Intelligent degradation

Recently, I came across the most interesting piece of news, which is roughly as follows:

When I saw this content, I really burst out laughing.

This actually means that people have spent a lot of time creating so - called agents, but all they've created is negative value.

Why is this the case?

The reason is very simple.

People from the past R & D camp always think that they can improve AI's performance in a specific area by adding what they think are useful knowledge (in the form of prompts, which are essentially rules) to make AI better.

These so - called "human knowledge" and "little tricks" are presented as prompts, but in essence, they are a bunch of rules for AI.

These things can help improve the accuracy under limited goals, but for a large - scale model, they are actually harmful and a kind of "intelligence degradation" behavior.

Where does the power of a large - scale model lie?

It lies in the fact that it has learned a vast amount of data and formed an extremely complex probability model that simulates the real world inside. It has an "emergent" general intelligence that we haven't fully understood yet.

The rules you add are like forcing an artist with infinite imagination to use a children's coloring book and requiring him to color within the lines without going out of bounds.

You think you're "optimizing" his ability to draw an apple, but actually you're ruining his chance to create a masterpiece like "The Starry Night".

When you're faced with demands that are vast, diverse, and infinitely close to the real world - like the daily work of a lawyer - your so - called "local optimization" becomes a losing proposition.

The loss caused by the part of general intelligence you've cut off is greater than the benefits you've gained from the part of the demands you've laboriously matched.

Ultimately, it boils down to three words from users: Not useful.

This is "intelligence degradation", which is a typical trap in creating agents now.

What's even more troublesome is that the faster the large - scale model itself and scaffolding like general search progress, the deeper the trap gets.

What to do?

To avoid the huge pitfall of "intelligence degradation", the core is just one sentence:

Stop trying to teach AI "how to think", and instead, give it "materials for thinking".

(The underlying principle is the intelligence - first principle we started talking about in 2023)

Humans have to admit that AI's "brain" (the base model) is already better than humans in terms of un - attributed intelligence. So, don't try to be a poor "teacher", but be its "intelligence officer".

Provide it with high - quality, exclusive "intelligence" that it couldn't access originally, that is, data and context. Then trust its intelligence and let it reason and judge on its own.

Of course, since you need to know more about what you want to do, you need a relatively complex evaluation system.

(This part is very complex and won't be elaborated on in this article)

Following this line of thought, the valuable directions for agents are actually quite clear:

Direction 1: Dig deep into "exclusive context"

General large - scale models understand the public domain.

They don't know what meetings your company held last week, who your most important customers are, and how your flagship product was developed.

These are your "exclusive context". This is your only and strongest moat.

As we used to say: The boundary of data is the boundary of application.

The primary value of an agent is to safely and efficiently feed the large - scale model with the internal, private data scattered all over the company.

To put it simply, it's like giving AI "intranet access". Let it see all the emails, chat records, meeting minutes, code libraries, product documents, and data in the customer relationship management system (CRM).

When AI can see all this, it will no longer be a "mouthpiece of the Internet" that just talks nonsense, but a "digital employee" that truly understands your business.

What humans need to do is not to use tricks to improve intelligence, but to make up for the lack of digitalization. If we dig deeper into this, it involves digital costs and production relations. The content often mentioned by Comrade Chen Guo is very useful here. For example:

What kind of product form can win? Let's look at Glean

After talking so much, what kind of products can avoid the pitfall of "intelligence degradation"?

We can compare two different types of products on the market, and it will become clear at once.

The failed form: The $50,000 contract AI

I call this type of product "workflow AI".

Its obvious problem is the lack of flexibility, and the lack of flexibility is due to the lack of in - depth integration.

Its logic is to preset an "analyze contract" process within a closed software, and then let AI fill in some parts of this process.

When people can interact with AI frequently, this workflow basically brings more harm than good.

Its problems may be:

1. Lack of context: It only knows the one contract you uploaded. It doesn't know the background of this contract, the negotiation process (which may be in the emails), or the relevant historical contracts (which may be in another folder). It's "blind".

2. Rigid process: It locks both you and AI into a fixed process. If a lawyer wants to ask questions from a different angle or let AI combine other information, it's impossible. This is a high - risk area for "intelligence degradation".

3. Isolated value point: Its value is limited to the single scenario of "analyzing contracts". It can't connect this ability with other workflows in your company.

The successful form: A product like Glean

The essence of Glean is a "context platform".

All its contributions are to ensure what we've repeatedly mentioned: In - depth understanding of reality.

It doesn't preset any processes. Its only goal is to break through all the data silos within a company, connect all the data in systems like Slack, Google Drive, Jira, Salesforce... to form a unified "enterprise knowledge graph" or "enterprise brain" that can be retrieved and understood by AI.

The most troublesome part of this work wasn't actually done well after the emergence of large - scale models.

What was called the data middle - platform in China back then is very similar to this work.

It's estimated that most of the companies that worked on data middle - platforms have failed, which also reflects the practical difficulty of in - depth AI application from one side.

Here is a picture of a middle - platform I randomly found. You can compare it with Glean's architecture:

(https://juejin.cn/post/6844904164292575246)

Here is the architecture diagram of Glean:

By comparison, there's a scary conclusion: To create an agent, first build a good data middle - platform...

If you can't build a good data middle - platform, the agent won't be useful either. At least, those involving production relations definitely won't be useful

Glean itself is the most awesome "intelligence officer". It doesn't teach AI how to think. It only feeds the most comprehensive and accurate "intelligence" to AI.

When an agent is built on a platform like Glean, it comes alive. You ask it: "What progress and risks were there with our most important customer, 'ACME Company', last quarter?"

A "silo AI" will be completely confused.

But an agent based on Glean does this:

1. Retrieve all the records of ACME Company in the CRM.

2. Retrieve all the emails and Slack chat records related to ACME.

3. Retrieve the internal meeting minutes and weekly reports about the ACME project.

4. Then, it uses the general intelligence of the large - scale model to synthesize these fragmented information and give a well - reasoned, insightful answer: "The progress is that the XX contract has been renewed, but the risk is that their key contact person recently complained in an email about our delivery delay, and the relevant discussion is recorded in the XX Slack channel."

You see, in this whole process, no one presets a rigid "customer risk analysis" process.

The form like Glean is less likely to have the problem of "intelligence degradation".

Because what it does is not "subtraction" (restricting AI with rules), but "multiplication" (broadening AI's vision with data). Its core value doesn't lie in designing amazing prompts, but in building an amazing "data path".

When the underlying large - scale model is upgraded from GPT - 4 to GPT - 5, that $50,000 "silo AI" may become useless.

But the value of Glean will skyrocket. Because a stronger brain combined with more comprehensive data can generate exponentially growing intelligence.

Again, I've simplified the above content to illustrate the direction. In actual development, dealing with cumulative deviations and other issues is a very complex process. It requires a combination of architecture, domain models, etc. I wrote some articles about this more than a year ago.

Summary: Intelligence - first and the unmanned company

Ultimately, this is a fundamental paradigm shift, a principle of "AI - first".

In the past, the thinking was "process - first". We designed the processes and let AI play a supporting role. This is essentially human - first, and the process is used to solidify certain expectations.

But "AI - first" is the opposite: We assume there is a smart "brain" in the center, and all our work is to build an environment for this brain to maximize its value.

As we used to say: Shift from the if - else thinking mode to the any - then thinking mode.

Pushing this principle to the ultimate form is the "unmanned company".

In the future, organizations won't rely on human employees to execute thousands of rigid SOPs (Standard Operating Procedures). Instead, they will encapsulate the entire company's business logic into a set of agent systems.

How to encapsulate? Not by writing rigid code, but by defining the following for the agent system:

1. Goals: For example, "Maintain the inventory of product A in the company between 1000 and 1200 pieces."

2. Context: Connect it to real - time sales data, supply chain data, and logistics information (in the Glean mode).

3. Tools: Give it the permission to call the procurement system API, send emails to suppliers through the API, and send early - warning messages to human supervisors through the API.

All these need a base, and this base is the data middle - platform that we almost gave up on.

Then, this agent system will run autonomously 24/7, make decisions on its own, and call tools to execute tasks.

The role of humans changes from executors to goal - setters and supervisors of the final results.

So, to avoid "setting rules" for AI and ultimately creating products with "intelligence degradation", first understand a basic question: What is AI and what does it do!

(It's quite embarrassing to say that after I've been writing about AI - native, AI - first, and the unmanned company for almost two years, many of my friends still don't seem to understand or agree very well)

This article is from the WeChat official account "Think about things", author: Lao Li talks about one, two, three. Republished by 36Kr with authorization.