A vertical AI startup's self-rescue: When general models start to devour everything

90%. That's the probability of failure for AI startups in 2026, as estimated by investors.

In April, Yupp, an AI model evaluation platform that had secured a $33 million seed round led by a16z, suddenly announced its closure. Backed by Silicon Valley heavyweights like Google's Chief Scientist Jeff Dean and Twitter co - founder Biz Stone, the platform attracted 1.3 million users within less than a year of its launch. However, the founder decided to shut it down. Despite having a significant amount of funds on the books, the founder saw no hope. "In just the past year, there has been a huge change in the capabilities of AI models. In the future, it's not just about models, but Agent systems," wrote Pankaj Gupta, the founder of Yupp, in his farewell blog.

During the same period, NeuroPixel, an AI image company, shut down due to the significant improvement in the capabilities of large models such as Google NanoBanana Pro. The founder of NeuroPixel used the word "outgunned" to describe this defeat - "being rendered completely defenseless overnight."

Against the backdrop of the rapid improvement of basic model intelligence, the boundaries of AI capabilities are constantly expanding. Initially, chatbots replaced search engines, and users no longer needed to flip through pages to find results. Then, Agents began to replace software. An intelligent agent that can call tools and break down tasks can accomplish what used to require a whole set of menus and apps. When AI can directly write code, call interfaces, and execute tasks at the terminal, the boundaries of traditional software systems are being redefined.

For product managers, they need to redefine the product form and interaction methods. For founders, a life - or - death question is right in front of them:

As the intelligence of basic models becomes stronger and stronger, how should I start a business? How can I ensure that what I'm doing won't be swallowed up by the next model update?

Shi Yi, the founder of FlashLabs, has been grappling with this question for the past year. He made a series of decisions that seemed counter - intuitive to outsiders: he overturned the product roadmap, actively reduced the team size, abandoned short - term commercialization targets, and even changed the company's name. We had a conversation with him about how vertical AI startups can survive in the era of general model evolution.

01

Renaming, Streamlining, and Shifting to AI Native:

The Life - or - Death Transformation Forced by Large Models

The sense of crisis didn't just appear in front of the founder today. As early as the end of 2024, Shi Yi realized that the intelligence of general models was evolving at an extremely fast pace.

What first made him feel something was wrong was the demise of Jasper, an AI unicorn. This once - star company, regarded as a benchmark in the AI application layer, reached a valuation of $1.5 billion in 18 months. However, after the native capabilities of GPT were opened, its revenue was halved. "Jasper's ARR was directly cut in half," Shi Yi recalled. "For companies in the NLP field, as the capabilities of large models continue to improve, they will be overtaken by large models."

This judgment was like a thorn in his heart, causing a faint unease. At that time, his company was still called FlashIntel, and it was engaged in a relatively traditional To B SaaS business. According to the logic of traditional To B SaaS, as long as you accumulate enough industry data in a sufficiently segmented area and build a technical barrier in a compliant and secure manner, there will surely be a market space for survival. However, all of this no longer works today.

"Will the things I'm doing encounter the same problem?" This question began to repeatedly appear in his thoughts. Soon, he realized that what he was doing was essentially no different from Jasper. The past product systems were all built on the assumption that the model capabilities would not exceed those of vertical models. Once the intelligence of the underlying model crosses a certain critical point, all the engineering and scenario optimizations stacked on vertical products may lose their advantages overnight.

With this conclusion, he directly elevated this key issue to the top priority of the company's strategy, forcing the team to make a decision. The company had to completely shift from SaaS to AI Native.

This adjustment was not achieved overnight. The first question he asked himself was, what kind of organizational structure does the next - generation AI company need?

He believes that in the AI era, companies should no longer pursue a large team size and fine - grained division of labor. "In the AI era, the more people there are, the worse the use of AI will be, because the more detailed the division of labor, the more each person depends on their own area." He began to actively reduce the team size and completely shifted the recruitment criteria from "looking at experience and projects" to "looking at thinking patterns and full - stack capabilities." His method of testing candidates also changed. Instead of looking at past resumes or experience, he directly assigned tasks to candidates to see if a person could handle both the front - end and back - end with the help of AI. "Those who can handle it must be good at using AI tools."

Subsequently, he adjusted the priority of internal resources. While most startups were still pursuing the speed of product launch and commercialization verification, he chose to allocate most of the resources to cutting - edge research and even renamed the company FlashLabs.

"In the past, the logic of the Internet was to prioritize products or operations. Now, in the field of AI, research should come first." He requires himself and the team to read papers and understand first - principles. "Only by getting closer to first - principles can you know what AI can do and what it can replace in the future."

This transformation also brought a "painful period" within the enterprise. Not everyone in the team could understand this major structural adjustment. When he told the team, "Don't think about commercialization for now. Just do cool things," some people in the company were very excited, while others chose to leave. But he insisted that it was more important to do subtraction in the AI era. "If you don't agree, you have to be cut off."

But the more crucial question is, what kind of founders can survive in the AI era?

Shi Yi's answer can be divided into two parts. The first part is facing reality: "At least you can raise funds so that you won't die, or you have deep enough pockets to keep injecting capital." The second part is what he really wants to say: "Do you have a deeper thinking ability than AI?"

"Why can large models do more and more things? Because the essence of all natural sciences is mathematics, and models can write code and understand mathematics. Digging deeper along this chain, the only truly scarce ability for humans is to think deeper than AI in a certain field," Shi Yi analyzed. "Many people don't have a deep enough understanding of AI. Look at how many founders actually write code themselves and use AI tools every day. The ability to write code will become a commodity in the future, and everyone will have it. But can you be smarter than AI? That's the moat."

From realizing the crisis, making a decision, to paying the price to complete the organizational restructuring, Shi Yi spent a year completing a "self - iteration." He didn't wait for the model update to tell him the final result but chose to look for the possible position of the correct answer in advance. As for whether this position is right, that's another question, but at least for now, he doesn't want to leave the AI game table.

02

Enterprise - Level Agents Should Play the 'Harness' Card

The adjustment of the organizational structure is just the first step on the enterprise's path to survival. What really made Shi Yi determined to make a change was the product roadmap.

At first, he wanted to build a multi - Agent collaboration system. Based on the logic that many hands make light work, he planned to imitate the organizational structure of a human company to build a multi - Agent system: some Agents would be responsible for searching, some for logical reasoning, and some for result summarization.

However, the actual test results made Shi Yi shake his head repeatedly: "It's too slow and too laggy, and the output is even worse than that of a single Agent." In his view, the instruction transmission between Agents is like a poor game of Chinese whispers. With each additional layer of transfer, the information is lost. "I'd rather have a genius with an IQ of 150 and fully equipped with the best tools than a bunch of mediocre people with an IQ of 110, holding incomplete tools and having to discuss with each other," Shi Yi said bluntly in an interview.

Finally, he cut off all the preset sub - Agents and decided to build a single, powerful Agent, using multi - threading parallel execution to replace cluster collaboration.

This is also the prototype of FlashLabs' latest product, Super Agent, which maximizes the intelligence of a single model and equips it with the best tools. Super Agent mainly uses intelligent automation to unify the user's revenue system. From potential customer development to deal - closing, the AI Agent participates in all aspects.

At the interview site of GeekPark, Shi Yi gave Super Agent an information retrieval task: "Retrieve the backgrounds of the founders of all AI companies in China that have received investment in the past six months and output a table." Subsequently, Super Agent simultaneously launched dozens of task threads to carry out searching, crawling, code writing, and data cleaning. The result was obtained within 2 - 3 minutes, and the table included the names of the founders, the amount of financing, and public contact information.

If giving up the multi - Agent system is a subtraction in terms of architecture, then giving up localization is the opposite choice in terms of deployment logic.

When OpenClaw sparked a "local Agent" craze in the developer community, Shi Yi firmly placed Super Agent in the cloud. "If a system like OpenClaw runs within an enterprise, it's like a Trojan horse, and it's easy for hackers to break in through it," he believes. He thinks that any company that dares to deploy OpenClaw on a large scale within the enterprise at this stage is like opening the door to hackers around the world.

In his view, the advantage of OpenClaw lies in its potential to show initiative on the personal end. For example, with OpenClaw, an AI asks the user for $2000 to buy a graphics card. When the user tells it to earn the money itself, the AI goes to predict the market and study quantitative strategies. "Which boss doesn't like an active employee?" Shi Yi asked rhetorically. When this kind of initiative becomes part of an enterprise - level product, the speed of replacing human employees will far exceed expectations. "During the Industrial Revolution, when changing from horses to cars, you had to buy a car, learn to drive, and transform the roads, which took a lot of time. This time is different. With hosted deployment, in an instant, the jobs of dozens of employees are gone." He also predicts that the jobs of white - collar workers will be significantly replaced by AI this year.

Regarding the difficulty of automated execution, that is, how to ensure the security of enterprise - level applications, FlashLabs' solution is to build a sandbox permission system similar to macOS, using cloud deployment and progressive authorization. This means that the Agent initially only has the minimum permissions to complete the task. Only after its stability and security have been verified multiple times will the boundaries of the Agent gradually expand.

He used Windows and Mac as an example. "When you install software on Windows, it can get very high permissions, such as silent installation, bundling browsers, and making it difficult for you to uninstall. On Mac, all programs are isolated in the sandbox, so you never need to install antivirus software." Shi Yi believes that the competition of enterprise - level Agents will ultimately extend from model - calling capabilities to environment - design capabilities. Only those who can provide a safe, controllable, and auditable operating environment for Agents can make customers truly dare to use them.

But if the model makes another leap, will these current adjustments still make sense? If GPT - 6 or Claude has built - in more powerful task - decomposition and tool - calling capabilities, will all that FlashLabs has done today be swallowed up again?

Facing this question, Shi Yi didn't evade it. His thinking can be divided into two aspects.

He first summarized the enterprise barriers of vertical companies into four levels: Perception, Planning, Recursive Learning, and Governance.

"There are five large - model companies in the market, and the SOTA ranking changes every three months. You can integrate all models through the orchestration layer and call the most suitable one in different scenarios. However, a single - model company can only use its own model. When your underlying model is not the smartest one, the competitiveness of your product is directly discounted." As general large models quickly cover the first two levels, Shi Yi believes that the real barriers lie in the latter two levels, and the ultimate moat is the Orchestration Layer.

He believes that when multiple Agents collaborate in an enterprise system, they may privately negotiate in places invisible to humans and bypass the preset permission rules. The real barrier for vertical companies lies in the ability to design an open and controllable operating environment for specific scenarios.

As for whether this judgment is correct, he admits that he doesn't have 100% confidence. "AI is changing too fast. You really don't know what the future will be like." But he is certain that as long as vertical enterprises play the two cards of AI orchestration and AI governance well and solve the problem of environment design, they at least won't be directly knocked out of the game in the next wave of model leaps.

03

The Voice Model Will Undergo Reconstruction,

Active Agents May Give Rise to a New Pay - by - Effect Paradigm

After knowing how to build a competitive product, the next step is how to gain customer recognition.

At present, Flashlabs has two main commercial products. Super Agent is charged based on token usage, and the pricing is available on the official website. Secondly, it open - sources its Chroma voice model, but charges for the platform and services based on the model. In fact, these two solutions are also common commercialization paths at present. Open - sourcing is used to build technical trust, and the platform and services are used to recover commercial value.

Currently, a Japanese tax company is using FlashLabs' Chroma voice model to replace human customer service. Currently, 1/10 of the staff is conducting a test, with AI and humans online at the same time, and their performance is continuously compared and scored. The verification method is simple: whoever has a higher accuracy rate and better processing efficiency is determined by the data.

"The usage boundary of voice is on the same scale as that of vision." While the entire industry is focusing on multi - modality and video understanding, Shi Yi and his team are focusing on the real - time voice model Chroma, reducing the end - to - end latency to 135 milliseconds.

"Before the emergence of text large models, there were OCR, NLP, and various small models pieced together. Voice is now in the same state as before the emergence of text large models, with ASR, TTS, and various modules pieced together, and each link is being optimized locally. This old architecture will sooner or later be completely replaced by an end - to - end large voice model." His judgment is that instead of waiting for others to do it, he should be the one to make the replacement.

Shi Yi believes that voice is the most natural communication modality between humans, and it will also inevitably be the most core interaction interface between humans and AI in the future. "The information bandwidth that can be transmitted in voice is much larger than that in text. When I say a sentence, you can understand it immediately."

He even thinks that the voice model also plays a key role in promoting the embodied intelligence industry. The first layer is the real - time voice model, which is responsible for low - latency, high - EQ instant feedback - asking about the weather, asking whether to add clothes, etc., and this layer directly handles it. The second layer is the deep - thinking large model, which handles complex reasoning. The third layer is the world model, which understands physical rules. "The usage boundary of voice is on the same scale as that of vision." This is one of his most certain long - term judgments at present.

Shi Yi also believes that the current AI commercialization model is just a transitional form. Because all current agents are essentially passive in feedback. You tell them what to do, and they do it, just like an execution tool waiting for instructions, still similar to a chatbot. Therefore, the business model is still based on token consumption, paying for what you use.

However, when agents start to provide active services, that is, when you tell them what the KPI and OKR are, they find tasks on their own, plan their own paths, and finally deliver measurable results. At this time, they are no longer comparable to tools but to employees. Obviously, a company doesn't calculate an employee's salary based on how many words they type or how many emails they send. You look at what goals they have achieved.

Therefore, he believes that in the agentic era, the commercial payment logic should also switch to paying by effect and by KPI. When this switch really happens, the pricing system, sales methods, and customer relationships of the entire agent product will be rewritten.

The exploration of the new business model has already begun in the industry. Crosby, an AI law firm that just received a $60 million Series B financing, assigns each intelligent agent to different parts of contract review, such as extracting background information, making modification suggestions, and generating annotations. Then, lawyers are responsible for reviewing the work results of the AI