Chapter 225 - 216: So You’re Just a Daydreaming Kid, Too

Chapter 225: Chapter 216: So You’re Just a Daydreaming Kid, Too

Li Dong politely put down his chopsticks.

"President Qiu, hello. I’m Teacher Gao’s student, Li Dong."

Qiu Mingli was taken aback for a moment.

Then, she said with a hint of pleasant surprise.

"So you’re Li Dong?"

"The Li Dong from the dimensional reduction algorithm?"

"My goodness, you’re so young."

A few others at the table who hadn’t made the connection before reacted upon hearing the words "dimensional reduction algorithm," and their gazes all turned to Li Dong.

Li Dong felt a little uncomfortable being stared at by so many pairs of eyes.

"You’re too kind, President Qiu. I’m just a university student."

"I was just messing around."

President Li laughed heartily beside him.

"Messing around? If what you did was just messing around, then all of us have been working for nothing."

A wave of good-natured laughter went around the table.

After the laughter died down, Qiu Mingli was the first to speak.

"Li Dong, I’d really like to hear your thoughts."

"You heard what we were just discussing—computing power, data... all that stuff."

"So, do you have any ideas on how we can catch up?"

To be honest, everyone present genuinely wanted to hear what he had to say. After all, they didn’t know where Li Dong’s algorithm had ended up, and that lack of knowledge was telling in itself.

That’s why they were particularly curious about Li Dong.

Li Dong was silent for a few seconds.

Truthfully, he didn’t really have a fully formed idea he could present.

He was just...

He was just thinking of Newton’s Xiaohai.

That thing, no matter how you looked at it, didn’t seem like it was built using the current large language model approach.

He chose his words carefully and began to speak slowly.

"I don’t have any highly professional opinions."

"I just have a layman’s question, so please don’t laugh."

The people at the table all smiled and said they wouldn’t.

Only then did Li Dong continue.

"Why... does artificial intelligence have to be developed along its current path?"

"What I mean is, why can’t it more closely resemble the way the human brain actually works?"

"Look, the human brain has over eighty billion neurons and consumes only about twenty watts of power per day."

"But it can hear, see, think, learn, recognize a face in under a second, and even learn a language within two years."

"And yet our most advanced large models today..."

"...burn tens of millions of kilowatt-hours of electricity for a single training run, and a single inference requires mobilizing an entire server room of cards."

"And when it comes to certain tasks, they’re still not as good as a three-year-old child."

"Where is the gap?"

"I don’t think it’s in computing power, but in... the fundamental approach."

To be honest, Li Dong didn’t really understand the field of AI that well, so he was just voicing his own thoughts.

No one at the table made a sound.

They just listened quietly.

Li Dong paused for a moment before continuing.

"While I was looking through some literature a while back, I came across a field called spiking neural networks."

"The idea is that instead of using the current continuous-value activation functions, it mimics the way neurons fire in the human brain."

"There are only two states: firing and not firing. It’s idle most of the time, only sending out a pulse when a signal accumulates to a certain threshold."

"Theoretically, the energy efficiency of this type of network should be orders of magnitude higher than current ANNs."

"There’s an even more radical approach called neuromorphic computing, which goes so far as to build the chip itself to mimic the structure of neurons."

"IBM’s TrueNorth and Intel’s Loihi are both attempts along this path."

"I was just thinking..."

"Could it be that one day, we could actually build something that can accomplish the same tasks not with ten thousand A100s, but with a single brain-like chip?"

When Li Dong finished speaking, the table fell silent.

It wasn’t that everyone was stunned, but rather that a single thought crossed their minds:

’Oh, so he’s just another kid who daydreams.’

President Li was the first to react.

The smile on his face was a bit complicated, but you could tell he genuinely wanted to share some insights with Li Dong.

"Li Dong."

"Your idea, it’s actually not new in academic circles."

"Not only is it not new, you could even say it’s very old."

"It was proposed as far back as the 1980s."

Li Dong nodded.

President Li continued.

"But why hasn’t this path taken off after all these years?"

"Because it’s just too difficult."

He put down his glass and spread his hands.

"Our understanding of the human brain itself is nowhere near as extensive as laypeople imagine."

"To date, the only connectome in the entire world that has been fully mapped..."

"...belongs to the nematode C. elegans."

"It has only three hundred and two neurons in total."

"Just mapping the connections between those three hundred and two neurons took scientists a full thirty years."

"A portion of the fruit fly’s connectome was just completed a couple of years ago. A mouse has about seventy million neurons in its brain, and we haven’t even mapped a fraction of them yet."

"The human brain? Eighty billion neurons, with each neuron connected to an average of ten thousand synapses."

"We are a million miles away from fully understanding it."

President Li looked at Li Dong and said.

"Even if we did understand the brain, we have no way to efficiently replicate it on silicon-based chips."

"The neurons in the human brain and our chips..."

"...operate on two completely different systems of logic."

"If you force a spiking neural network to be simulated on a GPU, its efficiency is even lower than just running an ANN directly."

"This is why, despite all the hype about SNNs for so many years, no mainstream AI company has truly gone all-in on them."

"The scale of tasks it can currently handle is far inferior to that of ANNs."

President Li sighed and continued.

"Also, we don’t have a mature training method."

"Why have ANNs made such rapid progress in the last decade? Because we have backpropagation."

"Backpropagation is mathematically very efficient."

"But does the human brain use backpropagation? No."

"So how do you train an SNN? Currently, the academic world has no universally accepted method that works as well as backpropagation."

"Many teams are trying, but none have succeeded yet."

President Li lowered his hands and picked up his glass.

"So, Li Dong."

"Theoretically, of course, your direction is correct."

"In the long run, neuromorphic computing is very likely the future."

"But..."

He emphasized his point.

"Within the foreseeable future—five, ten years—the only path that can be implemented, be profitable, and allow China’s AI to catch up to the top tier in the world is the current one: the transformer-based large model approach."

"Not because it’s the best."

"But because it’s the most realistic."

After President Li finished, the others at the table all nodded in agreement.

Clearly, this was the established consensus within the industry.

Li Dong didn’t say a word.

But a voice in his heart was firmly refuting it.

’Impossible?’

’I’ve "seen" someone build it with my own eyes.’

’That was Newton’s Xiaohai.’

’Since someone was able to build it back then...’

’...it proves that this path is viable.’

’As for the current difficulties...’

’Understanding the brain, the fundamental logic of the hardware, the training methods...’

’They can be overcome, one step at a time.’