Mid level disappearing, Token frenzy receding, Silicon Valley engineers’ eyes on the ‘second half of AI entrepreneurship’

The stage of savage growth has passed

This is the deepest feeling recently felt by Ma Peiyuan, who is on the forefront of AI entrepreneurship in Silicon Valley.

In December 2021, after graduating from undergraduate studies, he joined Quora, which is known as the “American version of Zhihu”, and later transferred to the AI chatbot aggregation platform Poe as a senior AI engineer. Two months ago, he joined Cognition, one of the hottest AI agent startups in Silicon Valley.

In May 2026, this company, which had been established for less than three years, announced the completion of over $1 billion in financing, raising its valuation to $26 billion. Eight months ago, its valuation was still $10.2 billion.

In addition to being an engineer, Ma Peiyuan also has another identity – a Venture Scout. He searches for AI entrepreneurial projects worth betting on for investment institutions, with an annual investment quota of $500000.

In the more than four years since graduation, it seems that he has always acted earlier than the consensus in the AI industry. Before the term “agent” was clearly defined by anyone, his team had already quietly initiated a project. When Anthropic was still unknown, they became Claude’s earliest deep users. He was the first AI certified engineer and the first AI agent engineer in the team, and later brainwashed the entire engineering team into an AI Native team using AI programming.

The most interesting evidence of this advanced cognition and action occurred two weeks ago when he casually stated during a conversation with the “Future Human Laboratory” that “people managers are disappearing”, which was soon confirmed by Tencent and ByteDance announcing the cancellation of mid-level news.

Changes happen too quickly, in Silicon Valley, we must always coexist with them.

During the two-week gap between the two interviews in this article, Ma Peiyuan found that the logic of Silicon Valley’s “pursuit of the most expensive and cutting-edge models” he had previously talked about began to falter. The current popular practice in Silicon Valley is to “merge scheduling” multiple cheap models and top models.

After chatting with Ma Peiyuan for nearly 5 hours, we found that he has been almost constantly overthrowing himself from yesterday, and Silicon Valley is no exception. I can’t say everything, but you might say it’s over 60%, and a month later it might be the opposite again

In the past few years, AI has almost rewritten all the default rules in Silicon Valley: what kind of people are more likely to stand out, what kind of organizations are more competitive, and what kind of money is worth spending.

Of course, these so-called ‘correct answer’ judgments are also rapidly becoming invalid. In the field of AI, one year is already a long cycle, and many judgments may be overturned within a month.

Here are 10 current observations of Silicon Valley AI in the eyes of Ma Peiyuan:

Ma Peiyuan

What kind of talent does Silicon Valley need?

Ma Peiyuan’s career portrait highly fits the favorite talent profile of AI companies in Silicon Valley. He is not from a traditional computer background, has not won any competition gold medals, does not have the aura of top universities, and even had no interest in computers at one point. However, in the past four years, he has gone from the advertising team of Quora to an AI engineer in Cognition, and his resume has hit every keyword in the talent portrait that Silicon Valley AI companies are most looking for now. His experience is both a personal growth history and a sample of changes in industry recruitment standards.

Observation 1: Generals crush specialists

The most sought after talent portrait in Silicon Valley now is not the expert who has delved the deepest into a certain field, but the person who “excels in every aspect and has no obvious shortcomings” – this is called a polyglot talent in the industry, or more commonly known as a “hexagonal warrior”.

In 2024, Poe launched an internally undefined “agent” project, and the reason why I was chosen as a newcomer in the team was very direct: “I have worked in both AI, full stack development, and iOS. At that time, most of the other members in the group had backgrounds in traditional machine learning and search advertising. Their people did not understand front-end code, and the agent happened to need to be able to write front-end interfaces. After the person in charge of AI prompt word engineering left the company, there was no one in the team who could fill this gap. That’s just right. Anyway, I am a quick learner and I am responsible for both the agent prompt word engineering and front-end coding.

This combination of “knowing a little bit of everything” abilities later became my most valuable asset when looking for a job. What new field can you give me that I can learn in about two weeks. And this is exactly what early-stage startups need.

Observation 2: AI native skills can crush seniority in a short period of time

The ability to use AI tools can directly surpass the qualifications accumulated by years in the traditional sense.

When I transferred from the Quora ad group to the iOS group, the reason was’ wanting to step out of my comfort zone ‘, but at this point, I knew nothing about iOS and had a complete’ 0-year experience ‘. I didn’t choose to make up for the missed classes slowly at that time, but instead made GPT-4 my partner. At that time, the people around me were all very experienced iOS engineers who may not believe in AI. I couldn’t write iOS or Swift code, so I crazily used AI to help me write code.

The result is that as a newcomer who has only been in the industry for three or four months, my output speed is on par with or even exceeds that of senior engineers who have been immersed in it for many years, and I have achieved a senior position within two years. This is almost impossible speed in the traditional promotion system. Previously, it may take at least 4 to 6 years for you to become a senior engineer at Facebook. Some people at Google may never upgrade to senior engineers in their lifetime.

I met a 19-year-old interviewer at Cognition. At first, I was a bit frustrated because a 26 year old person was being interrogated by a 19-year-old person, and I couldn’t answer many questions. Later on, I realized that this was just a sign that this company was indeed able to demote talent without any restrictions, and age was no longer the standard for judging talent in Silicon Valley.

Observation 3: Recruitment standards have been completely rewritten, and AI companies no longer use the same template for interviews

Before leaving Quora, I had a serious interview with nearly 20 companies, including 30 to 40 that I had called. At most, I am approached by over 30 headhunters a week.

After interviewing so many companies, I found that there is currently no unified interview template for hiring in AI companies.

But there are several obvious changes, the importance of LeetCode style algorithm questions is decreasing, and the best answer for AI system design is changing every month.

The interview structure is also changing. Online written exams and work trials have become a new combination. A work trial is a very short ‘probationary interview’. Some are 3 hours, some are 5 hours, and the longest is only one or two days. One interview I had was when I was given a code repository and asked to use any tool, including AI, to solve a practical problem. After completion, it is necessary to showcase the ideas and achievements. The whole process is very close to real work.

What’s even more interesting is that the examination of AI usage itself has been divided into two modes. Some rounds require the use of AI, it depends on how you “code” it; But there are also rounds that explicitly prohibit the use of AI, such as debugging, which tests whether you can identify errors without AI assisted bonuses. Essentially, it is testing your actual software engineering level and taste.

The interview itself is also diverse. For example, some companies ask me to conduct business analysis to determine which industry they should acquire.

The behavior round has also changed. It is no longer a traditional problem of “personal biggest weaknesses and strengths” or “how to solve conflicts with colleagues”, but rather a longer timeline to see a person’s growth trajectory, from high school to university, from internship to work, chasing down all the way, looking at your “life slope”, and judging whether you are a person who can exponentially increase.

One unexpected benefit of not having a standard answer is that no one can get an offer by answering questions. Even my Twitter followers on Cognition told me directly, “Don’t prepare, I don’t think there’s anything to prepare for

AI is restructuring organizational structure
In Silicon Valley, the most powerful people never belong to a single company, only to the ‘most powerful companies of that era’. Twenty years ago it was Google, ten years ago it was Uber and Robinhood, five years ago it was Coinbase, now it is an AI company, and Silicon Valley is a ‘silver market with flowing water and iron soldiers’. These talents screened out by the new standards will be more likely to stay due to a sense of mission rather than “treatment”. These organizations that gather the most powerful people are also undergoing tremendous changes under the reshaping of AI.

Observation 4: Excessive thirst for “AI native”

The desire of Silicon Valley companies for AI native is far greater than I imagined. Most of the companies I interview for are Series A, Series B, and there are also some seed rounds. A frequently asked question is: What AI tool do you use? I usually just throw my blog over, and the other party is basically stunned.

They will say that they not only want me to do AI application development, but also hope that I can drive the internal AI transformation of the company and make the team more AI native. Many companies are actually developing AI products, but their employees have not yet established their own AI workflows. Moreover, they also need people who spend a long time on Twitter every day and share some novel AI discoveries or efficiency improvement techniques in a timely manner.

This is a bit counterintuitive for me. Some of my past hobbies, such as tinkering with new tools like Claude Code every day and sharing AI usage in Slack (an office software), have instead become a new organizational role.

Previously on Quora, I successfully got the entire engineering team to use Claude Code. Whenever I encounter an engineer in the office, I always ask, have you used AI to program? I remember there was a network security engineer in the office who was initially very resistant, thinking that using AI to write code was too risky. I later asked him to change his mindset and not write code, only use AI to read code, understand code libraries, and do data analysis. Later on, he really started using it too.

Observation 5: “Better use of AI” is essentially an organizational transformation issue

When I was working on AI related work at Cognition, I increasingly realized a fact: better utilizing AI may no longer be a purely engineering problem, but rather an organizational transformation issue.

for instance. For Internet companies with traditional organizational structure, from discovery to repair, a bug often goes through several links, such as demand assessment, project scheduling, development, testing, and online. Writing code using AI may only take two to three hours, but the entire process often takes a month to complete.

This process used to be reasonable. Because fixing a bug may take an engineer three to five days, the value of the process lies in coordinating different teams and allocating resources. But AI has compressed this time.

Now, with the AI software engineer product Devin launched by Cognition, we can complete the work of the past few days in just two or three hours. In Cognition, we can even use Devin to find the person responsible for this bug and ask Devin to assist in fixing it. Many coordination and circulation links in between can be eliminated without any redundancy. Actually, AI is driving the transformation of the entire organizational structure.

Observation 6: People managers are disappearing, managers must be able to write code themselves

The so-called people manager is the engineering manager. In the United States, a prevailing view in recent years is that as a manager, you should do a good job in management, be more responsible for team coordination, organizational communication, taking care of employee emotions, providing growth support, and not have particularly strong engineering abilities.

But when it comes to cognition, I found that the entire organizational structure is quite flat. For example, the person I directly report to is even CPO, who is one of the co founders. He doesn’t care about your personal growth in detail, his role is more like determining the overall direction. It can even be said that in the engineering team of over 60 people at Cognition, there is only one true people manager, while the others, including CPOs, are both responsible for managing people and writing code. This is completely different from my previous experience on Quora. At that time, a manager may have to manage five or six engineers.

The Cognition team collectively put on red clothes to celebrate the birthday of a colleague who loves to wear red clothes on a daily basis

I think this may be an organizational change brought about by AI. I have observed that in any thriving AI startup, there is almost no manager who is only responsible for managing people – if he manages people, he must also have strong technical skills.

This judgment is not only happening in Silicon Valley, but also in the two-week interval between two interviews, domestic giants Tencent and ByteDance have announced the cancellation of middle-level positions.

In the era of AI, any judgment will be quickly overturned
The changes observed by Ma Peiyuan are not limited to the organizational level. A more macro industry judgment: “What kind of token spending is reasonable and what kind of model is a good model within Silicon Valley AI companies? ”It is also being overturned at a visible speed to the naked eye. This time, even the judgment he made in the first interview couldn’t withstand two weeks.

The AI industry in Silicon Valley itself is also overturning itself on a monthly or even daily basis, and there is no judgment that can keep it fresh for too long.