The Evolving Role of Foundational Models in Productivity Enhancement

I. Productivity Revolution

OpenAI's application of universal large-scale models through ChatGPT signifies the formal entry of the artificial intelligence field, which has been under development for over half a century, into a new era of widespread productivity enhancement. The productivity revolution brings about significant changes in the existing production relationships.

Machine substitution for human labor is not a new phenomenon and has a long history. From a societal perspective, the leap in productivity far outweighs the drawbacks, considering that things become more affordable. However, a certain portion of individuals will inevitably be liberated from their previous production positions, as the cost-effectiveness of their acquired production skills diminishes significantly with the emergence of machines. Regardless of the capital perspective, there will still be a group of people who become unemployed as a result. Although they are unemployed, the knowledge and skills they possess for their previous positions are still vital in the new production relationships constructed by large-scale AI models. The downside is that neither they nor their employers may realize this, and it requires individuals to have strong thinking abilities and habits of reflection.

Based on the inference from the previous interaction mode mentioned, there are two main stages in which humans play a role during the interaction:

1. Expressing their intention (aim).
2. Evaluating the output results (select).

Therefore, compared to individuals who only possess sufficient capabilities for decomposing actions, executing actions, and understanding output results, it is apparent that the cost-effectiveness of human labor is lower than that of large-scale models. Consequently, except for a few specialized positions, humans in such roles will gradually be replaced at the societal level. The liberated productivity will be redirected towards more "human" tasks - aiming and selecting.

Let's analyze specific industries to observe the changes:

1. The content industry

As Aim and Select gradually take center stage in the production process, the prominence of "individuality" within the production relationship starts to increase. This is because the aim and select actions are more directly driven by the subjective will of the action initiator, rather than being limited to considering the implementation process. Consequently, ordinary Aim and Select processes will be overtaken by the foundational model over time since average efforts are as unnecessary as directly using GPT for aim and select tasks.

Within the production relationship, those who lack their own methodology for Aim and Select will struggle against the foundational model. However, people will eventually develop a new evaluation system based on personal criteria, similar to today's User-Generated Content (UGC). This will result in works that were once considered average, scoring 60 points, failing to meet the new standards. Similarly, relatively outstanding works scoring 75-80 points will now be considered passing at 60 points. Meanwhile, consistently excellent works scoring 90 points or higher will stand out even more, amplifying the Matthew effect. Ultimately, this leads to an overall aesthetic improvement in society.

Simultaneously, while the reference system for entertainment remains unchanged, the distribution pattern of low-mid-high quality content will gradually evolve into a positive triangle. Content that lacks standardization in any aspect will gradually fade away. Just as it is challenging for us to write business trip tables within Excel today, delivering poor-quality work becomes highly costly in an improved infrastructure environment. Therefore, in terms of the application layer, our focus is likely to shift towards elevating the requirements of content creation infrastructure and enhancing operational efficiency.

2. The code industry

I would describe the code industry as the initial stage of product managers gradually disappearing (although it is unlikely that the role of product managers will change with the evolution of production methods and tools). This is because the primary task of a typical product is to understand and abstract requirements from the business side, then transform them into a solution, and communicate this solution to the technical development team.

Once the business side clearly expresses its purpose and requirements, the foundational model translates the purpose into a development project using semantic understanding and translation. It provides a development project plan, leaving only the project promotion tasks for the product manager. Although many conditions assumed in this process have a low probability of occurring in reality, at least they can still be achieved.

Moreover, in the process of writing code within a known system structure, the goal is to accomplish function Z in part Y using method X, synchronize the front-end and back-end to complete the build, and then run the code in a virtual environment to identify and fix bugs based on the feedback. This task, it seems, does not necessarily have to be performed by humans. Objectively speaking, both development and writing a product requirements document (PRD) involve implementing the concrete implementation after gaining a basic understanding of the matter. However, writing a PRD may only require typing and drawing, while writing code demands a deeper understanding of code logic and implementation methods.

The emergence of the foundational model makes it more cost-effective to "write code" and use code to build functionality. The foundational model excels in information integration and utilization, rendering the type of work that programmers previously jokingly referred to as "copy and paste" code more efficient and systematic.

As a result, most pure execution programmers will become less cost-effective (earning less than $20 per month). However, at the same time, those who previously focused solely on execution will be required to transition towards non-execution work. They will need a deeper understanding of business requirements, better abilities to construct code logic, and more experimentation with different approaches and prompts.

In this domain, the foundational model plays a similar role to the content industry, raising the bar for entry (which, of course, is an ongoing process). During this transition, individuals who do not meet the requirements will naturally be perceived by the market as less cost-effective. As the standards are elevated, more people will naturally be freed from purely execution-based work to fully leverage the "human" factor in the process: Aim and select.

3. In terms of application

Let's put forth a daring hypothesis: Will there still be a need for separate tool apps in the future?

Consider this hypothesis: In most daily scenarios, the majority of us have similar needs that can be fulfilled without requiring special methods or functions. If this hypothesis holds true, it naturally follows that a single feature capable of satisfying various needs would suffice for most individuals in their daily lives.

Therefore, it is possible that in the future, we may no longer need individual apps like GPT dialog box - Todolist, Calendar, Weather, Alarm clock, Meitu, Meituan, Eleme, Keep, and other apps that primarily provide fixed tool attributes for specific scenarios. Instead, we can simply instruct the foundational model to whiten a picture and adjust the leg position, and it can automatically perform these basic operations. Even in Dingtalk and Notion AI, task building and low-code app development are already being explored to meet the personalized needs of different individuals, rather than just functional requirements.

So, what will become of tool-based applications? Which applications will truly disappear in the future, and which ones will gradually transform into behind-the-scenes tools? How will people react to the advent of the foundational model era? Perhaps it requires individuals who are more insightful than me to contemplate these questions.

The aforementioned ideas are limited to our current understanding of daily work, while true pioneers will innovate and explore new implementation methods. Presently, the foundational model can only answer questions based on existing information and cannot create or illustrate information that is currently unavailable. We still rely on individuals to identify and expand the boundaries of the foundational model.

The shift of the foundational model from handmade to industrial does not imply the disappearance of handmade approaches, nor does it render previous handmade experiences meaningless. Those pioneers who are willing to explore improved handmade methods and more efficient industrialized production methods remain the driving force behind the industry's progress.

II. The Next Phase of the Foundational Model

The current state of the foundational model is not flawless; there are areas where we hope for improvements. It resembles a capable subordinate who excels at execution. When given instructions or assigned tasks, it performs them competently. Although there may be some minor issues that require monitoring the final outcome, overall, it understands what needs to be done and carries it out effectively. If we liken it to an employee in a familiar work scenario, we can consider it a reliable team member, but lacking in proactivity.

The current foundational model operates on a responsive basis rather than offering active suggestions. When providing solutions, it assigns a higher probability to scenarios that align with the majority of occurrences in the training set, even if the specific circumstances are not clear.

Additionally, the foundational model lacks the ability to discern when human judgment is necessary to limit potential output solutions. As a result, we can only provide feedback on the solution after it has been generated, rather than guiding the solution's production in real-time.

Therefore, in order to enhance the utility of the foundational model in the future, we need to focus on the following two aspects:

• Activeness: Encouraging the model to exhibit flexibility beyond mere implementation and avoiding erroneous judgments.
• Judgement nodes: Enabling the model to recognize situations where human judgment is required and providing guidance prior to the full generation of a solution.

If these requirements can be met, the realization of a sophisticated "AI personal secretary" is well within our grasp.