On February 29, Zhou Hongyi, the founder of 360 Group, gave a free lecture and opened the first lecture of the AI series - "Foreseeing AGI".
This course systematically shares the insights into the latest development trends of AI, the five levels of multimodal development, and the five stages of AI development, and other hard-core knowledge, and the following are 21 course notes to help you organize.
Today I will talk about four parts:
First, new insights into AI development.
Second, SORA inspires us.
Third, what do we do?
Fourth, 360's new AI products.
1. Learning methods. When I enter a new field, I first have to learn the framework, many new technologies emerge in an endless stream, first of all, I must have an overall grasp of it as a whole, and if I grasp this whole, there will be no deviation in the general direction.
2. Why do you want to come to my class? Everyone is equal before the skills and knowledge of artificial intelligence, which is a professional skill, just like learning to drive. Therefore, I believe that learning artificial intelligence is a basic ability that everyone must have in their future career development.
The three major methods of AI. Believe that the big model is really intelligent; Believe that the big model is an industrial revolution; Believe that all businesses will be reshaped; believe that companies that do not embrace AI will be eliminated; believing that employees who do not embrace AI will be replaced; It is believed that artificial intelligence is moving rapidly towards AGI.
Part 1: Top 10 AI Development Trends for 2024**
Top 10 AI Development Trends of the Year**.
a. The outbreak of open source large models; b. "Small models" emerge and run on more terminals; c. The rise of the large-scale model enterprise-level market is developing in the direction of industrialization and verticalization; D. 2024 is the year of large-scale model application scenarios, and TOC will have killer applications; e. Multi-modality has become the standard configuration of large models; f. Breakthrough progress in AIGC function of Wensheng diagram and Wensheng **; g. Embodied intelligence empowers the humanoid robot industry to flourish; h. Large models promote breakthroughs in basic science; i. Large models are ubiquitous and have become the standard configuration of digital and intelligent systems; j. Agent stimulates the potential of large models and becomes a super productivity tool.
Part 2: SORA's Innovative Implications
5. The appearance of SORA is indeed beyond my expectations, beyond everyone's expectations, it seems to be a tool for cutting **, but it is not.
6. The essence of SORA's innovation breakthrough is to understand the common sense of the world through observation.
7. In the training process of SORA, SORA should not only understand what is in the multimodal input, but also learn some of the laws behind it similar to human common sense. So, I think, it "understands" some of the laws of the world.
Clause. First, if you understand the law, you don't necessarily understand the formula. Second, there are some things in the ** made by SORA that it doesn't quite understand, which has something to do with training, because after all, the computing power is limited now.
8. Why did SORA hit Pika and Runaway with dimensionality reduction? The principle of SORA, through the learning and training of images and **, it knows the interaction of some common objects in the world. So, why is it a world simulator, you must first understand the world to simulate the world, and to understand the world is not necessarily to understand the language, but to understand the basic laws of the world.
9. How to define understanding? Sora's full ability** proves that it has understanding. Openal revealed: SORA can identify, process, analyze, understand and generate ** and images. Is Sora a pixel manipulation? I've proven that you can't make such a realistic picture if you only manipulate pixels without understanding them.
10. Five levels of multimodal development:
11. "Sora" is normal to roll over, just like human dreams, there are "hallucinations". If there are not enough training samples, there is no built-in 3D engine.
12. Conjectures about the implementation of SORA technology. There is a philosophy that is to vigorously produce miracles, the aesthetics of violence, that is, the computing unit can be very simple, but it can be infinitely superimposed, and this principle I think is exactly the same as the philosophy used by the Creator to create our world.
13. One of OpenAI's recent achievements is what does SORA prove? It actually uses an architecture to process text, audio and unity, which is remarkable, in the past we did multimodality, and many multimodalities are fake, that is, a model processing, a model processing, so that it can't get through to each other, and it can't help.
14. GPT solves the problem of communication between machines and people, and SORA solves the problem of interaction between machines and the world.
15. The emergence of SORA has accelerated the arrival of the AGI era. The sora looks like a hair dryer but is actually a razor. I understand AGI - the ability to communicate, break down tasks, and perform tasks like a human.
16. Five stages of AI development:
17. Why develop AGI? First, we need to make breakthroughs in basic science. Human basic scientific research is facing a huge bottleneck, and AGI is urgently needed to bring about substantive breakthroughs. Second, reverse the solution to the energy problem.
Part III: What Should We Do?
18. How can China embark on a distinctive large-scale model development path? First, the super general model; second, the development of enterprise-level large models; Third, accelerate the implementation of scenarios.
19. Advice to entrepreneurs. First, it makes no sense for entrepreneurs to stop touching the general model. Second, don't do some simple casing and thin applications on top of the general model, so that the traditional model only needs one tool, and you are finished.
It is recommended to do it in two directions:
First, find a direction at the enterprise level, because the enterprise market has a very rich scene. We have so many large central enterprises and large private enterprises, and we have to do digital transformation and intelligent reform, and there will be a lot of opportunities here. Therefore, in the past, many small SaaS companies used large models to combine their SaaS with both sides, which I think is a very important direction.
Second, if you are looking for this kind of to C scenario, you must make the scene very heavy, very vertical, and the scene that is very shallow is not valuable, because other companies add a function to the browser a little, and you are gone.
Part IV: 360AI Products
I think that the large model must be combined with the scene, otherwise the large model will always be a toy for large companies to constantly show their technical strength. So, we did these two scenarios: first, trying to reinvent the search experience so that users can find answers directly through conversations; Second, position the browser as a learning and productivity tool, which can help you quickly read long**, 10,000-word**, tome books and long web pages.
21. Why use large models to reshape search?
a. User pain points and rigid needs. Search is still a rigid demand of users, but the current search has several problems: First, if the keywords are not accurate, the results will be very different. Second, the search results need to be clicked one by one, and the desired results can be found in countless links. Third, users need to summarize the search results by themselves.
b. What can AI + search do? With the blessing of AI, search can be turned into a personal intelligent assistant.
c. AI reshapes the product theory of search. The first thing to subvert the big model should be search, because search has not changed since the advent of Google in 1998 and 1999, the same recipe, the same taste, and the same search box. And 60% of users are actually looking for "answers" when searching. So we came up with two theories that reinvent search: conversational search and answer engines.
Finally, after reading these course notes, which one gave you the most gain?