In fact, Meng Fanqi had fantasized about rebirth countless times. In his dreams, career and love always progressed rapidly and smoothly.
However, when he really had the opportunity to walk back to the university campus ten years ago, Meng Fanqi found that buying BTC was the only thing that was easier to do.
He does not have the memory to write "A Dream of Red Mansions" from another world and memorize 300 Tang poems and Song lyrics.
In the AI era, code and tools are updated so rapidly.
As of this time, not to mention that the two AI tools that will dominate the world in later generations have not been released, and even the earlier Caffe has not started development.
The entire academic world is still reeling from the aftershocks of the AlexNet earthquake. Many people in the academic world have not fully understood what a deep neural network looks like. Where can they find a ready-made framework for them to use?
Meng Fanqi was walking back to his dormitory and began to think about how to deal with the first problem after being reborn. As long as the server survives these few months, he can buy it with a portion of BTC.
And after winning the first place in the competition, Google will not be stingy about lending you a few machines. Meng Fanqi is very aware of this past incident. As long as the algorithm is successfully implemented, Google will pay him an astronomical number.
What is more troublesome is the code framework used to build this model. When he started to get involved, the hottest tool was TensorFlow (TF) released in 2015. Two years later, in 2017, Facebook proposed PyTorch, which was easier to use and gradually caught up with it. Meng Fanqi also switched to using PyTorch.
He had heard of some previous frameworks, but to be honest, he had never been exposed to them.
As we get to a later stage, various packages become better and better. Even an elementary school student can directly construct a high-performance cutting-edge model with two lines of code.
But that was ten years later, and now it is more than two years before TF releases the initial version. Meng Fanqi is a little irritable.
His current plan is to take a look at the source code of the original version of AlexNet last year, and then take a look at Alex's own homepage. There should be some gains.
If it doesn't work, I can only take advantage of the fact that Alex is now very proud and has not yet signed a contract with Google, so I can approach him in many ways and ask him for advice.
Since AlexNet is a milestone event in the deep learning era, and its model is named after himself, Meng Fanqi has a little understanding of his paper, the network and even himself. Click on the web page to search for his name, and you can easily find his concise personal homepage on the University of Toronto website.
Not only that, this personal homepage also details some of the codes he developed and accelerated while he was in school. Mainly C++ and CUDA code based on NVIDIA graphics cards, basically in the period of 10-12 years.
"Good people have a safe life!" Meng Fanqi clicked on the link and found that he not only recorded the code and results in it, but also clearly remembered what was corrected every time, and even left a sentence next to it, "If there is any If any code fails to run, please feel free to contact me.”
Meng Fanqi shed two tears in his eyes. It is precisely because of the selfless dedication and sharing of generations of algorithm pioneers that the AI industry has developed so rapidly in the past ten years.
"Good code, I'll copy it!" The data and code are being downloaded. If this cuda-convnet framework can be used, then half of the code problem will be solved. Meng Fanqi can initially implement some key issues based on this prototype framework. algorithm.
Thinking of this, Meng Fanqi had to sincerely thank later generations for the increasingly intractable algorithm interviews.
He has a very clear understanding of his abilities. Even ten years ago, his coding ability was not outstanding. His advantage lies in knowing the development routes of many technologies and companies.
The reason is that he knows these famous algorithm technologies well and can reproduce a considerable part of them. I would also like to thank the programmer for the unique interview mechanism. It is common for larger companies to interview you for three or five rounds of technology in one interview, test you on how to write code on-site, and ask you about classic technologies to help you implement them.
In addition, programmers especially like to share, and they love to write "interview stories", which are interviews and experiences. There are often well-intentioned people who provide very detailed and perfect answers to these questions.
This back and forth has been repeated for several years. In order to select candidates, large companies have asked more and more difficult questions, and the questions have become more tricky. This made Meng Fanqi familiar with the basic core technologies and classic papers of each route.
It's just that people in later generations have already written the tools well. Understanding these classic principles is of little use in work, but I didn't expect that they can be put to full use here.
In addition to the code, another important task is the details and reasoning of several important papers.
Meng Fanqi basically remembered the main context and logic, but the details were not comprehensive and needed to be polished hard.
If you want to share the biggest cake, there are still quite a lot of things to do. Meng Fanqi carefully calculated that many of the cornerstones of the future direction of AI were actually put forward between 2013 and 2015.
Moreover, if you frequently propose algorithms and publish papers one or two years in advance, it is likely to cause a butterfly effect, making progress in the field faster. As a result, you have less time than before.
Therefore, Meng Fanqi initially plans to implement or propose theories for the core algorithms in all directions in the first one to two years, that is, around 15 years.
In the following 15-18 years, we will mainly use the resources that have been exchanged, whether it is economic resources or traffic and popularity. Shift your focus and focus your energy mainly on those projects that break through the human level, such as AI Go, AI games, AI painting, and AI language. With AI based on its own algorithm, it frequently creates records that beat the top level of humans. The paper only needs to publish a few less important parts.
Once enough assets are accumulated in the process, you can hold large-scale shares in some technology companies in advance. By 2018-19, I can gradually transform, gradually leave academia, and focus on investment or AI entrepreneurship to avoid my gradual loss of technological advantages.
In addition, if there is anything missing, it can only be an independent residence. As a university with electricity in its name, Yanjing Electric Power University cuts off power to students every night. Sometimes model or algorithm testing needs to run continuously for several days, and Meng Fanqi doesn't want to be affected by lights out and power outages at night.
Moreover, if the computer is doing algorithm work, it will run under high load and make a lot of noise. Even if you can do this in the dormitory, I'm afraid it will make the roommates unable to sleep.
Rent a small bar near the school. The money I saved in the past two years has just been invested in BTC. I really have no choice but to ask my brother to borrow some.
If you ask your brothers to borrow money, your seniority will inevitably drop sharply. I haven’t even taken a few steps towards becoming an AI godfather, but “calling my father” has become a reality. How could those damaged roommates lend him money if he didn't call him dad twice?
"Cherish the last few months of searching for money." Meng Fanqi patted his empty wallet sadly, "I may never know what it feels like to be short of money in the future."
