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

1 p.m., Mountain View Computer History Museum.

There were over 500 people in the exhibition hall. Four cameras were set up at the back of the last row, with two CNBC cameras occupying the best positions. YouTube's live stream was being pushed out by the side door, and an engineer wearing headphones was constantly watching the bitrate readings on the screen.

Go forth, brave youth! A beautiful angel is calling you from afar, brave youth, go and create miracles!

Wearing a white sweater and blue jeans, Chen Yiran peeked at the audience from backstage and then instructed Jasmine, "I'm not 14 anymore! Even if we ever create an EVA, I won't be able to pilot it! I'm going up now, so start prompting me."

[As you wish, Your Excellency.]

Chen Yiran walked onto the stage from the right and stood in the center. She felt pretty good and didn't feel the nervousness she had imagined.

He started by joking with the audience, explaining that while 400 people were expected, 500 showed up, and many companies and bigwigs had come to ask for invitations, which he simply couldn't refuse.

"To be honest, I didn't expect that after only one year in Silicon Valley, I would have such influence. Microsoft and Oracle weren't on my invitation list, mainly because we haven't been discussing business for a while, so I forgot about you guys."

"Thankfully, you remembered me and proactively asked me for the invitation. And Mr. Musk even used a Tesla in exchange for the invitation—that's very generous of him."

Chen Yiran made no attempt to hide anything and directly extended an invitation to Musk: "I actually have a very high opinion of SpaceX. If the project has any funding gaps, you can come to me."

After all, as everyone knows, the media has described me as the new generation's walking money-printing machine. Actually, I'm not too happy with that title; a money-printing machine probably can't be as efficient as me.

At this moment, Musk did not yet possess the unbridled confidence of a top entrepreneur, but rather the cautiousness characteristic of a startup boss, nodding and clapping from the audience.

Upon seeing this, Chen Yiran sighed inwardly. As expected, it was the same old Ma who hadn't evolved, the shy internet addict who could be ridiculed by the CEO of Lenovo.

After the usual humorous opening remarks of the press conference, Chen Yiran got straight to the point, pressed a controller, and a slide appeared on the big screen.

"Human language is essentially an ordered arrangement of symbols. The meaning of each word depends on the words around it. 'I ate an apple yesterday' and 'The apple ate me yesterday' are the same words, but in different orders, they have completely different meanings."

He pressed a button on the remote control, and a traditional language processing pipeline diagram appeared on the large screen. A sentence, from left to right, went through word segmentation, part-of-speech tagging, syntactic analysis, and semantic understanding.

"In the past, computers processed languages ​​like this. It was sequential, word by word, with each word only seeing what preceded it. If the word segmentation was wrong, the subsequent syntactic analysis would all be wrong."

Chen Yiran continued her presentation after switching to a different slide.

"In a sentence, related words are often not adjacent. 'Yesterday I watched a very boring movie with my friend in San Francisco.' In this sentence, the relationship between 'movie' and 'boring' is much more important than the relationship between 'movie' and 'of.' In traditional serial processing, these connections all depend on the physical distance between words. Words that are far apart have weaker connections."

Let me be more specific. Traditional language processing has a problem called long-distance dependency. If a sentence has more than twenty words, it's difficult to establish a valid association between the first word and the last word.

Because information attenuates as it's passed down through layers of communication. It's like passing a message on to the next person; there won't be a problem, but after twenty people, the meaning could easily be misinterpreted.

Another image appeared on the large screen. This time, it wasn't a traditional assembly line; it was a completely different structure.

The image on the screen began to move; a sentence was broken down into all the words, and each word was no longer arranged in a sequence, but rather scattered into a circle, connected by lines.

"I used a different approach: parallel semantic association. After a sentence is input, all words are processed simultaneously. Each word is compared with other words in the sentence to calculate its relevance."

On the screen, the lines connecting the different words in this sentence vary in thickness.

"The relevance calculation method is automatically learned; it doesn't need to tell you what words should be associated together. It doesn't need a dictionary, a syntax tree, or any manual annotation."

It learns from massive amounts of text data. "Movie" and "boredom" always appear in similar patterns and positions, so it automatically links the two words together. But that's not the most crucial point.

Another screen appeared on the big screen, showing the same sentence, but with three independent sets of connections. Each set of connections was independent, did not interfere with each other, and ran simultaneously.

"Most importantly, this correlation calculation can be performed on many groups simultaneously. Three groups, eight groups, sixteen groups. The more the better. Each group focuses on a different correlation pattern."

He pressed a button on the remote. The first set of lines was highlighted, connecting the semantic relationship between "movie" and "boredom". The second set was highlighted, connecting the temporal relationship between "yesterday" and "watched". The third set was highlighted, connecting the interpersonal relationship between "friend" and "me".

"The first group focuses on semantics, the second on tense, and the third on relationships between characters. They work simultaneously without conflict. The machine doesn't need to understand what 'semantics' or 'tense' are. It simply needs to focus on the same sentence in different ways at the same time."

The technicians below remained relatively calm; these weren't particularly shocking things, but it was different if they had been successfully implemented technically. They began to look forward to the demonstration that followed.

Chen Yiran didn't stop, and continued talking with a different picture.

"There's another thing, position. The meaning of a word depends not only on what words surround it, but also on its position in the sentence. I chose to attach a positional code directly to each word."

This architecture, with its parallelized semantic association, multi-head attention grouping mechanism, and positional encoding, combined with the DeepFlow framework, is the product I'm introducing today.

Chen Yiran took out a Samsung Android phone from his pocket. Staff members came on stage and used a camera to film the phone's screen, which was then displayed on a large screen in real time.

The Samsung representative below immediately straightened his back.

Chen Yiran said directly into the phone, "Bun, help me find an Italian restaurant near Mountain View."

A text reply appeared directly on my phone: There are three Italian restaurants near Mountain View with four stars or higher on Yelp. The number one ranked one is Osteria, on Via Castro.

"I'm too lazy to read, just tell me."

Then everyone heard a gentle female voice, almost devoid of any mechanical quality, repeat the text.

"I don't want the first one, I like the second one. Give me the address of the second store."

The voice answered directly.

"Forget it, just call the third company directly."

The phone switched directly to the caller ID and dialed. After a few rings, the call was answered: "Hello, this is Sun Pizza. How can I help you?"

The audience immediately erupted in enthusiastic applause, and some scholars even stood up.

Chen Yiran quickly hung up the phone and began to explain, since not everyone on site was a professional with such good skills, and some people might not have realized what these three questions meant at first glance.

"It knew what I meant by 'the second one.' I didn't repeat the restaurant name, and I didn't say 'the previous one.' I just said 'the second one.'"

Chen Yiran lowered her head again and said, "Send a text message to Gao Dewei for me."

A gentle female voice asked, "What was the content of the text message?"

"Tell him we're having dinner at Sun Pizza tonight."

"Sent. Content: Dinner at Sun Pizza tonight."

Another round of applause and laughter erupted from the audience.

Chen Yiran didn't stop: "Check the weather in San Francisco tomorrow. If it rains, set an alarm for me at 7 a.m. The alarm will be called 'Bring an Umbrella'."

"It's 60% chance it will rain in San Francisco tomorrow. I've set an alarm for 7 AM called 'Bring an Umbrella.' Another alarm for 7 AM called 'Run' has been detected on my phone. Do you want to keep both alarms running?"

Chen Yiran laughed and said, "Never mind, I won't add it. Just set a reminder for my meeting tomorrow at 3 PM..."

He paused for two seconds before continuing to speak.

"No, let's change it to tomorrow at 2 PM."

The meeting reminder was set very accurately on my phone for 2 PM.

The applause was even louder this time. Chen Yiran didn't say anything more and just stood there.

The people in the audience thought it was over and started to stand up and applaud.

But Chen Yiran made a gesture and said, "You think this is the end? I have to say, although I am very proud of what I have shown so far, what I am most proud of so far is DeepFlow."

Unfortunately, although it's an open-source project under the MIT license, its development has been rather lackluster. This is quite disappointing. Today, I'll show you just how powerful DeepFlow truly is!

Next, let me introduce the complete form of "bun," the "bean bun." I call it the first intelligent agent program in human history.