It was 10 p. m., and Ryan was still sitting in his office.
Looking out through the transparent glass wall, the office area outside was also brightly lit, with researchers, programmers, product managers, and marketing team members sitting at their workstations, all working overtime.
Normally everyone would be off work by this time, but today is different.
This morning, NeusAI just finished a press conference that should have been recorded in the company's history.
The Venus model, which had been secretly trained over the past six months, was released, topping the leaderboard and representing a leap forward in the engineering capabilities of the Atlas Agent. This was further reinforced by strict access restrictions.
This should have been the headline news of the week.
However, just two hours later, OpenAI released GPT-V and GPT-VOmni.
Ne. usAI's Venus model launch event turned into a joke.
This sudden blow felt like a slap in the face to every employee. No one wanted to leave, and no one could really focus on their work. Everyone was burning with determination, wanting to regain the face they had lost that day.
Retracting his gaze, Ryan reached out and rubbed his temples, leaning back in his chair.
Then, forcing himself to stay awake, he straightened up and opened a new webpage on his computer. It was a third-party website that aggregated Large Model ability assessments.
Ryan couldn't explain why he wanted to watch it again.
He'd checked the list several times since it was released by Venus. For a brief two hours, Venus was at the top, but then it remained in second place, with the glaring name GPT-V in first.
Still second place----
Ryan glanced at the list again and suddenly felt something was wrong.
Venus is not second.
Second place goes to GPT-V.
Who came in first place?
Ryan stared at the unfamiliar name, TangYuan-0.9.
Then his gaze moved down to another name a little below on the list: KunYuan-1.0.
Seeing this, Ryan found it absurd and laughable.
No, what are these China companies up to?
Is it really worthwhile to engage in this overfitting and ranking manipulation?
Ryan Supreme Purity. The evaluation sets for these leaderboards are public; as long as the training data and evaluation questions are mixed together, the scores will naturally go up. This is called data pollution.
Everyone in the industry knows about it, but nobody dares to do it, because doing so would mean throwing their reputation into the trash.
usAI has developed so many generations of models, honestly pushing its actual capabilities to the top. Why? Because it's meaningless for a new company to reach number one on the leaderboard through minor tweaks; nobody will believe it. It's like this: if you're the world's richest person, even if you buy a fake luxury item from a street vendor, everyone will think it's real. But if you're just a low-paid office worker, even if you spend a year's salary on a genuine Louis, everyone will assume it's fake.
In this industry, sometimes reputation is more important than results.
A real-world example is right before our eyes: the last company to falsify its rankings was KunYuan, released by a China company at the beginning of the month. When the actual tests came out, the inflated scores were immediately exposed.
Ryan frowned as he stared at the screen. Which naive startup company was this? Didn't they even know this basic rule?
Then he thought again, this might not be a bad thing. Right now, OpenAI's GPT-V is at the top of the leaderboard, let Sam worry about that.
OpenAI had just dethroned NeusAI from the top spot and hadn't even had time to consolidate its popularity when, just a few hours later, it was overshadowed by a model that appeared out of nowhere.
This is hilarious.
Ryan was about to close the page when there was a knock on his office door.
"Come in."
John pushed open the door and walked in, his face even more grim than when he saw the OpenAI press conference that afternoon.
Ryan glanced at him and asked, "What's wrong?"
John didn't answer immediately. Instead, he walked behind Ryan and said, "Look at the link I just sent you."
Ryan opened Slack, John had indeed just sent him a link. He clicked on it, and it was a video.
The video footage is very basic; it shows a simple conference room with a China man sitting behind a long table, next to a man and a woman.
He had met this China man before; he was Louis Han, the CEO of Source Code Technology.
There are English subtitles at the bottom of the screen.
Ryan glanced at the title.
[TangYuan-0.9 Launch Event]
Glutinous Rice Ball) press conference.
Ryan's heart skipped a beat. Was this TangYuan, who had suddenly appeared, related to Riding the Wind? He asked, "What kind of press conference is this?"
Riding the Wind is now his biggest worry, and any news related to Riding the Wind deserves his close attention. NeusAI's engineering team implemented a new user authentication scheme at 00:00 Green on April 1st, and according to the plan, Riding the Wind should now be paralyzed.
"See for yourself," John gestured for him to press play.
The two of them, one sitting and one standing, one in front of and one behind, watched the video.
Ryan quickly realized that this was a crisis Public Relations by Source Code Technology in response to NeusAI's supply disruption, and that the Tangyuan Model was the key to their rescue.
The question now is whether this model is truly that powerful, or whether this is just a temporary emergency measure devised by Source Code Technology to release a fake high-scoring model to stabilize users before figuring out a more effective solution later.
If the Glutinous Rice Ball were released by another company, Ryan would naturally be 100% certain that they were fake.
But upon seeing Han Luyi, Ryan was no longer certain. After all, this remarkable China CEO had already accomplished far too many incredible feats.
But can Han Luyi really train a model of this caliber in such a short time?
Ryan shook his head slightly, hoping it wasn't true.
Because if this is true, it would be far more terrifying than losing to OpenAI.
"Have you verified it?" Ryan turned to look at John.
"Not yet," John replied. "I just saw it myself and sent it to you immediately."
Without hesitation, Ryan immediately closed the browser and opened Riding the Wind on his computer.
Watching press conferences or reviews is no substitute for trying it out yourself to be sure.
In fact, to assess the gap between Atlas and Riding the Wind, he himself was a heavy user of Riding the Wind. He was confident that he could tell immediately if his abilities were deteriorating.
Ryan pulled out his dedicated test project and entered a very complex modification request.
Riding the Wind began breaking down the requirements, and the waiting time became noticeably longer. However, Ryan was unsure whether this increase was due to the model's processing capacity or network latency. After all, if their new model were deployed in China, communication itself would cause significant latency.
Soon, Riding the Wind received a response and began working.
Soon, the work was completed, and a summary of the changes was displayed on the screen.
Ryan stared at the screen, his fingers hovering over the keyboard.
He was originally going to enter a second test task, but after a long time, he didn't move.
The results of the first test task are sufficient to illustrate the point.
Their abilities haven't declined; in fact, they've improved significantly.
At that moment, Ryan finally realized that his enemies were not only Google and OpenAI, both in California, United States, but also an even more terrifying enemy across the ocean, an enemy that had once been so close to him.
He began to doubt whether the user restriction policy he had just announced was the right decision.
"They actually did it," John whispered from the side.
On the night Riding the Wind was released, the only comforting thing they said when discussing Riding the Wind was, "They don't have a model."
Now they have it.
Ironically, working backward from the time required to train the model, on the night Ryan and John comforted themselves, the training of the model named TangYuan was nearing completion, something they were completely unaware of at the time.
At the same time, in downtown San Francisco, at the OpenAI office.
After the press conference, the employees returned to the office and held a small celebration party. After learning about NeusAI's press conference plans, Sam specifically scheduled his event for two hours after NeusAI's.
--
This arrangement indeed created a stark contrast, with OpenAI's two newly released models dealing a devastating blow to Ne. usAI's challenger status.
The employees are very excited, as this means they are still the company that is moving the fastest and furthest along the path of AI.
As the celebration party drew to a close, employees left one after another. Many of them came over to say goodbye to Sam and congratulate him before leaving.
Sam sat in a corner seat next to a floor-to-ceiling window, from which he could see the Bay Bridge over the sea.
Night had fallen, and the bay bridge was lit up with cool white lights.
Sitting across from him was Marcus, an old friend he had known since his time at Y Combinator and one of OpenAI's earliest institutional investors.
There were two cocktails on the table, almost empty.
Marcus was scrolling through his phone when suddenly his fingers stopped. He read quietly for a moment, then turned his phone around so Sam could see it.
It's a screenshot of a third-party AI model evaluation ranking list posted by someone on X.
The top spot on the list is held by TangYuan (-0.9).
"Have you heard of this company?" Marcus asked.
"Source Intelligence Technology." Sam pushed the phone back. "Source Code Technology is associated with a model company, made by Han Luyi."
Marcus raised an eyebrow, surprised that Sam had revealed the name so quickly.
Are you following this team?
"I've been working on this for two hours while they were having their victory celebration," Sam said, picking up his half-finished cocktail from the table. "Do you know what kind of reputation Riding the Wind has in the developer community?"
While everyone else is celebrating their success, you're researching a small company far away in China?
Marcus looked at Sam with surprise, without answering his question.
Sam waited a moment, then realized Marcus hadn't heard him, so he asked again, "Do you know what kind of reputation Riding the Wind has in the developer community?"
Marcus thought for a moment, then shook his head. He hadn't been investing in small companies much lately, and he even met with founders less often, so he had no chance to get in touch with developers.
"I had the team conduct a survey on the programming Agent field, specifically regarding Riding the Wind's current penetration rate in the United States and Atlas."
"They're neck and neck, and they haven't done any advertising or promotion in the United States at all." Sam looked at Marcus and said, "Do you know what that means?"
As a seasoned investor, Marcus certainly knew this.
Despite insufficient marketing and promotion, similar usage demonstrates a significant lead in product capabilities, allowing organic growth driven by user referrals to achieve the same effect as marketing.
"So," Marcus asked, "That score, you think it's real?"
"Whether it's real or fake," Sam said with a smile, "It doesn't really matter. From now on, competition between Large Model won't be based on rankings."
After saying that, Sam stood up, walked to the French windows, and looked at the night view outside.
"What I'm more worried about is their ability to apply it."
Marcus also stood up and walked over to Sam.
"What do you think is ChatGPT's biggest moat?" Sam asked.
Is it model capability?
Sam shook his head.
"Is it user volume?"
Sam shook his head again and gave his answer, "It's a habit."
"Hundreds of millions of people open ChatGPT every day, not just because it's the smartest model, but also because it's the first AI they learn to use. Once this habit is formed, it's very difficult to replace."
He paused, his tone turning somber. "But Riding the Wind is different. They've entered a completely different market. While we're targeting ordinary users, they're targeting programmers."
Marcus frowned slightly.
"Developers put their projects, context, version history, workflow configurations—everything's inside. Switching to another tool isn't just about changing the chat window; it's about changing your entire work environment." Sam turned to Marcus. "The migration cost is an order of magnitude higher than replacing ChatGPT."
"So what you're worried about is the lock-in effect," Marcus said.
"More than that," Sam said. "This actually exposes a problem we have: without productivity tools, there's no strong binding. Code development needs to be accelerated. We need to release at least the version for programmers that's comparable to Riding the Wind first, or we'll lose to them sooner or later."
"I will have the team release a code preview version within a week to deal with the impact of Riding the Wind."
Sam only said these things to Marcus; he didn't show this sense of crisis to other investors or his employees.
Marcus looked at his old friend, now one of the most powerful and influential people in the tech world, yet he never let his guard down in the face of the threat of challengers.
"You think—they're a threat?" Marcus asked. Throughout their previous conversation, Sam had consistently mentioned NeusAI.
It is considered the greatest threat.
Sam remained silent for a long time before speaking.
"Before today, I had never heard of this company," he said. "But from today onwards, they are a greater threat than Neus."
Marcus was silent for a moment, then slowly raised his empty glass, and the two gently clinked glasses.
"For OpenAI," Marcus said.
"For AGI," Sam said.
Santa Clara, Nvidia headquarters.
Nvidia CEO Jensen Huang was still sitting in his office. By this time, most of the people in the office building had left, but his secretary knew that he could work as long as needed.
There were two other people sitting in his office.
One is Philip, the vice president in charge of the Data Center business, and the other is Travis, the head of the strategic research department.
Travis placed a newly compiled briefing on the table in front of Jensen.
--
"Today, a China startup called Source Intelligence Technology released a Large Model," Travis said. "It's called TangYuan, version 0.9. The release was very low-key, just an online live stream, no dedicated launch event, but after the data was released, it's now ranked number one on third-party charts."
Jensen didn't pick up the briefing; he sat quietly in his chair.
The name TangYuan doesn't mean much to Europeans and Americans, but Jensen knows what it is. He loved eating it when he was a child, with black sesame filling and red bean filling.
It's interesting that someone named a model after a snack.
"Where does their Computing Power come from?" Jensen asked the question that concerned him most.
Rankings aren't important. Today in Silicon Valley, two of Nvidia's customers got into a fight over rankings.
For Nvidia, this competition is a good thing, as long as they are all using Nvidia GPUs.
The larger the model, the more intense the competition, and the higher the Computing Power requirement.
Regardless of who wins in the end, more GPUs will be needed, which is exactly the situation Nvidia wants to see.
However, a previously unheard-of China company suddenly released a Large Model with capabilities surpassing OpenAI and Neusoft.
An AI model must have a huge training cluster behind it.
How many cards does this cluster have? What are their models? Where is it located? Who provided it?
Most importantly, has it bypassed Nvidia's ecosystem?
This is what Jensen really needs to know.
Philip took on the question; he had done his homework before coming.
"Based on the information we have, they purchased GPUs that had just entered mass production from another China startup. The specifications were even lower than our H100, and they only bought 500 of them."
Jensen, who had been leaning back in his chair, suddenly sat up straight upon hearing this, startling his two subordinates with his sudden movement.
"Impossible!" Jensen blurted out.
Only 500 images? Training a state-of-the-art (SOTA) Large Model? This doesn't fit his understanding.
"No one can train Large Model on this scale," Jensen said confidently.
Travis cleared his throat beside him.
"I have some other information here, which I collected through special channels," he said. "This Source Intelligence Technology should be using Dingsheng's Computing Power."
"How much did Dingsheng purchase?" Jensen asked.
Philip quickly checked his computer and found that Dingsheng was one of the largest clients in the China market, with aggregated data from various channels being updated in real time.
"Through all channels, there are at least 30,000 A100 tickets, 80,000 A800 and H800 tickets, as for the H100—"
Philip didn't have the data because H100 is strictly prohibited under the ban, and all its distribution to Dingsheng went through some gray channels.
Jensen knew his concerns but didn't press him. Knowing a rough number was enough for him.
If it uses the Computing Power resources of Dingsheng, then it makes sense.
At least it's still within Nvidia's ecosystem. This relieved Jensen.
It's important to understand that Nvidia's high profits and soaring stock price today all stem from a consensus: at the GPU level, Nvidia has achieved a de facto monopoly thanks to its CUDA system, which it has accumulated over nearly two decades.
This consensus entitles the holder to a monopoly premium.
But if this consensus is broken one day, Nvidia's story will no longer make sense.
This is why Jensen is so opposed to blocking Computing Power.
If you don't let me sell to them, they'll develop their own!
What will I do then?
Those idiots in Washington.
Unfortunately, even Jensen Huang, who ranks among the top ten richest people in the world, has moments when he has to give in.
Bringing his thoughts back, Jensen dismissed his two subordinates from the office.
After the two left, Jensen sat in his office for a while longer.
He suddenly thought of something.
A month ago, Dingsheng also released a Large Model, which only stayed on the list for a short day before being knocked down.
With the same Computing Power, Dingsheng produced Kun Yuan, while Source Intelligence produced Glutinous Rice Ball, and the results were even better than Dingsheng's.
This team has very strong technical skills; we need to keep a close eye on them.
Thinking of this, Jensen picked up his phone and sent Philip another message: "Keep a close eye on that new China model-making company's GPU procurement. Report any unusual activity immediately."
Then he put down his phone and closed his eyes.
Suddenly, the image of a bowl of Glutinous Rice Ball appeared before my eyes, floating directly above Silicon Valley.
He opened his eyes; if this continued, they would really do it.