The ride-hailing driver overheard their conversation but didn't find anything strange about it.
Liu Huo sat in the passenger seat, not daring to utter a sound.
One can only sigh silently in their heart: youth is so wonderful; youth is truly wasted on young people.
Gu Siyue and Han Luyi's planes were not in the same terminal. After dropping Gu Siyue off at Terminal 2, the car continued to take Han Luyi and Liu Huo to Terminal 3.
The Capital Airport was still very busy in the afternoon.
Long queues formed outside the security checkpoint, business travelers hurried forward dragging their carry-on suitcases, flight information was repeatedly broadcast over the loudspeaker, and the air was filled with the smells of coffee, perfume, and disinfectant.
Han Luyi found a seat in the business class lounge, Liu Huo went to get two free cups of coffee.
The business class lounge is, well, most of the time, more crowded than the waiting area outside. Nobody there is actually resting; most people are either on the phone discussing work or working overtime on their computers.
I think it would be better to call it the "Ox and Horse Overtime Room".
Han Luyi was no exception, except that he worked overtime in a more discreet way.
While sipping his coffee, he used his mind to operate his phone and communicate with Zhao Wenyuan on Lark about work progress.
Like ducks on Stanford Lake, they appear relaxed and carefree, but their flippers are actually treading water frantically underwater.
This method is efficient, but you have to avoid people around. However, even if someone sees a phone being operated by itself, they might just assume it's being remotely controlled.
Modern technology has long blurred the lines between magic and reality; technology is the magic of Muggles.
After finishing his conversation with Zhao Wenyuan, Han Luyi opened his eyes.
I'm a little tired.
I'm mentally exhausted.
While NeusAI's supply cut was expected, their release of new models, along with OpenAI's, was unexpected.
This dramatic event drew a lot of attention to Riding the Wind's API supply disruption. In order to stabilize customer confidence, Han Luyi decided to release Glutinous Rice Ball -0.9 and hold a brief press conference as a response.
Unexpectedly, this brief press conference propelled Source Intelligence to unprecedented heights.
It's been less than a week since the release of Glutinous Rice Ball, and Han Luyi is already incredibly busy with this matter.
The hype exploded, investors were clamoring for quotas, the government extended an olive branch, a trip to Beijing was made to report the domestic adaptation plan to relevant leaders, there was cooperation with Shuimu Forum, and the Strategic Emerging Fund actively invested.
These are all good things, but they are happening too much and too fast.
Han Luyi is readjusting to this rhythm and has found a key point that allows him to take the initiative.
The key point is Ximing.
Domestic adaptation is still basically in the stage of going from nothing to something. In this process, various domestic manufacturers are exploring their own paths, some fast and some slow, but no one has been able to make it work yet.
Jiang Songran recommended Ximing's L100 at the time, partly because its software foundation was poor. If Source Intelligence started from scratch, it would have no burden.
The cluster built by Source Intelligence with 500 L100s can now efficiently run the inference of Glutinous Rice Ball -0.9 and steadily handle the traffic of Riding the Wind.
This progress puts them at the forefront of domestic manufacturers.
The next challenge for Source Intelligence is to conduct training within domestically developed clusters.
Although both are part of Large Model ecosystem, the training difficulty is more than ten times that of inference.
Inference is essentially "using a model," loading pre-trained weights, taking a question as input, and outputting an answer. As long as the graphics card has enough memory, a single card can run it, though the response time will be slower. With a cluster, the response time can be faster, supporting more users and requests, but this is a quantitative change, not a qualitative one.
Training is about "creating a model." The Computing Power and memory of a single GPU are definitely not enough. To enable the model to achieve gradient descent and eventually converge, a cluster of at least several thousand GPUs is needed. This is the difference between being able to do it and not being able to.
More importantly, it's about time and money.
The Glutinous Rice Ball -0.9 uses the Computing Power of Dingsheng, and in just over two months, the quota of 100 million yuan has been almost used up. This is still using the mature Nvidia cluster in the ecosystem.
If we switch to a domestically produced cluster, the time will conservatively be more than three times longer.
Using non-Nvidia GPUs for training is a challenge that all Chip companies are trying to overcome, but so far only Google has succeeded.
Source Intelligence aims to become the next company.
Following the path of L100 is the fastest way. Moreover, software Hardware are inseparable. Once the training system is established by Source Intelligence, Ximing's Next-Generation Chip will need to work more deeply with the software, which must be considered from the initial design stage.
This is the best route for Han Luyi to achieve China and adaptation.
Finding other domestic manufacturers to cooperate with is also an option, but that would be slow and would take us out of our control. Acquisition is much more convenient, faster, and more reliable.
Strictly speaking, Ximing is not a good company.
The company was in poor financial condition, had a chaotic financing structure, and the founding team was left with almost no room to maneuver by investors. Its product line had also experienced several failures.
But these shortcomings became advantages for Han Luyi.
Buying Ximing means directly purchasing their inventory of over 8,000 L100 chips, which will greatly accelerate the Computing Power required for Source Intelligence's model training.
In addition, many people are now eyeing Glutinous Rice Ball.
Besides congratulations, there were also those who questioned, envied, and slandered.
This is normal.
When a China startup suddenly produces a model that ranks first on a list, especially against the backdrop of increasingly stringent Chip bans, it's natural for outsiders to be skeptical.
Suspecting it manipulates rankings and that it has selectively optimized certain ranking capabilities, yet it's not suitable for large-scale commercial use.
There is only one way to respond to these doubts and get the market to accept Glutinous Rice Ball.
Heap Computing Power, open API.
Make sure enough people can use Glutinous Rice Ball glutinous rice balls).
Let them use it in development, let them use it in universities and research institutes, let them use it in various industry scenarios, and let them use it in daily chat.
Once the number of calls increases, the effectiveness will naturally be verified.
Rankings are merely proxy indicators for judging the quality of a model; real-world usage is the ultimate measure of its effectiveness.
Therefore, the key point now is to acquire Ximing.
Given Han Luyi's past personality, he would have handled such an important matter himself from the very beginning of the negotiations.
But he was simply too busy to handle everything and didn't want to drag things out.
So he sent Lu Mingzhou.
The first lesson I learned after starting my own business was decentralization.
Last time, he entrusted the daily management of the Source Code to Su Niannian.
This time, he entrusted the negotiations with Ximing investors to Lu Mingzhou.
As it turns out, the results were quite good.
Boarding had begun, Han Luyi stepped out of the lounge and looked out the floor-to-ceiling window.
An Air China plane is taking off from the runway and soaring into the blue sky.
Looking outside, Han Luyi's thoughts became clearer and clearer.
Source Code Technology cannot be just an application company, and Source Intelligence Technology cannot be just a model company.
If Glutinous Rice Ball is to continue moving forward, it must have its own Computing Power foundation, model capabilities, application ecosystem, and Industry Alliance.
Therefore, Ximing's acquisition is not an ordinary one; it represents the first step that Source Intelligence is taking from a model company to a full-fledged Artificial Intelligence industry chain.
©