Silicon Valley AI startup adapts as token spending and management models evolve

Based on over four years of industry observation, Silicon Valley AI engineer Ma Peiyuan identifies ten significant shifts: talent hiring is moving from specialists to generalists, with AI tool proficiency transcending years of experience; organizational structures are becoming flatter, requiring managers to possess technical skills as the role of people manager fades; token usage is shifting from quantity-focused pursuit to efficiency-driven evaluation, with large companies beginning to reassess their spending; the cult of model supremacy is being overturned, as Silicon Valley increasingly favors orchestrating multiple affordable and top-tier models; new AI infrastructure is emerging, and vertical AI agents retain investment value by addressing “untrainable” real-world pain points. Ma emphasizes that one year is a long cycle in AI—most assumptions can be overturned within a month.

Article author and source: 36kr

All consensus is rapidly expiring.

Everyone is trying to form judgments about AI, but all consensus is rapidly becoming outdated.

Interview | Ba Rui, Sea Breeze

By Hai Feng and Liu Sijie

Edited by: Liu Sijie

Cover image source | Photo by interviewee

The phase of wild, unregulated growth is over.

This is Ma Peiyuan’s most profound recent feeling, as someone on the front lines of AI startups in Silicon Valley.

In December 2021, after graduating with his bachelor’s degree, he joined Quora, known as the “American version of Zhihu,” and later moved to Poe, an AI chatbot aggregation platform, as a senior AI engineer. Two months ago, he joined Cognition, one of Silicon Valley’s most popular AI agent startups.

In May 2026, the company, founded less than three years ago, announced it had raised over $1 billion in funding, raising its valuation to $26 billion. Eight months earlier, its valuation was $10.2 billion.

In addition to being an engineer, Ma Peiyuan has another role—Venture Scout. He identifies promising AI startup projects for investment firms and has an annual investment budget of $500,000.

In the four-plus years since graduation, he consistently acted ahead of the AI industry’s consensus—his team quietly launched a project before anyone had clearly defined the term “agent.” Even when Anthropic was still unknown, they became some of Claude’s earliest and most dedicated users. He was the team’s first AI-certified engineer, its first AI agent engineer, and later transformed the entire engineering team into an AI-native team that codes with AI.

The most interesting validation of this forward-thinking awareness and action occurred two weeks ago, when he casually mentioned during a conversation with "Future Humans Lab" that "people managers are disappearing"—soon after, news broke that Tencent and ByteDance had announced the elimination of middle management.

Changes happen too quickly; in Silicon Valley, you must always be ready to adapt.

During the two-week gap between the two interviews, Ma Peiyuan realized that the logic he had previously described—that Silicon Valley chases the most expensive and cutting-edge models—was already beginning to fall apart. The current trend in Silicon Valley is to combine and dynamically schedule multiple affordable models with top-tier models.

I spoke with Ma Peiyuan for nearly five hours, and we found that he was almost constantly contradicting himself from yesterday—just like Silicon Valley. “I can’t say everything, but you might say over 60%—a month from now, it could all be reversed.”

Over the past few years, AI has nearly rewritten all the default rules of Silicon Valley: who is more likely to stand out, which organizations are more competitive, and where money is best spent.

Of course, these so-called "correct answers" are also rapidly becoming obsolete. In the AI field, a year is already a long time, and many judgments may be overturned within a month.

Here are the 10 current Silicon Valley AI observations from Ma Peiyuan's perspective:

Ma Peiyuan

Ma Peiyuan’s professional profile perfectly aligns with the ideal candidate profile favored by Silicon Valley AI companies. He doesn’t come from a traditional computer science background, hasn’t won competition gold medals, and lacks the prestige of a top-tier university—even once having no interest in computers at all. Yet over the past four years, he progressed from the advertising team at Quora to becoming an AI engineer at Cognition, embodying every keyword currently sought by Silicon Valley AI firms. His journey is both a personal story of growth and a case study in the evolving standards of industry hiring.

Observation 1: Generalists outperform specialists

The most sought-after talent in Silicon Valley today isn't the expert who delves deepest into one field, but someone who is competent in everything and has no obvious weaknesses—industry insiders call this a polyglot talent, or more colloquially, a "hexagon warrior."

In 2024, Poe internally launched an “agent” project whose definition was still unclear at the time. I, as a newcomer, was selected for the team for a straightforward reason: I had experience with AI, full-stack development, and iOS. Most other team members came from traditional machine learning, search, advertising, and recommendation backgrounds—they didn’t understand frontend code, yet the agent project required someone who could build frontend interfaces. After the person originally responsible for AI prompt engineering left the company, no one else could fill that gap. Fortunately, I learned quickly, so I took on both the prompt engineering for the agent and its frontend development.

This broad set of “jack-of-all-trades” skills later became my most valuable asset when job hunting. Give me any new field, and I can learn it well enough within two weeks—and that’s exactly the kind of person early-stage startups need.

Observation 2: AI-native skills can quickly overshadow years of experience.

The ability to use AI tools can directly overshadow qualifications accumulated through years of traditional experience.

When I switched from the Quora ad team to the iOS team, my reason was “I wanted to step out of my comfort zone,” but at the time, I knew nothing about iOS—I had zero experience. Instead of slowly catching up, I used GPT-4 as my partner. Everyone around me was a highly experienced iOS engineer who may not have trusted AI. Since I couldn’t write iOS code or Swift, I relied heavily on AI to write code for me.

As a result, I, a newcomer with only three to four months of experience, matched and even surpassed the output of seasoned engineers around me, reaching a senior role within two years—an almost impossible pace in traditional promotion systems. Previously, at Facebook, it typically took at least four to six years to become a senior engineer; some people at Google never advanced to that level in their entire careers.

I once encountered a 19-year-old interviewer at Cognition. At first, I felt frustrated—here I was, a 26-year-old being grilled by someone seven years younger, and I couldn’t answer many of the questions. Later, I realized this was precisely a sign that the company truly values talent regardless of background, and age is no longer the standard for evaluating talent in Silicon Valley.

Observation 3: Hiring criteria have been completely rewritten—AI companies no longer use the same template for interviews.

Before leaving Quora, I seriously interviewed with nearly 20 companies, and over 30 to 40 after including phone calls. At one point, I was contacted by more than 30 headhunters in a single week.

After interviewing with so many companies, I’ve noticed that AI companies today don’t have a standardized interview format.

However, there are several noticeable changes: the importance of LeetCode-style algorithm questions is declining, and the best answers regarding AI system design are changing every month.

The interview structure is also evolving. Online written tests combined with work trials have become a new standard. A work trial is a very short “trial period interview,” lasting anywhere from 3 to 5 hours, and at most one or two days. In one interview I experienced, I was directly given a codebase and asked to solve a real-world problem using any tools I wanted, including AI. After completing the task, I had to present my approach and results. The entire process closely mirrored actual job responsibilities.

More interestingly, the evaluation of AI usage itself is divided into two modes. Some rounds require the use of AI, testing how you “code by AI”; other rounds explicitly prohibit AI use—for example, during debugging—assessing whether you can identify errors without AI assistance, essentially evaluating your actual software engineering skills and taste.

Interviews themselves also vary widely. For example, some companies have asked me to perform a business analysis to determine which industry they should acquire.

The behavioral interview has changed too. It’s no longer about traditional questions like “What are your biggest strengths and weaknesses?” or “How do you resolve conflicts with colleagues?” Instead, it looks at a longer timeline, tracing your growth journey—from high school to college, from internships to your career—digging deeper to understand your “life slope” and determine whether you’re someone capable of exponential growth.

One unexpected benefit of there being no standard answer is that no one can land an offer by memorizing questions. Even my referral at Cognition told me directly: “Don’t prepare—I don’t think there’s anything worth preparing for.”

In Silicon Valley, the most exceptional people never belong to any single company—they belong to “the most outstanding company of their era.” Twenty years ago, it was Google; ten years ago, it was Uber and Robinhood; five years ago, it was Coinbase; today, it’s AI companies. Silicon Valley is “ever-changing plates, but unshakable talent.” These individuals, filtered out by new standards, stay not for compensation, but out of mission. The organizations that bring together the best talent are also undergoing massive transformation under AI’s reshaping force.

Observation 4: Excessive craving for "AI-native" solutions

The demand from Silicon Valley companies for AI-native talent far exceeds what I imagined. Most of the companies I interviewed with were Series A or B, with some even at the seed stage. A very common question was: “What AI tools do you use?” I usually just shared my own blog posts, and they were often left speechless.

They said they don’t just want me to develop AI applications, but also to drive the company’s internal AI transformation and help the team become truly AI-native. Many companies are already building AI products, but their employees haven’t yet established their own AI workflows. Additionally, they really need someone who spends hours daily on Twitter, promptly sharing new and interesting AI discoveries or productivity tips.

This feels a bit counterintuitive to me. Some of my past hobbies—like experimenting daily with new tools such as Claude Code and sharing AI usage tips in Slack—have unexpectedly become a new organizational role.

Previously at Quora, I successfully got the entire engineering team to adopt Claude Code. Whenever I met an engineer in the office, I would ask, “Are you using AI to code?” I remember one cybersecurity engineer who was initially resistant, believing it was too risky to use AI for writing code. I later encouraged him to shift his approach—instead of writing code, he could use AI to read code, understand the codebase, and perform data analysis. Eventually, he started using it.

Observation 5: "Using AI better" is essentially an organizational transformation issue.

While working on AI-related projects at Cognition, I increasingly realized a truth: better utilizing AI may no longer be purely an engineering challenge, but rather a transformation issue at the organizational level.

For example, in a traditional internet company with a hierarchical structure, fixing a bug typically involves multiple steps: requirement evaluation, engineering scheduling, development, testing, and deployment. While using AI to write code might take only two or three hours, the entire process often takes a full month.

In the past, this process made sense because fixing a bug might take an engineer three to five days, and the value of the process lay in coordinating between teams and allocating resources. But AI has drastically reduced this time.

Now, with AI software engineering products like Devin, introduced by Cognition, we can complete work that previously took days in just two or three hours. At Cognition, we can even use Devin to identify the person responsible for the bug and then have Devin assist them in fixing it—eliminating many of the coordination and handoff steps along the way, with no redundancy whatsoever. In fact, AI is driving the entire organizational transformation.

Observation 6: The people manager is disappearing; managers must be able to code themselves.

The term "people manager" refers to an engineering manager. In recent years in the U.S., a popular viewpoint has emerged that as a manager, your primary role is to focus on management—primarily coordinating the team, facilitating communication, supporting employees' emotional well-being, and providing growth guidance, without needing to be highly skilled in engineering itself.

But upon joining Cognition, I found the entire organizational structure to be extremely flat. For example, my direct report was even the CPO, one of the co-founders. He didn’t focus on the details of your personal growth; his role was more about setting the overall direction. In fact, you could say that in Cognition’s current engineering team of over 60 people, there is only one true people manager—everyone else, including the CPO, is “both managing people and writing code.” This was completely different from my experience at Quora, where a manager might oversee five or six engineers.

The Cognition team wore red clothes together to celebrate the birthday of a colleague who regularly wears red.

I think this may be an organizational change brought about by AI. I've observed that any vibrant AI startup operates this way—there are almost no managers who only handle people; if someone manages people, they are also highly skilled in technology.

This decision is not unique to Silicon Valley; within the span of two weeks, major Chinese tech companies Tencent and ByteDance also announced the elimination of middle management roles.

The changes Ma Peiyuan observed extend beyond the organizational level. Even broader industry assumptions—such as what constitutes reasonable token spending within Silicon Valley AI companies, or what makes a model truly good—are being overturned at a visibly rapid pace. This time, even the judgments he made during his first interview couldn’t withstand two weeks.

The AI industry in Silicon Valley itself is constantly overturning itself on a monthly, even daily basis—no judgment remains valid for long.

Observation 7: Token Maxxing has diverged; large companies are beginning to refocus on efficiency.

At the beginning of this year, "Token-Maxxing" briefly became popular in Silicon Valley. Companies like Meta even created leaderboards ranking token usage—those who used the most were considered representatives of "AI-native" work. I think this was fundamentally ridiculous—token usage does not equate to an engineer's efficiency.

Google and Meta’s friends told me they were all caught up in an “AI rush.” Companies would hold dedicated “AI Weeks,” during which everyone had to put aside their current work and write AI skills, embellishing their abilities extravagantly.

Initially, nearly all companies adopted this approach: encourage widespread use of AI to see just how far it can be applied.

But once it becomes a public ranking, human behavior becomes distorted. Engineers may use the system just to increase usage, even leading to absurd situations—such as scheduling AI to run aimlessly every night, performing meaningless tasks solely to boost usage metrics. This is similar to using GPA to measure student ability; a single metric is easily distorted.

Later, many senior executives also began to recognize this issue. More specifically, they were alarmed by the bills. For example, Uber burned through nearly its entire AI budget for 2026 within just the first four months. Many large companies have started reimposing limits on token usage. Boards and CEOs are now asking: Is this token consumption truly worth it? Does it really deliver such significant efficiency gains? For a $100 million investment, has it truly generated $100 million in returns?

This anxiety was quickly transformed by Cognition into a commercially verifiable proposition. On June 4, Cognition launched its “AI Productivity Guarantee”—offering up to $10 million in compensation if enterprise customers using Devin (Cognition’s AI software engineer product) did not achieve the promised efficiency gains. The research behind this was conducted by a colleague of mine who has won three gold medals at the International Olympiad in Informatics. He analyzed 233 real customer cases, randomly sampled conversations between clients and Devin, and asked clients: “How long would it have taken a human to complete this task without Devin?” Using this data, he trained a predictive system that directly converts each Devin interaction into “equivalent engineer hours,” then calculates the total cost savings by applying standard engineer hourly rates.

But many YC companies may continue the “token maxxing” model, as they benefit from generous free token allocations and are small in size—perhaps only three people—making AI governance largely irrelevant; managers can directly observe whether each individual is using AI to slack off.

Observation 8: The worship of models is itself being overturned.

During the two weeks between the two interviews in this article, I realized that the points I made last time were already beginning to seem incorrect.

Two weeks ago, I was still discussing which model was the most cutting-edge right now. Now, Cognition has internally released a feature without a Chinese name yet, codenamed Fusion Mode—intelligently combining and scheduling multiple affordable models alongside top-tier models, rather than blindly relying on the most expensive one. You wouldn’t drive your Porsche to buy groceries every day—you’d use your Honda or a more affordable car for that. The pace of model iteration over these past two weeks has also exceeded my expectations; I’m no longer confident declaring who the “number one” is.

I can't say for all of them, but you might say over 60%—many could reverse again just a month later.

This is the pattern of every technological wave: the collapse of old consensus often marks the beginning of a new paradigm.

As the market moves away from blind worship of single models and compute-intensive spending, demand surges for solutions centered on “how to deploy AI stably and controllably” and “how to tackle the hard problems that large models cannot solve.” Observing from both an engineering and venture recruitment perspective, Ma Peiyuan notices that a new wave of businesses—unreliant on the “myth of large models” and focused on solving real-world, complex pain points—is quietly emerging in Silicon Valley’s non-consensus spaces.

Observation 9: New AI infrastructure is emerging

Around “token-maxxing,” a new wave of startups has emerged, specializing in monitoring teams’ AI usage and analyzing where tokens are being wasted.

A deeper issue behind this is that AI agents themselves are fragile; even a minor change in a prompt can degrade the agent’s performance. As a result, new AI infrastructure is emerging, such as sandboxes, to isolate agent operations and prevent them from directly accessing sensitive systems like bank accounts or production environment data.

Meanwhile, the act of “observation” itself has become more complex. Traditional systems can be monitored simply by detecting errors, but agents don’t “error out.” They can run normally while still producing serious issues at the output level—such as bias, discrimination, or causing users to have clearly negative experiences. This requires new software to detect when an agent’s “mind has gone wrong.”

Observation 10: "Things that cannot be trained" are becoming a new business

There’s a growing sentiment in the investment community that vertical agents have lost their investment value. But I don’t fully agree.

Silicon Valley investor Sarah Guo has proposed a concept called “the untrainable”—things that cannot be trained. The core idea is that in many vertical industries, language models will never fully automate processes, because these industries are filled with complex workflows that cannot be abstracted away and often require human intervention to resolve one by one.

For example, U.S. tax filing involves 50 states, each with its own tax rules that must be individually and carefully integrated; the same applies to legal domains—immigration law, civil law, commercial law—each with its own complex system, which an AI agent cannot fully master in one step.

The value of a vertical AI agent company lies in continuously uncovering deeper, real-world pain points through interactions with customers and FDEs (Frontline Deployment Engineers). These insights become embedded in the company’s products—something no large corporation can replicate or absorb through generic models.

Therefore, I believe vertical agents still hold investment value; they may not deliver 1000x returns, but rather more stable, modest returns of 20x, 50x, or even 100x.