Sensing Meng Fanqi’s unique insights, Jeff also shared some of Google’s current internal ideas and goals for AI:
First, actively use AI to solve various problems, including translation, maps, languages, images, etc., try to incorporate AI into existing Google products and services, and use AI technology to improve product capabilities and user experience.
Second, we will continue to launch a variety of AI software and related tools to make AI technology easy to use and accessible, continue to share the latest research results, and continue to make papers to the public and open source code. Benefit more researchers and allow developers to use the AI tools created by Google to better develop their own machine learning models.
Third, use AI to solve big problems facing the world, such as medical care, energy, environmental protection, etc., and cooperate with other organizations to jointly solve challenges.
Since there is currently not much content about AI that can really be applied to industry, Jeff’s pieces of cake and these routes are a bit fake and empty.
After Meng Fanqi heard this, he directly started to draw a cake for Jeff, and the drawing was more realistic and specific.
"I plan to quickly launch AI models and technologies to solve various practical problems. The long-term goal is to solve them in a universal way. For example, language, translation, question and answer, classification, generation, all can be solved with one model.
Not only the language model must be a universal model, but the visual speech model must also be an ultimate behemoth, and it can directly generate classification, detection, segmentation and generation for it as a four-in-one model.
First, it sets off a visual storm and revolutionizes image tasks, then unifies the paradigm of language models, and then integrates multi-modality, where language pictures and sounds are directly integrated. "
Meng Fanqi talked eloquently, leaving Hinton, who was still thinking about how to design the backbone neural network better at this time, and Jeff, who was leading such a large group for the first time, stunned.
Jeff made everyone dizzy when he heard this, "Good guy, your pie is bigger than mine?" The little thing is blowing well.
I'm going back to Silicon Valley in a few days, and I'm going to brag to the CEO of Google at this pace!
It took one year to revolutionize machine vision, two years to unify the language paradigm, three years to integrate multi-modality, and in the fourth year a large model has unified the world.
Is it the fifth year to achieve full-level autonomous driving, the sixth year to be a fully intelligent robot, and the seventh year to create a Terminator in 2020?
The tenth anniversary is a direct virtual invasion of alien spacecraft.
Jeff never thought that Meng Fanqi was really not drawing cakes.
Although the time is a bit bragging and exaggerated. For example, a large model with hundreds of billions of parameters at the ChatGPT level would be more difficult to train without NVIDIA's A100.
If the card is not good enough or advanced enough, the more cards will be needed for distribution. The greater the number of cards, the more likely it is that there will be hardware problems. Once a few cards have errors, the entire process may be affected.
Judging from NVIDIA's release time, if you want to produce a conversational robot of ChatGPT's level, it is likely to be limited by hardware and will not be available until the end of 2020, which can only be about two years in advance at most.
But in order to launch this series of technologies in advance again, Meng Fanqi may also consider sacrificing some of ChatGPT's performance slightly, thereby pushing forward the emergence of ChatGPT by one or two years.
Counting from now on, four years may still be a bit bragging, but five or six years is still very promising.
But Jeff is also an honest man, and he knows very well that he is suspected of being a charlatan.
The status of the Google Brain team was not high at first. Although Andrew Ng founded the project and proved the potential of deep learning, the team had no output for a long time.
Many senior people at Google believe this is a waste of money and are skeptical and negative about the team's direction.
The breakthrough results of the Hinton team in 2012 and the addition of the Hinton team at the beginning of the year have largely alleviated this problem. Everyone still has great respect for this old godfather of AI.
However, although Hinton is responsible for the research, the efficiency and quality have improved a lot.
The major matter of technology implementation has never achieved a phased result, and there are no actual projects that really use new technologies.
The recruitment of Meng Fanqi this time, in addition to once again using his amazing results to improve the status of the Google Brain team, Jeff was very interested in the real-time detection algorithm just now.
This is likely to be a large-scale technology application that can completely wipe out the decline of the Google Brain team that has only been working in small projects.
"About the real-time detection algorithm, Mr. Li of Baidu must have summarized all aspects of the situation in great detail." After discovering that Jeff was also thinking about this matter, Meng Fanqi laughed and carefully selected the current situation. The technology that realizes the fastest deployment in stages is indeed quite attractive.
"Mr. Li's PPT and technical conference must be very thorough. His presentation is much more detailed than what I can say directly. Let's just watch Baidu's technical conference on Friday."
Are you kidding me? You haven't seen the sharing contract yet, but you want me to pay for the technology? Meng Fanqi remained calm and started practicing Tai Chi.
Although it is a foregone conclusion to sign with Google early because of the need for Google's framework and TPU, it is impossible to discuss technical details before signing the contract.
If Google's share agreement turns out to be really unsatisfactory, Meng Fanqi can simply not hand over a lot of technical content. Even if there are contractual provisions for such pure technology and ideas, it is not that difficult to bypass restrictions and publish them through friends. .
Google here has just become interested in the real-time detection algorithm and launched an early attack. Baidu, which has reached a technical cooperation with Meng Fanqi, has already discussed the ideal stage with many leading companies, and the progress is very rapid.
Since the service content this time is very technical, the actual needs of customers are many different from each other. It is often necessary to go to the customer's site for on-site operation, debugging and demonstration according to the requirements. Most of Baidu's employees are not qualified to do this in a short time. Work.
But Baidu Deep Learning Research Institute, this part that was originally a research institute, was fully mobilized, and the researchers were directly transformed into sales + customer service. This is also a microcosm of domestic major manufacturers' eagerness to realize technology.
As of now, the Ahri and Penguin series have no special needs, but they are beginning to realize the importance of AI technology, so they hope to have technical exchanges. The two parties have already talked about the third phase.
Several large-scale manufacturing and coal and metal industries still need to provide some more data and do some model training work. However, the current negotiation situation is quite good, and the expectations of both parties are very high.
The most important thing is the government’s road monitoring and security-related needs. If you win this part of the order, it will be a big deal.
Government orders are stable and long-lasting and generally do not change frequently. In particular, the order volume in this area is much larger than that of enterprises, and the demand is usually lower than that of enterprises. There may be more procedures and trivial matters, but it is completely acceptable.
The person responsible for connecting with relevant government departments is a group of elite candidates personally led by Vice President Yu Kai.
Naturally, this includes Wang Kai, who had direct contact with Meng Fanqi and contacted him many times for technical advice.
At this moment, he was sitting in the Public Security Bureau, looking left and right.
