Yan Junjie, founder of MiniMax: AI is evolving into a more powerful productive force.

One year ago, the industry was still discussing the parameters and capabilities of foundation large models. This year, what can be seen everywhere at the World Artificial Intelligence Conference (WAIC) are lightweight Agent systems, AI application solutions... AI manufacturers are showing a more practical and down - to - earth approach. All exhibition booths strive to fully demonstrate their commercialization capabilities, product implementation capabilities, and customer cooperation cases.

At MiniMax's booth, the main products on display are also AI application products such as MiniMax Agent, Conch AI, MiniMax Audio, and Xingye. These include AI intelligent hardware products like smart home devices, wearable devices, intelligent cockpits, smart speakers, smart headphones, and interactive devices, as well as applications in fields such as culture and tourism, e - commerce, office work, education, gaming, healthcare, and finance.

Photo by:

The reason behind this is that the large - model industry is undergoing a structural transformation. The inference cost has dropped by an order of magnitude within a year. The open - source ecosystem is booming, approaching the performance of closed - source models. The commercialization of Agent applications is accelerating, and emerging manufacturers focusing on vertical fields have the hope of forming differentiated competition with industry giants.

Against this backdrop, on July 26th, Yan Junjie, the founder and CEO of MiniMax, delivered a keynote speech titled "AI for Everyone" at the opening ceremony.

The key points are:

1. As models get better and better, artificial intelligence is gradually becoming a productive force in society.

2. An AI company is not just a replication of an Internet company. AI is a more fundamental and basic productive force.

3. There will definitely be multiple players in the AI field.

4. AI will become stronger and stronger, and this enhancement is almost endless.

5. A large amount of innovation can make AI R & D a less money - burning industry.

The following is the content of Yan Junjie's speech (abridged):

AI is evolving into a stronger productive force

In the past 15 years when I was engaged in AI research, every day when I was writing code for tasks, reading papers, and conducting experiments, I always wondered: What exactly is this highly - regarded artificial intelligence? What kind of relationship does artificial intelligence have with society?

As models get better and better, artificial intelligence is gradually becoming a productive force in society. For example, when we are doing AI research, we need to analyze a large amount of data every day. At first, we had to write some software to analyze this data. Later, we found that we could let AI generate a software to help analyze all the data. As a researcher, I am also very concerned about all the progress in the AI field every day. At first, we thought about whether we could develop an APP to track the progress in various fields. Later, we found that it was more efficient to use an AI Agent for automatic tracking.

Yan Junjie. Source: The interviewee

While AI is gradually evolving into a stronger productive force, it can also produce stronger and stronger creativity. For example, 15 years ago, the mascot of the World Expo held in Shanghai was called "Haibao". In the past 15 years, Shanghai has achieved all - round development. If we want to use the "Haibao" IP to generate a series of derivative images with more Shanghai characteristics and in line with the current trend, AI can do a better job.

Another example is the recently very popular Labubu. Previously, it might take about two months and cost tens or even millions of RMB to produce a creative video of Labubu. With the increasingly powerful AI video models, a similar video can be generated in basically one day at a cost of only a few hundred yuan.

Through high - quality AI models, most of the content and creativity on the Internet will become more and more accessible. The low threshold allows everyone's creativity to be fully unleashed.

In addition to releasing productivity and creativity, we have found that the use of AI has exceeded its original design and expectations, resulting in various unimaginable application scenarios: such as deciphering an ancient character, simulating a flight, designing an astronomical telescope... With only a small amount of collaboration, everyone's ideas can be turned into reality.

Facing so many changes, an idea began to emerge in my mind: An AI company is not just a replication of an Internet company. AI is a more fundamental and basic productive force, which continuously enhances personal and social capabilities.

Two key points are: First, AI is a kind of ability. Second, AI is sustainable. It is difficult for humans to break through biological laws, such as learning new knowledge non - stop and continuously becoming smarter, while AI can do it. When we build better AI models, we also find that AI is progressing with us humans and helping us create better AI. Inside our company, employees used to write a lot of code and conduct a lot of research experiments every day. Now, about 70% of the code is written by AI, and 90% of the data is analyzed by AI.

In this situation, how does AI become more and more professional? One year ago, training models still required a large amount of basic annotation work to be done manually, and annotators were an indispensable profession. This year, professional AI can complete a large amount of mechanical annotation work, and annotators can focus on more valuable expert - level work to help the models become better. Annotation work is no longer simply giving an answer to AI, but teaching AI to learn the human thinking process, so that the AI's capabilities become more generalized and closer to the level of top human experts.

AI has now advanced to the point where it can learn a great deal in an environment. In the past six months, as long as the environment can be defined and there is a clear reward signal, the AI placed in that environment to learn can gradually solve the corresponding problems.

Based on these observations, we have a very definite judgment: AI will become stronger and stronger, and this enhancement is almost endless.

AI will not be monopolized by a single organization

As AI has an increasingly greater impact on society, will it be monopolized in the future? Will it ultimately be in the hands of a single organization or multiple organizations?

We believe that there will definitely be multiple players in the AI field. There are three reasons:

First, all the models we currently use rely on model alignment. Obviously, the alignment goals of different models are different. For example, if the alignment goal of a model is to be a reliable programmer, then it will be particularly strong as an Agent; if the alignment goal of a model is human - interaction, then it will have a relatively high emotional quotient and be able to conduct smooth conversations; some models may be full of imagination... Different alignment goals reflect the values of different companies or organizations, which will ultimately lead to very different performances of the models and enable different models to have their own characteristics and exist in the long - term.

Second, the AI systems we have used in the past six months are no longer single models, but multi - Agent systems, which involve multiple models, and different models can use different tools. This will make AI more intelligent and capable of solving more and more complex problems, but it will also lead to a result that the advantage of a single model in a multi - Agent system will gradually weaken.

Third, in the past six months, many very intelligent systems are not owned by large companies. The reason behind this is that in the past year, open - source models have emerged like mushrooms after rain and become increasingly influential. In the past year, although the best models are still closed - source, there are more and more good open - source models, and they are constantly approaching the best closed - source models.

Based on these three reasons, we believe that AI will definitely be in the hands of multiple companies. Meanwhile, we believe that AI will definitely become more and more accessible, and the cost of use will also become more controllable.

In the past year and a half, we have had more available computing power, but the size of AI models has not changed significantly. The reason behind this is that for all practical models, the computing speed is a relatively critical factor. If the computing speed of a model is very slow, it will reduce users' willingness to use it. Therefore, all companies focus on the balance between the number of model parameters and the intelligence level.

Previously, the growth of models was basically proportional to the progress of chips. The progress of chips doubles every 18 months, and models also maintained a corresponding growth trend. Now, although everyone has more computing power, the model parameters have not become larger. So where does this increased computing power go? I think there are the following points:

First, in terms of training, in the past six months, the growth rate of scale has become relatively slow, and the cost of training a single model has not increased significantly. More computing power is spent on research and exploration. Research and exploration depend not only on computing power but also on efficient overall experimental design, an efficient R & D team, and some genius ideas.

This leads to the fact that the difference in training between companies with a large amount of computing power and those with less computing power may not be very large. Companies with less computing power can make their experimental exploration more efficient by continuously improving their experimental design, thinking ability, and organizational form.

Second, at the inference level, in the past year, the inference cost of the best models has dropped by an order of magnitude. We believe that within the next one or two years, through a large number of computing network systems and optimization algorithms, the inference cost of the best models may be reduced by another order of magnitude. In summary, the cost of training a single model will not increase significantly.

A large amount of innovation can make AI R & D a less money - burning industry, but the use of computing power will still increase. Although tokens will become very cheap, the number of tokens used will increase significantly. Last year, a single conversation of a ChatBot only consumed a few thousand tokens. Now, a single conversation of an Agent may consume millions of tokens. And because AI can solve more and more complex and practical problems, more and more people will use it in the future.

Making AI affordable for everyone is our judgment on the development of AI. We believe that AGI will definitely be achieved and will definitely serve and benefit the general public.

If AGI is achieved one day, the process must be realized by AI companies and their users together, and this AGI should belong to multiple AI companies and a wide range of users, rather than just a single organization or company.

This article is from the WeChat public account "China Entrepreneur Magazine" (ID: iceo - com - cn). Author: Kong Yuexin, Editor: Ma Jiying. Republished by 36Kr with permission.