The CEO who was bedridden due to a fracture increased efficiency by 100 times. Here comes the internal verbatim transcript.

From recognizing the professional and vertical capabilities of AI that far exceed those of humans, to understanding the underlying logic of large models and mastering the golden rules for using AI effectively, then to unlocking the core design and full - scenario implementation methods of Lobster, and finally pointing out the future direction of AI - native organizations, Fu Sheng has covered it all.

Whether you are an entrepreneur or startup founder aiming for ten - fold growth, or a manager looking to improve work efficiency, you can find directly implementable methods here to quickly get started with and make good use of AI, seizing the opportunity in the new era.

Below are the key points of this sharing. I hope it will inspire you.

Before the Spring Festival, we set our company's core strategy - fully centering around EasyClaw.

As early as when Manus emerged in 2025, I was certain that agents would see large - scale popularization. I firmly believed that the local version of Manus deeply integrated with computers must be the future direction, and we also invested nearly a year in research and development for this. When I saw the Openclaw product, I slapped my thigh in realization. It turned out we were on the same path.

Today, I'd like to share with you all my practices and thoughts during this period without reservation.

I. Face the Present: The Capabilities of AI Have Already Surpassed Your Perception

Nowadays, when people scroll through their WeChat Moments every day, they always come across various remarks about AI. I'd like to clarify a core fact with you first: Today's AI has comprehensively surpassed humans in logical reasoning ability.

1. Professional Abilities: In Closed - ended Questions, AI Has Set New Human Records

There is a very challenging test called GPQA (Note from Notesman: A high - difficulty Q&A benchmark used to test the ability of artificial intelligence models to solve complex problems, which can be understood as a "graduate - level exam" for AI).

You can think of it as a professional ability test covering all disciplines, with only multiple - choice questions and no subjective questions.

With the help of search engines, human doctoral students can achieve an accuracy rate of about 65 - 70% in their professional fields. If it's a non - professional field and without the help of search engines, even a doctor can only score 30 points.

I've seen this set of questions myself. I know every single word, but when they are put together, I have absolutely no idea what they mean. The difficulty is extremely high.

However, in the first half of 2025, the o1 to o3 versions of GPT - 4 scored over 80 points in the GPQA test. More importantly, when AI scores 80 points, it's not just in one subject, but all subjects meet the standard.

In front of AI, there is no such concept as "professional barriers". This means that when facing a closed - ended question with clear boundaries, AI's completion rate, rigor, and comprehensiveness have exceeded those of most people.

There is now a question bank called "the last test for humans". Currently, AI can only score 20 - 30 points in it. It is a fortress that AI has not completely conquered.

Why is it called "the last test for humans"?

Because this set of questions is the most difficult that human intelligence can design. If AI can conquer this set of questions, humans won't be able to come up with more difficult questions to test it. This is the current development status of artificial intelligence.

2. Vertical Fields: AI Has Achieved Breakthroughs in Core Production Processes

Many people's discussions about AGI (Artificial General Intelligence) always get stuck in the anxiety of "whether it will eliminate humans". Regarding when AGI will be achieved, some say in 3 years, some say in 5 years, and some say it will never be achieved.

But the core problem is that so far, the academic community has not given a clear and unified definition of "intelligence" and "AGI".

Academician Zhang Bo, the 90 - year - old founder of artificial intelligence in China, once said that artificial intelligence is exploring in uncharted territory because we can't even accurately define "what is intelligence".

If we define AI as "the ability to surpass humans in a certain field and continuously iterate", then it has achieved this goal in many fields.

The first field is autonomous driving.

I had an in - depth experience of Tesla's FSD (Full Self - Driving) in the United States. On the old highways in Los Angeles, it hardly needed any human intervention throughout the journey. In Shanghai, I also drove it for two hours, and its visual autonomous driving ability has reached a critical breakthrough point.

Those who haven't experienced it may think it's a boast, but those who have truly experienced it know that this day is not far away.

The second field is programming.

When GPT first came out, I hadn't touched code for 20 years and didn't even know what Python (a high - level programming language) was. I also did poorly in the information management major in college.

But by chatting with AI, it taught me step by step to choose a language, set up an environment, and write code. Finally, I wrote a complete Snake game program. At that time, I posted a video saying "the time left for human programmers is running out".

How powerful is AI's code - writing ability now? When you give it a requirement, the code it writes is more structured and concise than that of many ordinary programmers.

I publicly said in the company that the programmer position will definitely be restructured. Just like during the Industrial Revolution, the blacksmith profession inevitably disappeared. This is the trend of the times and is not subject to human will.

The third field is office work and tool usage.

We started using Lobster before this Spring Festival. The biggest feeling is that it is a tool that grows with you.

Using it on the first day and the tenth day is like using two completely different products, just like watching an employee grow with you. More importantly, it can call almost all the tools on the Internet, and its ability boundary is infinitely broadened.

3. Impact on the Workplace: AI Is Restructuring Positions and Organizational Hierarchies

A very obvious phenomenon has emerged in Silicon Valley: Large companies' profits are rising, but they are continuously laying off employees. The essential reason is that AI is replacing white - collar workers and middle - level managers in the traditional sense.

Let's start with white - collar workers. Sam Altman said two years ago: AI will replace those who sit in front of computers.

The core of this statement is that if your job essentially involves translating requirements into outputs that computers can recognize - your boss gives you a requirement, and you write an article, make a PPT, create a webpage, or write code - then your job is losing its meaning.

AI is inherently the best translator. Its earliest ability was text - to - text translation, and now it has extended to text - to - program, text - to - picture, and text - to - video.

The essence of most white - collar workers' jobs is "requirement translation", and AI can do this better, faster, and at a lower cost than humans.

I used to need a special team to make a PPT, and it would take a month of revisions to get the final version. Now, with AI, in just over a day, without even opening the PPT software, I can get a finished product of more than 50 pages. The efficiency difference is huge.

Now let's look at middle - level managers. There is a very popular article on Weibo called "AI Is Eliminating Middle - Level Managers". What is the core function of middle - level managers? It is to convey information up and down, decompose instructions, and follow up on progress.

There is a classic theory in Western management: One person can manage at most 8 - 12 people because human energy and communication ability have limits. So, enterprises have to build multi - level organizational structures, and middle - level managers become the core nodes for information transmission.

However, today, AI's efficiency in information organization, transmission, and coordination far exceeds that of humans. In the future, an organization can completely have a manager directly lead a group of front - line employees without so many intermediate levels.

This is the new possibility of organizational form in the AI era.

II. Underlying Logic: Understanding Large Models Is the Key to Using AI Effectively

Here is an underlying logic: Only by understanding large models can you truly use AI effectively.

Many people can't use AI well and always think AI is "stupid" and doesn't understand their needs. The core reason is that they don't understand the underlying principles of large models.

Only when you know why it can do something and why it makes mistakes can you truly control it.

1. The Essence of Large Models: Probability Prediction of the Next Word

You must remember: No matter how human - like AI seems, it is not human. Its core logic is always the generation and prediction of the next word.

When we see AI outputting words one by one, it's not just a display effect. It is really making predictions word by word.

When you ask it a question, it activates the corresponding text vector. Based on all the previous content, it calculates through weights to predict the next most likely word, then adds this word to the previous content and continues to predict the next one, and so on.

This is why long texts and long conversations consume more GPU computing power. It's not a simple addition of the number of words, but an exponential multiplication operation.

Now, most short - video generations only last for 15 seconds, based on the same logic: Each second of generation requires calculations based on all the previous seconds. The cost of an additional second may be the sum of the previous 10 seconds.

The amazing thing is that the content predicted word by word by AI finally forms a sentence that you can understand and that fully meets your needs. This is the emergence of intelligence.

So far, there is no complete mathematical solution to why large models generate the emergence of intelligence. We only know that if you feed it enough text, it will show intelligence. This is the scaling law.

It wasn't until GPT - 3.0 and 3.5 came out that the entire industry truly verified this law. And to this day, we are still exploring on this path, and to a large extent, we rely on the belief that "it can continue to make breakthroughs".

It is also based on this core principle that large models show extremely strong multilingual abilities.

In the past, when doing AI translation, you had to give it corresponding Chinese and English texts for it to learn word by word.

But now, for large models, if you feed it enough English corpora, it can master English proficiently. Then, if you feed it a small amount of Chinese corpora, it can naturally achieve Chinese - English translation, even without corresponding texts. If you feed it Japanese corpora, it can also master Japanese translation.

Because through a large amount of corpora, it has mastered the positional relationship and semantic association between words and has constructed a cognitive understanding of the world through language.

2. The Golden Triangle for Using AI Effectively: A Good Model, Good Context, and Good Questions

These three elements determine the final effect of your use of AI, and none of them can be missing.

① Choose the Right Model

Models are not omnipotent. Different models have different areas of expertise. We can divide them into three tiers:

First Tier: Claude Opus

It is at the level of a Stanford doctor and is currently the top - notch model.

You can entrust it with complex reasoning, in - depth analysis, and task planning without having to repeat the process. My personal experience is that using Opus is like having a different person helping you. It's expensive, but it's worth the price.

Second Tier: Sonnet / GPT - 4o

It is comprehensively sufficient. For programming, GPT is the first choice. It can fully handle daily work, writing copy, and data processing, and it has high cost - effectiveness.

Third Tier (for frequent use): Domestic + Flash + Haiku

It is cheap, but the gap is obvious. It can be used for simple tasks, drafting, and batch processing, but don't expect it to handle complex tasks - there is a visible gap in its ability ceiling and hallucination control compared to the first - tier models.

Do you want to use a Stanford doctor - level model or just go for the cheap one?

Our EasyClaw has a wide variety of models to choose from. You can access them all through one entrance and switch between them at will without having to register accounts everywhere.

② Manage the Context (Context)

Many people think that the more context and longer memory you give to AI, the better its performance will be. This is completely wrong.

The core logic of context is that the more content you load for it, the lower the weight of your current words and requirements in the entire context, and the more difficult it is for it to focus on your core needs.

It's like a person whose mind is full of small talks with relatives. When you talk to him about serious matters, it's difficult for him to focus.

The natural defect of large models is that they don't have real long - term memory. Even though many products now have a long - term memory function, when using it, they still extract part of the content from long - term memory and put it into the current context.

Therefore, it's not that the more context the better, but rather it should be precise and focused:

In daily use, try to streamline the conversation history and clean up invalid content in a timely manner;

The content closer to the task execution is more important. Before starting the task, repeatedly confirm and clarify the core requirements with it to make the key content have a high weight in the context;

Even though Claude supports a context of one million tokens, don't use the full amount in daily use. Less but more refined is always better than more but more complex.

③ Ask Good Questions

How good an answer AI can give completely depends on how good a question you ask.

Large models are trained with hundreds of billions of parameters and dozens of terabytes of corpora. They have the knowledge