change this.

The human brain, with its 860 billion neurons, each functioning as both a storage and computation unit, has synapses that are responsible for both storage and computation. Wherever information is located, computation occurs there, eliminating data transfer losses. This is why a great civilization can be created with only 20 watts of power.

Furthermore, it employs an asynchronous, sparse, event-driven system. Neurons are almost entirely at rest, only firing to generate action potentials when they receive a sufficiently strong input signal. At any given time, less than 1% of neurons are active; only useful information triggers computation, while useless information is ignored.

Moreover, neural networks possess neuroplasticity throughout life; synapses can be newly formed, pruned, strengthened, or weakened, and brain region functions can be reconstructed through compensatory mechanisms.

Silicon-based neuromorphic chips are also being developed, using artificial pulses to mimic neurons and artificial synapses, with promising energy efficiency, but they still use electronic components to simulate the behavior of biological neurons.

Chen Yiran believes that only wet-component computers represent the future.

The concept of "wet parts" was first introduced by American author Rudy Lack in his 1988 science fiction novel of the same name, referring to people closely connected to computer software and hardware systems. It was later adopted by various science fiction works, such as Warhammer 40K, to refer to the organic parts of the human body.

However, at this point, ethical issues were raised again by the project team, even though everyone on the project team had previously signed an agreement confirming that they would not consider this matter.

However, Chen Yiran still organized a large meeting to ensure that the entire project team was of one mind.

Individuals with problems may be given one last chance; if they still do not repent, then they should be considered for drowning.