After Raising $220 Million, He Wants to Build the "ByteDance of AI" | A Conversation with DeepWisdom CEO Chenglin Wu

By Zhang Zhuo | Produced by AI Now

01

DeepWisdom, an industry star that has raised a total of 220 million RMB, finally launched its most important product last night at 9 PM Beijing time: Atoms.

To sum up Atoms in one sentence: users only need to input text to go from idea to business, end to end. Behind the scenes, it deploys an AI organization — intelligent agent roles ranging from researcher, product manager, architect, and engineer to SEO specialist and data analyst.

Chenglin Wu, CEO of DeepWisdom, told AI Now that Atoms is built on his reading of what the next phase of competition will look like.

By 2026, the value of Agents will shift from "improving personal efficiency" to "directly delivering results." To understand Wu, you first need to understand his imagination of the future — he habitually works backward from endgame to starting point.

In his view, the fundamental unit of the future will no longer be the company, but multi-agent organizations. Not everyone will need to found a company; instead, everyone will be able to mobilize an AI team at any time, "as efficient as ByteDance." Wu said, "Because silicon-based productivity will, in terms of speed and scale, comprehensively surpass carbon-based labor. Human value in the future will no longer come from how many concrete tasks one can complete, but from judgment, taste, and choice."

Atoms was designed precisely from this premise.

• Atoms interface

• A user-built headphone e-commerce site

02

Unlike most Agents, Atoms is built around real business.

Step one is research. Atoms first completes business research and competitive analysis. According to official disclosures, in benchmark tests designed to evaluate commercial research capabilities, Atoms's modules outperformed comparable models from Gemini and OpenAI. Step two is delivering a complete system. Atoms directly builds out the full infrastructure a real business needs: payments (Stripe), permission management, deployment — all done in one go. What the user ultimately receives is a system ready to launch and start charging.

Step three is parallel execution. Atoms allows multiple AI teams to simultaneously advance the same idea, offering different implementation paths, with the user selecting the most effective option to continue executing. This design attempts to trade greater compute resources for higher probability of commercial success. DeepWisdom's data shows that with comparable or even better results, Atoms's overall costs are approximately 80% lower than mainstream closed-source alternatives.

If you break down DeepWisdom's products over the past few years into a continuous logical chain, Atoms didn't appear out of nowhere.

The earliest product, MetaGPT, solved the organization problem: could you code the structure of a software company — including product manager, architect, engineer, QA — and have AI collaborate according to SOPs? Then in February 2025 came MGX (also the predecessor to Atoms), which delivered that collaboration directly to users, making AI work like a real development team. This product once ranked first on Product Hunt's weekly list, with ARR reaching $1 million.

Atoms goes further: directly solving "helping users start a business."

DeepWisdom founder Chenglin Wu started his company in 2019. Before that, he worked on machine learning at Tencent and Huawei, and from 2019 to 2022 mainly struggled with B2B AI services.

The turning point came in 2023. After the explosion of large models, he read dozens of open-source projects, "to see how the smartest people in the world were discussing AI's endgame," and reached a conclusion: AI would not appear as monoliths, but as teams, groups, organizations.

It was based on this conclusion that he decided to enter from the agent angle, having DeepWisdom advance open-source frameworks, agent collaboration protocols, and Vibe Coding products.

Wu explained with evident probabilistic thinking: "In entrepreneurship too, you should do products with high win rate and high payoff."

The following is AI Now's interview with Chenglin Wu, conducted in October last year and January this year. A quant trading expert by background, he is calm and rational in temperament. The only moment his tone showed slight fluctuation was when discussing "the first failure in entrepreneurship": "Many things don't bend to human will. I've learned to accept reality."

• The DeepWisdom team in Shenzhen; Wu in the blue jacket

• Wu at the Bund Summit

Ten Questions NOW!

AI Now: When you came to Beijing to talk with us about Atoms last October, you had planned to launch in mid-October, but it was delayed until this January. What happened in between?

Chenglin Wu: We made quite a few optimizations to elevate user experience. At the same time, we developed some new agents, especially an SEO agent that can bring traffic to users, hoping to achieve a complete closed loop and directly deliver results — so we delayed.

AI Now: While Atoms was delayed, many similar AI products for "one-person companies" had already appeared in the market. Some in the industry say you got up early but arrived late.

Chenglin Wu: I'm very clear about my team's strengths and weaknesses. Our advantage is serious cognition and judgment; our weakness is that we're not very good at marketing. DeepSeek also faced much skepticism in its early days. I understand the market, and I believe value always returns.

AI Now: The future you're convinced of is one where everyone has the chance to become a one-person unicorn. Atoms equips them with an AI team. When thinking about how to build agent teams, you once said you referenced ByteDance's organizational form?

Chenglin Wu: It's not that I referenced ByteDance, but that ByteDance's organizational culture essentially already follows "multi-agent" collaboration. It's the human organization closest to the agent collaboration model.

ByteDance's core culture has several tenets: context transparency, atomic contribution, mechanism over hierarchy, rapid feedback, and critical thinking. These are precisely the capabilities an efficient multi-agent system must have.

Context transparency is the agent's complete perception of the world. If you want an agent to truly understand its environment, you must tell it the full state of the world: data, SOPs, industry knowledge, rules, including the behavior of other agents.

Second, every idea is equally important, because the history of human technology is a process of accumulating random mutations. Even the most absurd idea, in the right historical context, could become a breakthrough point for the system. The same goes for agents. Suggestions from all roles must be allowed into the loop, to be evaluated and filtered by the system, not filtered out by authority.

Most importantly, an organization's long-term competitiveness comes from the academic cycle of "critique — generate — critique again." This is very important. Similar to academic work, a paper is reviewed by 3–5 reviewers; errors must be corrected by others, ultimately forming stable consensus.

AI Now: So to put it in plain terms, Atoms helps individuals build an AI version of ByteDance?

Chenglin Wu: Yes, that's accurate.

So in the future, human ideas and taste are what matter most. For example, you love reading novels; a novel mentions a sword, and you feel something. You can immediately use our product to turn that sword into a sellable product, completed in seconds. Or say you're a content creator with 500,000 followers — you can one-click use our product to make all kinds of merchandise.

AI Now: During testing, what were typical use cases for Atoms?

Chenglin Wu: Mostly one-person or few-person organizations — for example, a bookstore owner in the United States, a jeweler in Malaysia, a small toy e-commerce shop owner. Their need was to digitize their business.

Every user's idea was quite different, but the commonality was that they all wanted to make money.

AI Now: What's the hardest technical challenge to crack at this stage?

Chenglin Wu: Model memory capabilities and reward mechanisms. They're the foundation of all agent capabilities, but language models have inherent deficiencies in both.

First, language models are naturally bad at "remembering" and "learning." This is because the underlying structure of language models is Transformer — good at processing "current input," but unable to have long-term memory, stable concept storage, or continuous learning like humans.

A concept: humans learn it in one look, dogs in five, but language models need 1,000 looks to learn a new concept. So it's like a person with both hyperthymesia and amnesia.

Additionally, human reward systems often don't rely on external factors, but on internally secreted hormones — dopamine, serotonin, endorphins — forming internal feedback. But models don't have this, causing agents to lack self-calibration ability and rely only on external rewards for adjustment.

At the same time, current models' multimodal understanding capabilities aren't strong. They can't understand time and space like humans do, completing world modeling.

Intent recognition is also problematic. Models often can't truly understand what people mean. For example, if I say I want to buy Apple, the model doesn't know if I mean the phone, the computer, or the fruit.

Human language is a probability cloud; it's fuzzy. Models aren't good at clarifying fuzzy questions. So a very common approach is to design counter-questions, requesting more context.

AI Now: Your solution?

Chenglin Wu: Our current approach is to abandon traditional vector retrieval and instead use small models to "scan context," automatically slicing to find truly critical information.

Additionally, we're building an "active memory management" system where agents can actively choose when to write to memory, when to delete information, and optimize their own memory structure.

This is a crucial step from passive retrieval toward active learning.

AI Now: A sharp question: as a company that invested in Agent R&D as early as 2023 — it's said that Mauns referenced your open-source architecture — entering 2026, what are your advantages?

Chenglin Wu: Three areas: effectiveness, cost, and features.

On effectiveness, everyone used to think open-source models couldn't surpass closed-source models, but the past six months have completely overturned that. On our own benchmark: European and American competitors average 0.4+; we've achieved 0.8–0.9+.

Using combinations of open-source models, our results already significantly and directly surpass Claude series performance on comparable tasks. This was unimaginable before.

Second is cost. We can surpass all competitors on the market at 1/10 the cost. This is our latest data.

Finally, features. Our multi-agent + full-stack capabilities give us application depth that competitors completely lack.

AI Now: Because you have so many products, we can't find a company the general public would understand as your benchmark?

Chenglin Wu: Before the AI era, no company like us existed.

The history of the internet is the history of information distribution. Yiming Zhang used recommendation to solve information distribution. But I believe the more direct solution is to directly supply intelligence. What we solve is precisely intelligence supply.

Entering the AI era, intelligence is absolutely abundant. Before, if you wanted to find a top-tier product manager, you'd wait for them to reply; even if you found them, they might not be who you wanted. Now you can directly use an agent.

AI Now: Last question — you're keen on reading top conference papers. Share your view on key developments for Agents this year?

Chenglin Wu: The following are entirely radical personal opinions.

Agent modules will continue evolving — memory, emotion, environmental perception, etc. Within two years, GUI Agents on phones will be blown open. Within six years, embodied robots in local scenarios will also be broken through.

Image sources | Provided by interviewee, Unsplash

Join Community