The rules of the game for large models: it’s not about specialization, but the base that determines life and death

In the past year, a trend of "professionalization" has emerged in the field of AI - medical big model, education big model, legal big model... It seems that as long as you add the word "big model" to a certain vertical industry, you can immediately have a technological moat, capital support and market voice.

But is it really like that?

In the face of more and more practices and cases, the answer seems to be gradually emerging: companies that focus on "professional models" may be taking a path that is destined to be difficult or even wrong.

The big model is not a "module" but a "base"

What is the biggest feature of a large model? It is not that a certain function is particularly strong, but that the underlying capabilities are extremely powerful and the generalization ability is extremely amazing . The emergence of ChatGPT is the most intuitive proof of this point.

Whether it is writing code, editing essays, translating, solving math problems, analyzing legal clauses, or simulating psychological counseling conversations, it does it all in an orderly manner. More importantly, it is not a "professional model" trained separately for these tasks , but a "generalist" that is constantly polished on a powerful and unified base model through massive data, complex training, and RLHF (reinforcement learning with human feedback).

And this is the biggest weakness of the "professional model": if your basic model is not strong enough, there is no question of being "professional".

The pseudo-barrier of “professional model”

Some people will say: "We have industry data, and doctors, lawyers, and teachers are involved. We understand vertical scenarios better than general models."

It sounds reasonable, but it doesn't stand up to scrutiny.

First of all, data is not a barrier, especially in AI, a game that depends entirely on data quality and scale . If industry data is not cleaned, labeled, and structured with high quality, it cannot be transformed into the "cognitive ability" of the model.

Secondly, if a "professional model" only makes some "fine-tuning" to the general model or adds some prompt engineering, then to put it bluntly, it is just a "plug-in" based on ChatGPT, Claude, and Gemini, and does not have real technical independence.

Moreover, these general-purpose large models already have very strong professional capabilities . For example, GPT-4’s medical test scores are close to or even exceed the level of American residents, and its problem-solving ability, logical ability, and cross-domain knowledge integration ability are far superior to most individually trained “medical models”.

The big model is that everyone prospers together and everyone suffers together.

We are standing at the entrance of an era of "platform-based intelligent entities" - the future AI will not be a collection of functions, but a cognitive system and an operating system.

From this perspective, the development path of the big model is very similar to that of the Internet: the winner takes all, and the strong will always be strong.

Whoever has the strongest basic model capabilities can continue to expand the boundaries and penetrate into more industries, scenarios and user needs. Players who try to "avoid the main channel" and break through with "small and beautiful" "professional models" are often just telling a "seemingly reasonable" story if they do not have real technical originality or irreplaceable data.

ChatGPT has "eaten" countless "professional models"

Let’s take a brief look back:

• After GPT-4 was launched, a large number of AI tools focusing on code generation became unpopular.
• Although large-scale educational models are just getting started, ChatGPT can already provide detailed teaching feedback.
• While the legal big model is still polishing the knowledge graph, GPT can already help users write contracts and analyze cases.
• The medical big model is still seeking hospital cooperation pilots, and ChatGPT has been implemented on multiple overseas health platforms.

What we are witnessing is not a "hundred flowers blooming", but a "grand unified" reconstruction of the intelligent ecosystem. The big model is not about "specialization", but about who can become an "expert among generalists".

Last words

It’s not impossible to make a professional model, but its existence must be based on a stronger, more stable and smarter basic model.

Otherwise, the so-called "professionalism" is nothing more than a gimmick that can be easily crushed by general large models; the so-called "barrier" is nothing more than a self-satisfying bubble.

The wave of technology will never slow down for the sake of "small and beautiful". Those who can truly win this AI competition are not "segment players" who set limits for themselves, but "platform players" who dare to build a world-class foundation.

ChatGPT has given us a clear answer.