It is normal for this kind of negotiation to not go through the entire process at one time. Even if it is a visit with a clear purpose like Robin Li's, he will carefully check the contents of the contract again after the verbal agreement.
Google is preparing to draft a sharing contract within one to two days, and it is moving very quickly.
The overall intention has basically been achieved, and the matter of joining Google is still a two-way process.
Meng Fanqi needs to use the deep learning framework TensorFlow (TF) to be developed by Google. The current code implementation method is too inconvenient.
Google's massive computing resources and later tensor computing unit clusters are also indispensable tools for Meng Fanqi to achieve various pioneering achievements in the future.
Thousands of computing units, whether NVIDIA's GPU or Google's TPU, are always more comfortable in large factories.
If he builds a cluster by himself, not to mention the cost of hundreds of millions, it will be difficult for him to afford the losses just to support the continuous operation of these devices.
So at this point, Meng Fanqi must join Google. Google has also accepted the split method, but Google needs to carefully evaluate how to divide it.
Now that the general direction of employment has been decided, there will naturally be some bragging in the remaining time.
Don’t underestimate this part, especially before the contract is officially signed. The big pie is spent well, but the budget is cut to pieces.
I'm really fooling you. With some value and emotion, it's very possible to win people over with a low budget.
"I've been at Google for almost fifteen years, where I've seen the rise of the internet, and now I think I'm leading the rise of artificial intelligence."
Now this is the time to discuss the past and future vision after talking about the business. Jeff hopes to introduce some of his and Google's plans and blueprints for the future of AI.
“Helping computers recognize objects, understand language and speech, and even conduct conversations, things that seemed like a fantasy before, are now gradually becoming a reality.”
"Just take the direction of computer vision. In the past five to ten years, computers have rapidly developed the ability to [see], and judging from your latest achievements, it has rapidly reached humans. level."
Technology in the AI era is developing too fast, which is the core reason why Jeff is willing to spend a lot of money to recruit talents.
"Google now has many scenarios where it hopes to develop AI technology. We want to achieve mutual translation of more than 100 languages so that people can communicate better; we want to intelligently analyze medical images to predict and diagnose diseases more accurately. All these applications Among them, the most core things are actually two things, algorithms and computing power.”
Jeff's summary is very concise. Modern AI is mainly based on the ancient algorithm of neural network. If you put aside computing power to talk about AI algorithms, it is completely a castle in the air.
"Google is determined to build the most powerful computing platform in the world, and we will definitely allow excellent and intelligent algorithms to maximize their value."
Meng Fanqi had no doubts about Jeff's determination, which was why he initially chose Google.
“The meaning of computing power is relatively pure and easy to understand. But algorithms have too many meanings. In fact, I personally feel that the design of the network structure itself is not the focus and core.
If you really want to change the world, you need a framework and platform that is simple and easy to use, easy to deploy, and optimized and accelerated in the data types of operations. "
At this stage, the industry is paying great attention to the design of neural networks, specifically how to design which layer, and what operations are better.
During this period, the benefits of doing so were also huge. For example, last year's AlexNet and this year's DreamNet both had terrible improvements.
However, in Meng Fanqi's view, the structure in the late AI era has not changed much. The most important thing is to work hard to achieve miracles. In addition, he knows which structure is better for which tasks, and the structural design is too simple for him.
“When the competition comes to an end, large model training technology and massive high-quality resources are more critical.”
Jeff and Hinton secretly exchanged glances, feeling strange.
Originally, it was Jeff who came to show this undergraduate who was still exposed to research in school Google's ambitions, such as the multi-field blossoming and the largest computing platform.
Why do I feel like this guy has such a clear understanding of the main pain points of the AI industry? He doesn't look like he's conducting research in an ivory tower.
The academic community studies AI mainly to verify a certain conjecture and improve specific indicators.
Industrial AI is more pragmatic. How can it be implemented with fewer resources, how can it be built faster, and how can it be deployed on different devices?
The two sides often dislike each other. The academic community feels that the industry is just a bunch of wage earners doing dirty work without any innovative breakthroughs. The industry feels that academia will write papers and brag, and no one will use the things they produce.
Jeff and Hinton can be said to be representatives of industry and academia. Even when Jeff was studying, the graduation thesis he wrote was industry-oriented, parallel training of large neural networks.
It was only 1990, and Jeff had already begun to study the most core technology in 2023, the training method of large models.
"I have to say, I thought you would be a more academic person who has continuously made breakthroughs in algorithms." Jeff's expression was filled with surprise, "I didn't expect your way of thinking about problems and the needs of our industry. Very much in line."
Jeff has come into contact with many outstanding scholars, and even Hinton has the thinking inertia of the academic world. Therefore, within Google Brain, Hinton does not participate in any management and decision-making work, and is only responsible for academic research.
Perhaps this time, I am not just recruiting an outstanding algorithm researcher, he may also be able to give me a lot of help in the company's AI strategy.
Jeff had a vague hunch.
He had been the technical backbone of Google for more than ten years and had not participated in many management projects. But the direction of AI was promoted by him with strong support from Andrew Ng, so many things in this area are led by him.
As a leader, Jeff likes different perspectives and new things.
For example, neural networks and AI, although he also studied them in the 1990s, later his work at Google was more about architecture, search, and advertising.
In fact, no AI knowledge has been updated.
Until 2011, when Ng was working on a project with Google, he suggested to Jeff that the situation was changing rapidly and Google should pay attention to AI technology.
Jeff quickly embraced this change, and it can even be said that he was naturally interested in this potential solution that he was not familiar with.
Once he understands the solution to the problem, he will lose interest.
Meng Fanqi's thoughts on AI strategy were different from his own, but Jeff was even more delighted, secretly thinking that he had really found a treasure this time.
