With the recent release of more and more Agents products, as well as the amazing performance of chatGPT-4o image mode and o3, there are more and more discussions about models as products and even models as agents, which also triggered some thoughts on what we should do in the era of big models.

Based on some recent product launches, I would like to make an update on the cognitive level. The content may not only be about the model as the product, but also some thoughts on what products, organizations, current enterprises and even individuals should do in the era of AI big models. I would like to add that technology is developing too fast, and many current opinions are likely to be wrong, but at least they can bring some inspiration.

What is Model as Product talking about?

This issue seems to have been discussed when chatGPT was first released. Everyone found that the interaction and page of chatGPT were very simple. The function and value of the entire product actually lie in the big model. It is the ability of the big model that determines the logic of this product, and the big model brings value to users.

Especially if we think with the historical experience of the Internet era, we will be more uncomfortable with this change, because we are used to various innovations at the product and operation levels. The big model completely breaks this cognitive inertia. We need to rethink what is more important in the big model era.

This is also the point where everyone started to debate about shells, and even now there is still a debate, because if you are a shell, then what is your core ability? What do you have left after removing the model?

My feeling is that everyone seems to be right, because large model technology has been changing rapidly in recent years, whether in terms of capabilities, costs or even technical routes, different stages may require different things.

Why do people start to focus on the model as product at this point in time? It seems that the technology has basically converged to the route of large-parameter pre-training + RL post-training, and everyone has found that this path seems to be generalizable to a wider range of scenarios.

In the pre-training stage, give the big model some prior foundation and knowledge, and then find the right environment and positive function in the post-training, so that the model can explore solutions by itself, even solutions known to humans.

The rapid development of agents is also promoting this understanding. Agents are models + tools, which is a general background that everyone basically agrees on. After all, we humans have completely parted ways with animals after we were able to use tools.

The use of tools has actually evolved. At first, everyone discovered that the model has the ability to write code, so they tried to let the model write code to use the tool. Later, OpenAI introduced function calls and GPTs to better help the model use tools. With the recent MPC and A2A standard protocols, how we can better use the tools has become more and more mature.

Regarding the model usage tool level, I feel that what is more important is not the quantity, but how to use it more smoothly. The protocol actually solves more problems of the quantity we can use and the data exchange between agents.

What tools are used for the model and when they are used may have a greater impact on the final effect of the agent. We see that o3 is actually a bit like an agent. It can call some tools based on user needs, such as code interpreters, retrieval, etc.

We can clearly feel the difference in effect. This kind of model that directly targets the tool for RFT in the post-model training stage is much better than using a large model to perform engineering arrangement at the prompt word level.

Just like a person cramming for an exam and truly applying what he has learned are completely different.

Therefore, on the eve of Agent’s explosion, everyone has begun to refocus on the relationship between models and products, because this determines how you will act as an Agent and how you will compete.

Organizational changes

This is a video about Research x Product released by OpenAI a year ago , which introduces how close cooperation between R&D and product teams can bring better performance and experience. The speakers are the head of the post-training team and the product manager responsible for model behavior.

They introduced the specific work related to post-training and products, and how they collaborated. From this process, we can also get a glimpse of some changes in the organizational structure under the new AI technology.

If we follow the previous thinking inertia, an AI company will probably have two major departments, one is the product department and the other is the model training department. These two departments are basically independent, and there should not be too much interaction between the two sides, because one is responsible for training models and the other is responsible for packaging and designing products.

This approach will have problems with the current development of AI technology, because it is easy for the product department to use some engineering methods to solve the model problem while the training team directly solves the problem, or the goals set by the training team cannot meet the needs of the product team to serve users, etc.

So in this speech, the head of OpenAI's product model behavior posted this picture

We will find that regarding models and products, OpenAI not only does this in terms of product and model thinking, but also designs the entire organizational structure in this way.

The design of the organizational structure is essentially the specific implementation and implementation of the company's strategy. From the collaboration between product and model departments in some current AI companies we mentioned earlier, we can see that there is a problem with the understanding of this point.

In the actual final product performance, they did a lot of work and finally found that after OpenAI was released, all their customers left. From their perspective, OpenAI is a company that provides models, but in OpenAI’s own cognition, the model is the product.

In this video, the product manager also mentioned three special aspects of model and product collaboration:

• Extremely vague, because the technology is advancing too fast, and the model has strong generalization ability, in fact, many dimensions cannot formulate strict rules and logic, just like we have seen OpenAI news recently mentioned that chatGPT has been a bit too flattering recently, in fact, it is also the same logic, that is, how the model performs more reasonably, in different scenarios and different people's understanding is completely different, so they can only get feedback through quick actions, learning by doing . So we often see OpenAI release some semi-finished products, it's not that they don't know, it's just that this is their strategy
• Technology-driven. We have seen that many product innovations may be driven by specific scenarios. However, the current AI big model technology is too powerful and is developing and evolving too fast. Therefore, we cannot just start from product needs, but need to start from technology. We cannot avoid the feeling of using a hammer to hit a nail, because the hammer has evolved too fast, from a hammer to a Swiss Army knife. You cannot turn a blind eye to this development. Therefore, the product department needs to think more about how to achieve the so-called TMF, how to better match technology with corresponding demand scenarios, and how to better present technology.
• Research cooperation, and finally, closer cooperation between products and technology. The work of the two departments needs to promote each other. The needs and data collected by the product need to be able to feed back model training with technology, and new model features also need to be better designed to provide users with use. These two things need to be done at the same time, but they are completely separated. It is a process of mutually promoting iterative development.

Under the current trend of technological development, companies that make AI products cannot simply consume the API of model companies. In other words, simply wrapping is definitely not enough. Of course, wrapping can be used as your phased strategy to develop rapidly, but from a longer-term development perspective, new technologies not only require thinking about specific product forms, but also require new paradigms and changes in organizational and collaborative levels.

Product working differences

Kevin Weil, Chief Product Officer of OpenAI, mentioned in a recent interview, “We try to keep it lightweight because it is impossible to be completely correct. We will abandon some incorrect practices or research plans halfway because we will continue to learn new things. We have a philosophy called iterative deployment. Instead of waiting until you fully understand all the capabilities of the model before releasing it, it is better to release it first, even if it is not perfect, and then publicly iterate. "

This new way of product update and iteration actually poses new challenges to the work of product managers.

• You need to be more proactive, because many product features and performance may require you to take the initiative to discover, there are too many problems to solve, and more exploration is needed to understand the technology and create valuable products based on more feedback.
• Adapt to ambiguity. Because of the ambiguity of technology, you need to understand and adapt. At the same time, how to better digest the ambiguity of technology and then provide users with more certainty at the product and scenario levels. You cannot wait for the answer to appear automatically, nor can you expect to get a perfect and completely certain answer.
• Finally, you need to act quickly because everything is developing rapidly and is vague. The only thing that can provide certainty comes from more user feedback. You need more exploration to establish a clearer cognition and understanding. Only by getting started yourself can you really feel the differences and characteristics of different technologies.

At the same time, for product managers, how to effectively establish a connection and balance between technology and users is also crucial. You need to think about how to better present the characteristics of technology to users, and how to make the best use of strengths and avoid weaknesses in specific scenarios to truly solve problems and create value for specific users.

I saw an interesting point of view before. In the AI ​​era, what and how may not be the most important, but why is the most important.

Because you need to go back to the starting point and think about why you need to do this and whose problem you are solving.

At the same time, based on the new technical paradigm, how to effectively define and evaluate problems is one of the most core capabilities of product managers.

Moving forward in blur and chaos

Sutton's recently released " Welcome to the Era of Experience " claims that people are standing on the threshold of a new era of artificial intelligence, and are expected to reach unprecedented levels. Like his previous classic "Bitter Lessons", we imagine that we are entering a new era

However, in this new era, we also have various problems. Bitter lessons seem to have boundary problems. As mentioned in the Moravec paradox, although we have made great progress, there are still more problems waiting for us.

Moravec's Paradox refers to the counterintuitive phenomenon observed in the field of artificial intelligence (AI) and robotics : things that are easy for humans (such as walking, identifying objects, grabbing a cup, and understanding basic language) are very difficult for artificial intelligence and robots . Things that are difficult for humans (such as playing chess, performing complex mathematical calculations, and logical reasoning) are relatively easy for artificial intelligence .

Of course, the new paradigm based on LLM+RL does bring us more exciting hope and dawn. We also need to evolve from the "human data era" that imitates humans to the "experience era" that surpasses humans.

What we can do

For startups, the development trend of model-as-product seems to be particularly unfriendly to startups. After all, large-scale model technology requires more capital and the threshold is too high. However, I think the viewpoint in a previous podcast is quite inspiring. Startups cannot use end-game thinking , that is, you cannot think about how to compete with large companies and barriers right from the start. You need to think about how to survive longer with your limited resources and how to move forward step by step. However, what is the technical route and product form, and how to stay at the table is the most important thing.

For enterprises that need to integrate AI into their specific businesses, the most important thing is to better accumulate past experience and data and combine them with big models. Most of the successful experiences and paths of enterprises may still be in the minds of some people. How to effectively sort them out and continuously amplify this value through the capabilities of AI is the core issue.

For each of us, volume intelligence may no longer make sense. Whether it is us or our next generation, it is more important to raise questions rather than solve them.

As an independent individual, your experience, taste, and independent thinking are the most important part. You can create greater value by better collaboration and integration with AI.