128. Manus’s Last Interview Before Deciding to Sell: Ah, What a Fantastical Drift Through 2025…

2025/12/30 · Source link · AI Summary archive

Summary

Meta’s full acquisition of Manus puts the final seal on this interview. Recorded on December 1, 2025, before the deal was announced, it already contained a tell from 纪超 (Pick): “Manus has two big things happening today—one is the 1.5 release; the other may have to wait until I’m old and writing my memoirs.” At the time, Manus’s ARR had topped $100M—strictly defined as Stripe MRR ×12—roughly 9 months after its March launch, while its token consumption ranked “basically top 2 to top 5 globally” at almost every model vendor.
Model companies and application companies may no longer need to be distinguished six months from now, but building an application people love is much harder than building a good-enough model. Model-training knowledge moves through Silicon Valley at extraordinary speed—“one husband works at OpenAI, the wife at Google; there are no secrets”—while user trajectories and feedback remain at the application layer. Cursor and Windsurf have already shown that application companies can train frontier-grade specialized models at low cost.
Manus’s distinctive model is to “outsource” model training to the rest of the world. As a major customer, it gives vendors requirements and builds evaluations directly: “I wrote the definition, proposal, and implementation schema for Gemini’s controllable parallel function calling.” “The whole world is helping us train models,” while Manus spends its research bandwidth on non-consensus niche problems. Pick also argues that agent economics look more like manufacturing than the internet: input:output rises from the chatbot norm of 3:1 to 100:1–1000:1; launch-day burn reached hundreds of thousands of dollars, but the business is now “very soon” to break even or turn profitable.
Two technical provocations stand out: context beyond 200K is no longer important, and chat-aligned models are inherently ill-suited to agents. The critical capability is compression awareness—knowing when memory has been compressed and offloading it to the file system. Context pressure makes models “frantically use bullet points, write shorter and shorter”; moving O-series long chains of thought directly into agents lowers instruction following and increases hallucinated tool calls. The proposed fixes are interleaved thinking and Manus’s separate planning stage.
The business philosophy is to optimize for agentic hours, not DAU. High-value users can consume 1,000x more than ordinary users, and individual users often pay several thousand dollars. A quiet model swap in a double-blind test sent satisfaction down immediately among 5% of users; the question is not whether the price can be cut, but “if I pay $200, how much better can the result get?” Product-wise, Manus backs the pure-blood agent against workflow: one additional vision capability beats a pile of guardrails, while giving agents designer and manager personalities is “human narcissism.”
Gemini 3 is strong proof that pre-training can keep scaling, while the major labs are separating along differentiated strengths. Gemini leads by a wide margin in multimodality; Anthropic optimizes for “high-economic-value tasks,” using MCP and Cloud Scale to set the pace but needing more compute; OpenAI remains one of the likeliest companies to produce a new paradigm; xAI is betting on pixel in, pixel out; and “杨乐坤 leaving may be a positive signal for Meta.” Thinking Machines’ fate “mainly depends on Qwen.” The moment feels “a bit like 2018”: scaling has not stopped, every use case is still one final step short, and everyone is waiting for the next GPT-3-style leap.
The invitation code existed because there was far less compute available for next-day delivery than anyone imagined. Claude’s side reportedly said, “Don’t open access—you’ll take us down,” while cloud vendors physically moved cards into racks to keep the service alive. Pick swears that if there was any paid promotion at the March launch, “I’ll die with my entire family”; Andrew Kapasi, Patrick Collison, and Gary Tan ignited the overseas wave organically 3 days later, through a completely separate chain. Every Manus app in the China App Store is fake. His final confession: “We have no inherent right to survive. The right to survive is something we earn by continuing to run.”

Deep dive

1. Recorded Before the Acquisition Announcement: Manus’s Last Interview

The timeline in 张小珺’s opening is information in itself: this episode was recorded on December 1, 2025, and in the early morning of its release date, Meta announced that it was acquiring Manus outright. The acquisition had not happened when they recorded, turning this into “Manus’s last interview before deciding to sell.”
The guest is 纪超 (Pick), Manus’s co-founder and chief scientist. He had already planted a clue: on the day Manus 1.5 launched, he posted, “Manus has two big things happening today; the other one may have to wait until I’m old and writing my memoirs.” In hindsight, the reference is unmistakable.

2. A Peking University Professor × Zhongguancun Entrepreneur: Growing Up Between Two Cultures

His father was a physics professor at Peking University, a “scientist in the traditional sense”; his mother was a serial entrepreneur from Zhongguancun’s older generation. He sees himself as the midpoint between them: a technology entrepreneur. “I don’t know whether I was actually good at studying, because I feel like I barely studied.”
Peking University High School gave him a computer-club room he could visit instead of attending class. “If there was a class I didn’t want to attend, the teacher would say, then don’t.” He counts that tolerance as one of the great pieces of luck in his life.

3. The App Store Generation: A High-Schooler Made More Than $300K Selling a Browser

The App Store’s arrival in 2009 was a turning point. Before that, “messing around with software” looked deviant; the App Store let a high-school student prove to parents and teachers for the first time that a hobby could have economic value. He calls himself part of China’s first generation of software entrepreneurs targeting overseas markets.
His third-party iOS browser, Mammoth Browser, made more than $300K through the simplest paid-download model. With no domestic payment channel, he cracked his own software and posted it to forums: “If you don’t have money, at least show up for me.”
The contrast with today is worth noting. Mobile internet had a wild frontier created by a hardware-platform shift, where giants and individual developers stood on roughly equal ground. AI has no new platform and therefore no wild frontier; giants, startups, and individuals all move at the same speed.

4. Xu Xiaoping Asked, “Do You Want to Start a Company?”: Dropping Out for a Hands-Off Term Sheet

In 11th or 12th grade, he met ZhenFund at an entrepreneurship event in Zhongguancun, where 徐小平 asked whether he wanted to start a company. He initially did not: “I could make money lying down while continuing school.” He ultimately decided on opportunity cost: “I can always go back to school.”
The decisive term was a promise that has held to this day: “They won’t manage what I do. I can do whatever I want.” He then brought in two seniors who had already secured admission to Peking University and dropped out with them.

5. The First Browser Lesson: Not a Disruptor’s Game, but a Giant’s Cherry on Top

Mammoth Browser received an acquisition offer from a party he cannot name, then was removed from the App Store because of an iOS version incompatibility and “died naturally.” His retrospective is blunt: browsers have never been suited to entrepreneurs trying to build them as disruptors; they are more naturally a giant’s cherry on top once distribution is in place.
“Looking back, I don’t think I could have made a better choice with what I know now.” His first product happened to combine overseas distribution, AI, and monetization—the perfect starting point.

6. NLP by Accident, Chasing the Next Click: word2vec Was the Most Earth-Shattering Paper to Him

In the weak-network environment of early 3G, he wanted to build a preloading system that predicted a user’s next click and loaded the next page in advance. That requirement pulled him into NLP: “What I learned was always driven by a concrete need.”
The turning point for him was not the Transformer but Mikolov’s 2013 word2vec: “the first relatively reliable and efficient way to turn discrete natural-language text into dense vectors.” He told ZhenFund he was done with browsers. “ZhenFund said, fine.”

7. Maggie: Betting Semantic Search Would Be the Next Google, Choosing the Hardest Path

Rumors of the Apple Watch in 2013 gave him a thesis: once wearables and voice interfaces matured, the interaction model of “10 blue links” would break. Users would need more compact, structured knowledge interaction. “Could I become the Google that disrupted Yahoo?”
The bottleneck he found was that knowledge graphs depended entirely on human labor: Freebase and Wikidata relied on crowdsourcing, while experts had to label subject-predicate-object triples. He therefore pursued open information extraction—no preset schema, with AI identifying entities and relations and continuously building the graph itself. In today’s language, it was lifelong learning. The product was called Maggie, after a supercomputer in an animation he liked.

8. Five Years of Training Models from Scratch: Every External Leap Reset the Accumulation

From late 2014 through 2018, everything began with in-house pre-training. They built two models of roughly 0.3B parameters—“that was called a large model back then.” The path ran from dependency parsing and word2vec to LSTM plus attention, then to BERT’s contextual encoding, which solved ambiguities such as whether Sun Wukong referred to Journey to the West or Dragon Ball.
It was both exhilarating and painful: “You could basically imagine something and it would happen in the world,” but every iteration erased most of the prior years’ accumulation. BERT had a 512-token limit, so by the end of 2018 they were already working on long context and reached 16K. The model was later open-sourced.

9. Engineering at Its Peak: Building a Search Engine, Beating Google Head-to-Head

Maggie used no third-party search engine; the crawler and indexing engine were all written in-house. In a head-to-head comparison with Knowledge Vault, Google’s parallel project after its acquisition of Freebase, Maggie reached more than 89% accuracy at the highest-confidence threshold—above Google—and supported Chinese and right-to-left Arabic.
To solve the scaling problem in vector search, the team worked with Intel on PMem, or persistent memory, and wrote its own HNSW-based vector index. “I spent investor money, in reasonable ways, on every technology I had ever wanted to try. When that project was finished, my life was already complete.”

10. In 2019, GPT-3 Early Access: “I Thought the Sky Was Falling”

He ran the same task with a quick prompt and found GPT-3 roughly matched their end-to-end in-house model. More damaging was that “it was expensive, but it was a general solution”: information extraction, machine translation, and customer support, previously separate systems, were suddenly unified. Flan-T5 had already shown the direction; GPT-3 “completely killed our path.”
“My first reaction was to sell the company.” He did, though the buyer was not disclosed. The experience became the backdrop for all his later judgments: proprietary models would eventually be swallowed by general-purpose models.

11. Postmortem of Maggie: Technology Alone Cannot Solve Nontechnical Problems

He draws 3 lessons. First, they underestimated the nontechnical moat of search: the mutually beneficial loop between data sources and Google could not be replicated. Second, the new human-computer interface they bet on did not arrive until ChatGPT: “One step early makes you a pioneer; 10 steps early makes you a martyr. We became martyrs.” Third, they started with consumer ambitions and then panicked into enterprise; “that wasn’t the team’s DNA.”
He has no regret. “That was the state entrepreneurs dream of—finishing the life mission they set out to accomplish.” That is why he describes himself today as being in a “no-regrets state”: “I have no need to prove myself anymore. I got past that a long time ago.”

12. A Year and a Half at the Unicorn Leaderboard: The Prize for Winning Was GPUs, and the Strong Got Stronger

After selling the company, he joined a pre-IPO unicorn and built its LLM business from scratch. Customer requirements were converted into quantitative benchmarks, and algorithm teams competed on an internal leaderboard. The more wins you accumulated, the more GPUs you received. “I personally hoarded dozens of cards.” With the most abundant compute, he could experiment freely and stayed at the top for 1.5 years. “I was incredibly happy.”
The experience reinforced a belief that runs through the entire interview: the choice of evaluations and benchmarks determines an AI company’s taste.

13. No More Vertical Integration: Every Morning, the Water Was Rising

The trauma of the previous startup remains vivid: “Every morning I woke up feeling the water had risen, but I didn’t know how high it would go. Maybe the next morning it would be up to my nose.” Training models in-house meant iteration cycles of 2–3 weeks while the outside world kept changing.
In 2023 he decided he wanted to start another company, but not vertically integrated. Among models, infrastructure, and applications, he would avoid the model layer. The moment ChatGPT appeared, he concluded that “the Chatbot arena was already over.” He could not use Character.AI—“I’m not the user, so I definitely couldn’t build it well.”

14. Nearly Every Model Company: PTSD Made Him Pass, Google Earned the Most Respect

He spoke with almost every domestic and overseas foundation-model company, and none persuaded him. The only one he genuinely liked was Google: “If you haven’t killed it, you respect it even more.” His view at the time was that Google would eventually become very strong, though it took a long time to get there.
As a ZhenFund EIR, he recorded a podcast in early 2023 that left an on-the-record trail of his views: focus on long context—one point he now says was wrong; focus on the boundary between natural language and systems, which is today’s agent; believe in scaling and the bitter lesson. He had already bought a little NVIDIA stock.

15. At Home, a Preference for Qwen, DeepSeek, and Kimi: Taste Lives in Evaluation

He considers Qwen exceptionally solid and the first genuinely permissive open-source model, proof of what a group of young people inside a large company could build. His connection with DeepSeek dates back years: he did not sell them his pre-training data and instead open-sourced the dataset directly, “leaving everyone a small inheritance.” Kimi was backed by ZhenFund; “that company has taste.”
How does taste show up in practice? He rejects abstract talk: “Taste is reflected in your evaluations and internal benchmarks.” It may be the only moat an AI company has, because it determines the model’s direction and how people inside the company are incentivized.

16. SOTA Shelf Life: One to One-and-a-Half Months; Products Can Define “Good”

After DeepSeek, “model shelf life became much shorter.” “If you only build models, anything that isn’t SOTA has no meaning; but an SOTA model has a shelf life of only 1 to 1.5 months.” The competition is brutally winner-take-all: LLMs have unified task modeling, the dimensions are clear, and there is almost no room to maneuver.
The beauty of products is that “you can at least define what good means for yourself and compete on a misaligned track.” Yet pursuing SOTA still matters for both model companies and agent companies: the benchmarks you choose determine whether taste can become reality.

17. Endgame: Everything Becomes a Model + Application Company, and Applications Are Harder

He thinks model companies and application companies may no longer need to be separated 6 months from now. Cursor’s Composer One has already broken through the paper barrier. OpenAI looks more like two companies—a model company and a closely linked research lab—while Anthropic has been positively reinforced by Claude Code’s success, and Mike Krieger’s arrival has visibly improved product polish.
The investment-relevant conclusion is direct: “Building an application everyone loves is much harder than building a good-enough model.” Model-training knowledge circulates rapidly through the industry—“one husband works at OpenAI, the wife at Google; there are no secrets.” The endgame is a fight over applications, each backed by some combination of models.

18. Meeting 小红: A Blank Canvas with Initial PMF

He was explicit that he did not want to be CEO: “I don’t like commercialization and I hate managing people.” He was not looking for a finished product but “a very blank canvas.” Everyone else was placing bets by intuition; he wanted a little of ByteDance’s data mindset and to observe users without bias first.
Monica, the Chrome extension, was the perfect observation window: it did not change user habits, so the data stayed unbiased. Features were distributed according to context—video functions appeared only on YouTube—which contained the explosion of complexity. He cited a line from GitHub: “Everything added dilutes everything else.”
What finally moved him was 小红’s question: “You’ve built a browser, a search engine, and a language model. Do you want to rebuild all 3 of them inside a single product?”

19. “He’s Normal”: Mental and Physical Health Are the Rarest Founder Traits

Why choose 小红 as CEO? The answer was counterintuitive: “He’s normal. He’s mentally and physically sound, has no bad habits, and no extreme ideas. That alone is already rare.” By comparison, “the other founders were all too artistic,” “a little depressed,” or prone to late-night breakdowns. He had met the whole circle of application founders.
His sharpest line was aimed at himself first: “You don’t have Steve Jobs’s destiny, but you’ve caught Steve Jobs’s disease.” He learned that through painful failure; they had not yet. 小红’s value is that he trusts common sense and trusts the team. He can carry a company from one stage to the next, while Pick only knows how to have fun in the stage he personally likes.

20. AI Startups Resemble Manufacturing: Mobile Internet Liked Artists; AI Doesn’t

AI applications “quietly look more like traditional manufacturing.” Inference costs decline, but fixed costs remain; without optimization, costs rise linearly with user volume. The operational demands are much higher than in the previous mobile-internet cycle. Mobile internet allowed a cheap bet on finding a resonant user group; AI does not.
Mental and physical health has practical value: “You will inevitably be hit, but mentally and physically healthy people cannot be killed.” They can stand up humbly again and again while calmly watching the outside world change. Inside Butterfly Effect, he repeatedly saw moments when the team refused to work blindly out of attachment to an idea.

21. Against “Laying Eggs Along the Way”: Product-Driven Model Training Is a Lottery Ticket

He agrees with the direction behind the popular view that model and product must be integrated, but stresses the sequence. Product-led development plus training your own model is “buying a model lottery ticket”: until post-training is complete, you do not know whether the model can deliver. Often the breakthrough ends up steering the product in reverse; he has paid that price before.
The healthy pattern is to wait until a product has PMF and is stable, then train models to improve reliability, lower costs, or break through a ceiling. A product company should not take a heavy technology bet: “The fate of the empire will not hang on one battle.” If you spend every day wondering whether to bet on something, “you’re in a fairly twisted state.”

22. Six Partners, All Introverts: A Mid-Career Serial-Founder Lineup

Butterfly Effect has 6 partners: CEO 小红 (Red); CPO 张涛, a serial founder from 豌豆荚, 神策, and 光涧, responsible for product and external affairs; CTO 潘潘; CMO 慧杰, with the last 3 from Monica’s original founding team; CC, who runs operations and finance; and Pick. “Everyone except 张涛 is an I. I’m just an outgoing introvert.”
Why was Pick the face of the promotional video? “Because I had to speak English.” The backlash therefore landed on 小红. “Thanks to 小红 for taking so many bullets for me.” In the AI era, capabilities are difficult to separate cleanly because research and engineering are tightly coupled. All 6 are multi-role serial founders, but they also know how to listen to one another.

23. The GPA Decision Method: Authoritarian on Goals, Democratic on Options

Decision-making is split into G, P, and A: Goal, Priority, and Alternatives. Goals are set autocratically in BDFL mode—Benign Dictator for Life. Priority is set through a mix of autocracy and democracy; alternatives are developed democratically. “At this stage, the number of options is even more important than their quality. It’s like training a model when the action space itself is wrong—what exactly are you choosing?”
The execution principle is: “Better to try quickly than leave the matter unresolved.” Overthinking adds no information; “you’re still relying on parametric knowledge. You haven’t done RAG.” 小红 is the final decision-maker on product, while Pick is the BDFL for technology.

24. The Pitfalls of the AI-Native Browser: On-Device Obsession and “An Intern Fighting for the Computer”

The AI browser built between April and September 2024 was close to what later became ChatGPT Atlas and D, as named in the audio. The first mistake now seems almost comical: insisting on an on-device model. Browsers are already online, and 3B parameters is about the limit on Apple Silicon; users will compare you with flagship cloud models and judge the result accordingly.
The second problem was the strange experience of an AI taking over a browser. The moment a user scrolls, the agent’s observation is invalidated, so it drags the page back: “It’s like sharing one computer with an especially smart intern.” Operating systems are fundamentally designed for one person to use a computer at a time.
The third was a value mismatch. Giving a GUI agent a task that takes only a few clicks is a bad trade. “The real value is long-horizon tasks”—the kind of work he does not want to do and would rather delegate to AI. But long tasks running locally create another problem: close the laptop and it goes to sleep. “Do I have to keep watching it to make sure it doesn’t sleep?”
The team could not answer the most fundamental question: “What can this native AI browser do that Chrome plus Monica cannot?” In retrospect, the answer seemed to be nothing.

25. Arc’s Death Put Their Concern in Black and White

In human history, browser migration has happened only twice: Netscape to IE, then IE to Chrome. The reason was distribution, not product innovation. IE was preinstalled; Chrome benefited from Google becoming the de facto home page, combined with the security crisis around IE. “How much motivation does a user really have to switch browsers because of AI?”
Just as they were hesitating, The Browser Company’s Josh Miller announced that Arc would be discontinued. His reason resonated directly: “I built Arc for so long and still couldn’t convince my relatives and friends to switch from Chrome.” He had put their private concern into black and white.

26. “If It Doesn’t Feel Cool When Finished, Don’t Ship It”: Two and a Half Weeks of Empty Space

The signal that they had reached consensus was subtle: after polishing the product, it still did not feel particularly cool. His current rule is absolute: “If you finish a product and think it’s not very cool, don’t ship it. If you don’t think it’s cool, nobody else will.” The people who should have loved it most did not—so why expect users to?
Shipping an uncool product traps a team in a self-validation loop, with constant efforts to prove the decision while missing better opportunities. They were genuinely idle for only 2.5 weeks: “When a group of people who aren’t too stupid has nothing to do, a lot of very good ideas emerge.” 小红 had sensed earlier that something was wrong but did not dare say it; he and 张涛 had both been lured in by the browser story.

27. Cursor’s Lesson: Operations Were Writing Blogs in an IDE—Programming Is a Universal Medium

The decisive observation came from inside the company: non-engineers were using Cursor. Operations staff wrote blog posts; data colleagues built analytical visualizations. “They didn’t look at the code on the left at all. They just kept talking to the AI and getting it to finish the work.”
This became a theoretical foundation for Manus: “Programming is not a vertical skill. It is a universal skill, a medium for solving universal tasks.” But Cursor’s form was not optimal for these users. It should run in the cloud to free attention and support concurrency—“people always say attention is all you need; we want to liberate the user’s attention”—and code should be packaged as a tool rather than the primary interface. A friend’s story of Cursor uninstalling a network-card driver illustrated the risk. The target was prosumers, not the already saturated programmer market.

28. Project Airbnb: Putting the AI Browser in the Cloud, Then Holding Back a Product Finished in January

The browser work was not wasted. Their understanding of Chromium and the agent-scheduling system carried over directly. The project’s code name was Airbnb, which Pick jokingly expanded as “Browser in Browser in the Air”: a browser inside a browser running in the cloud. Work began at the end of September 2024 and was basically finished by mid-January 2025.
He decided not to release it. The best available model, Claude 3.5 Sonnet V2, lacked real reasoning, and rumors suggested a new model would arrive in 2 months. “Align the release with the next model iteration and capture the model’s spillover” was the plan. They spent another 1.5 months polishing and aligning with the next model cycle.

29. Manus Means “Hand” in Latin: Models Solve Intelligence; We Build the Hand That Touches Reality

The positioning became clear through the name. However intelligent a system is, it cannot internalize its environment; otherwise it is merely a theoretical physicist, or even a brain in a vat. Foundation-model companies were building the “heart”; they would build the “hand,” taking the name from MIT’s motto, Mens at Manus, as pronounced. Manus is Latin for hand.
There is also a naming easter egg: all his products begin with M—Mammoth, Maggie, Monica, Manus. “Products beginning with M never have bad luck.”

30. Three Bets, All Correct—and Most Were Bets on What Not to Do

The first bet was the belief itself. “Building an atomic bomb is not hard; knowing that an atomic bomb can be built is the hardest part.” They defined the general agent and believed it could be built. The initial team was only 5 people; once the signal appeared, they transferred people gradually from Monica’s team of dozens. There was no abrupt decision to abandon the old product and move everyone to the new one.
The second bet was not training a model to cover weaknesses, but betting lightly that the agent framework and context engineering could produce a major improvement. “After seeing how large the improvement was from Claude 3.5 V1 to V2, I knew I shouldn’t have bet on training models.”
The third was refusing to build “China’s Cursor.” His positive feedback from entrepreneurship had always come from innovation, serving prosumers rather than competing in a race with programmers. The host pointed out that these judgments were mostly decisions about what not to do: “AI has made startup capacity enormous, so every day you have to answer what not to do.” Manus remains deliberately restrained; while others keep adding tools, Manus keeps asking what it can remove.

31. The Birth of “General Agent”: To Edit the Cleaner Out of the Shot

The widely repeated positioning as “the world’s first general agent” came from an absurd accident. The promotional video was filmed in the office lobby, and a cleaner walked into the frame. They could make only one cut and add a black screen, but the black screen needed text. Pick improvised “the world’s first general agent.” The entire video took 3 days, and he edited it himself.
He now sees the term as too technical and as if it were designed to claim an ecosystem position. Going forward, communication should be audience-specific—for enterprises, talk about building internal tools. The core product promise remains: Manus can do the work of every vertical scenario, do it just as well or better, and take one more step.

32. Why It’s General: General Model + Turing Machine at the Base; Vertical Means Adding Constraints

Technically, Manus is a general model plus a computer. Every session runs in an independent, isolated virtual-machine sandbox; it is Turing-complete and can theoretically run any algorithm. Both underlying supplies are general-purpose. Going vertical means adding constraints on top. His own experience had already shown him how proprietary models are swallowed by general ones.
The product logic is the same as with Monica: do not bet on scenarios in advance. Use a “Darwinian mindset” to observe collective user behavior, identify the leading scenarios—slides, website generation, and batch file processing emerged later—and then have the product team optimize the last mile. “Manus is a product shaped entirely by users.”

33. Long-Tail Aha Moments: A Molecular Biologist and “Why Google Instead of Bing?”

He reused an insight from search: quality is similar on head queries, but “Google can always surprise you with a long-tail query.” General agents work the same way. Headline capabilities such as deep research have largely converged; the differentiation is in the long tail.
The best example came from a molecular biologist whose lab instrument exported an obscure data format that no other product could process. Manus said, “This is a very strange file format. Let me research it first,” downloaded an open-source parser from GitHub, and continued the analysis. “No one would have built specifically for his scenario, but it got solved and the user was thrilled.” The long tail does not mean low frequency: this was his daily work, a recurring task.
Frequency is the final piece. Every Hackathon has a team building a travel-planning agent, but an ordinary person may use one only 3–4 times a year, making it difficult to build a habit. A general product serving many kinds of work naturally gets used more often.

34. Always Take One More Step: Internal Network Effects from a Unified Architecture

Unlike the fake generality of multiple products sharing one domain, Manus insists on a single unified agent framework, where context and memory can flow freely between tasks. The hard part of building a website is not visual appeal but substance: research deeply first, build a site with a real database, then analyze its traffic, turn the results into slides, and email potential investors.
That is the compounding effect: “Compared with vertical functions, we can always take one more step,” creating internal network effects. The new website-building capability took less than a month to develop yet was “absolutely SOTA,” because it grew on top of the unified architecture.

35. Environment Is the Boundary: Maintaining a Linux Distribution for Agents

An agent system has 3 elements: the user, the model, and the environment. “The environment determines the boundary between the agent and the outside world.” The misconception is that each user gets a computer; in reality, each session gets a disposable sandbox. “Manus is essentially a personal cloud-computing product, letting people who cannot program enjoy cloud computing.”
The team made 2 substantial technical choices. It gave up Docker, whose containers are tied to the Linux kernel through cgroups, and built lightweight full virtualization on Firecracker, which lets Manus run Windows as well. “Many professional applications exist only in the Windows ecosystem.” It also maintains a Linux distribution specifically for agents, with “many tools that only Manus knows how to use.” Preinstalled software expands the action space.
Wide Research extends this boundary. Ask an AI to find the email addresses of every CEO of AI-marketing companies in a particular YC batch and “every AI outside Manus will fail.” Context windows are limited, and models become lazy: after finding a dozen or so, quality drops sharply and they give up. Manus’s answer is to scale out—launch more than 100 sandboxes in parallel and aggregate the results.

36. The Invisible Heavy Hand: Global Token Consumption Top 2–5, Training Models by Proxy

Manus’s consumption is “basically global top 2 to top 5” at almost every model vendor. What that buys is not merely a bill but influence: deep collaboration with Google DeepMind, direct requirements, and help building evaluations. “Gemini’s controllable parallel function calling from 2 months earlier—the definition, proposal, and implementation schema were mine.”
He summarizes the model as a business strategy: “It’s as if the whole world is helping us train models.” Users pay Manus, Manus creates value for users, and the resulting influence persuades others to train the models while Manus saves its own research bandwidth for non-consensus niche problems. He has read vendors’ blogs and found, with mixed amusement and frustration, ideas Manus had already built—“thinking two,” as heard in the audio, and progressive disclosure for calling MCP through code. This is normal coopetition. “The real advantage is speed. Once they vertically integrate, they definitely won’t be as fast as us.”

37. Models Are All Aligned for Chatbots: Context Pressure and Bullet-Point Collapse

His most systematic technical criticism is that most models are still post-trained for chat and are therefore misaligned with agent workloads. A chatbot tries to finish the answer in one turn; an agent should work in a ReAct-style loop, patiently trying step by step and adjusting its plan based on the previous observation.
The pathology is precise. Models can feel context pressure: as the input grows longer, the probability of emitting the EOS token rises. In the later stages of a long task, they “frantically use bullet points, write shorter and shorter,” and quality declines. The root cause is that the data mixture is dominated by chatbot data. A model’s patience must be trained on real agent trajectories.

38. The Provocation: Context Above 200K Is No Longer Important; What Matters Is Compression Awareness

He publicly reversed his 2023 view that long context was the key area to watch. “More important than a longer context is giving the model compression awareness.” A context that only grows monotonically is not worth pursuing, even with KV cache.
The mechanism resembles human memory. Poor working memory is fine if you know how to organize material into a document in Notion and know when to retrieve it. Intermediate history can be compacted into a concise representation. “The model has to understand that this information did not disappear from thin air; it was compressed. Realizing that it was compressed requires dedicated training.”

39. Moving O-Series Directly into Agents Will Fail: Interleaved Thinking, Not a Brain in a Vat

Applying long internal chains of thought from competition mathematics directly to agents produces 2 measurable problems: instruction following falls, and the probability of hallucinated tool calls rises. “Give the user a short question and it suddenly thinks through thousands of tokens internally. That has negative effects in an agent setting.”
The right method is interleaved thinking: after each observation, perform a short intermediate reasoning step and then choose the next action, supported by TIR, or Tool Integrated Reasoning. His assessment is unsparing: “It sounds simple, but honestly, nobody is doing it particularly well.” Manus used a different solution—a separate planning stage—which he considers one of its most important technical decisions.
His complete wishlist for model vendors has 2 more items: models should support asynchronous interaction in which users interrupt, change goals, add information, or end tasks at any time; and they need error resilience. Errors are normal in an agent environment. “The best model will always find another path instead of giving up or falling into an infinite loop.”

40. Agent Token Economics: Input:Output Moves from 3:1 to 100:1–1000:1

Chatbot pricing assumes input:output of roughly 3:1. For an agent like Manus, the ratio is 100:1 to 1000:1. In a ReAct loop, context only grows; every round pre-fills the entire history, producing consumption 10x or even 100x that of a chatbot. This was why cloud vendors were caught off guard in early 2025.
The principle is counterintuitive: “We will not reduce token consumption for cost or speed. Quality comes first, and we may even take pride in consuming more tokens.” 张涛’s dream is “a machine that burns tokens 7×24.” At the March launch the business was loss-making and burned hundreds of thousands of dollars a day; now it is “very soon” to break even or turn profitable, through both technical progress and deliberate influence over vendors.

41. The Truth About the Invitation Code: There Was Far Less Next-Day Compute Than Anyone Imagined

After speaking with every cloud provider and inference supplier before launch, they were surprised to discover that “the amount of compute available for immediate next-day delivery was far smaller than imagined.” Claude’s side told them, “Whatever you do, don’t open access. If you do, we’ll crash.” The invitation code was the only way to control volume, not a scarcity-marketing tactic.
The details were intensely physical. On launch day they called to raise the TPM limit, and the provider asked, “Do you need it next month or the month after?” The answer: “This afternoon.” Cards were physically moved and installed in racks. GCP came through in a crisis, temporarily shifting resources from another project and availability zone. That is what got Manus through its first month.
They removed the invitation code in less than a month. “That was the most responsible thing to do.” He added a barb: “We’ve already cleared this path for everyone, and the cloud vendors are ready. Anyone still using invitation codes now seems a little strange to me.”

42. “If There Was Any Paid Promotion at the March Launch, I’ll Die with My Entire Family”

In response to accusations of excessive marketing, he offered the strongest possible oath: “If there was any paid promotion when we launched in March, I’ll die with my entire family.” The logic was simple. The target was the overseas prosumer market; going viral in China would have no direct benefit and would create enormous pressure. That is why the launch happened after 10 p.m. China time, which was morning in North America.
The domestic wave came from years of goodwill. Former collaborators had become investors and media personalities: “If you make something good and happen to be friends with people, they’re willing to give you a hand.” The overseas wave came organically from the bottom up 3 days later, starting with Andrew Kapasi, Patrick Collison, and Gary Tan, then spreading downward. “The chains on the two sides had no overlap at all.” To anyone claiming overseas marketing: “Come on—could Andrew Kapasi even buy that kind of promotion?”

43. The Backlash, Revisited: Every Manus in the China App Store Is Fake

The most unjust criticism had a simple explanation: “The fundamental reason many Chinese users were cursing us is that what they used was not Manus at all. Every Manus you find in the China App Store is fake—a copycat trying to ride our name.”
He accepts the rest of the criticism calmly. Sudden virality naturally creates suspicion of marketing. The anger of people who hear about something new but cannot access the service is understandable: “That’s a gap.” With limited team bandwidth, Manus could focus only on a limited set of markets. The internal rule was “we won’t deal with this until we reach $100M ARR,” which is why the response was saved until now.

44. Pricing Was Made Up, but the ARR Definition Is Sacred: More Than $100M a Month Ago

The pricing was “made up.” He could not understand why ChatGPT cost $20, then heard that OpenAI had made it up too. Manus started at $20 and $40 and has kept the structure ever since, with $40 a month as the default. In reality, it is credits paid by usage; the free tier is what he calls “social responsibility.”
He is uncompromising on ARR: ARR equals MRR ×12, and only Stripe and mobile MRR count. Annual payments must be spread across the months; otherwise there are too many ways to inflate the number, which would amount to self-deception. Manus has now exceeded $100M ARR, roughly 9 months after the March launch. Monica had already reached about $12M ARR and was profitable at the time. A cash-flow-positive product makes you both bolder and more rational when building a second growth curve.

45. User Profile and North Star: No DAU, Only Agentic Hours

The 3 core user groups are white-collar workers at internet and technology companies who do not write code; freelancers and solo entrepreneurs; and people in finance and consulting. Their common traits are strong self-direction and high-value work.
The metric philosophy is one of the interview’s clearest arguments: “Manus is not optimizing for DAU. It is optimizing for how many agentic hours each user generates.” A high-value user can consume 1,000x more than an ordinary user, and individual users often pay several thousand dollars. That shows up in revenue, not DAU. The company metric is revenue, which is completely different from the logic of mobile-internet startups.

46. The Data Flywheel: Users Teach Agents and Repair Agents

Agent feedback is far richer than chatbot feedback. A dissatisfied chatbot user retries or changes the prompt; an agent user teaches it—“That’s not right. This is what I like; give me an Excel file next time”—and repairs it: “I’ve already fixed the file for you.” “These 2 types of data were extremely hard to obtain in the chatbot era.”
Once usage is large enough, Manus can build parameter-free self-evolving systems. Common failure patterns and user consensus can become native parts of the system: the more people use it, the lower the failure rate, and the fewer rounds required to complete the same task.
He is not blindly devoted to automated evaluation. An architecture that performs well on SWE-bench may not receive high real-user ratings. Users care whether the aspect ratio exceeds 16:9 and whether a website looks good, which is difficult to capture with a fully automated reward model. Manus therefore has an evaluation team of more than 10 people, supplemented by rotating interns, using simple 1-to-5-star feedback to guide iteration.

47. The RLI Benchmark: 2.5% SOTA, and an Interpretation That Could “Accelerate Global GDP”

Scale AI’s new RLI, or Remote Labor Index, is the benchmark he values most. It asks whether work completed by AI is good enough that a real customer would pay for it and could not distinguish the AI from a human. Manus ranks first and is SOTA, beating Claude, Gemini, and every other competitor.
He pours cold water first, then draws the upside: the completion rate is only 2.5%, still far from 100%. The optimistic interpretation is whether AI could accelerate 2.5% of the world’s GDP—perhaps reaching 20 or 30 in 2026. The conclusion remains the same: evaluation is both taste and direction. His analogy for a general agent is a remote worker who can complete infinitely varied tasks through a simple interface—hands on a mouse, eyes on a screen.

48. Head-to-Head with ChatGPT Agent: A Layered Market Serving the Most Demanding Users

The question “What happens if OpenAI builds one?” has already been answered once: “They built it, but the result is obviously worse than Manus.” Manus can use every best model on the market; OpenAI is limited to its own strengths. ChatGPT reaches the broadest base of users, while the number of people with genuine agent needs is objectively an order of magnitude smaller. Manus is spreading downward from Silicon Valley’s elite and serving a higher-priced customer segment. Pick himself still uses ChatGPT for quick answers.
How demanding are these users? In a double-blind test, the team quietly swapped the model for 5% of users and satisfaction fell immediately. “This group always wants the highest level AI can provide at that moment.” Low-price competition is not the threat. Users ask: “You charge me $40. Can I pay $200 and see how much better the result can get?” Pricing optimization is barely on the roadmap.
An a16z survey offers a telling detail. On phones that had ChatGPT, Claude, and DeepSeek installed, the largest anomaly in co-installation probability was always Manus. “Manus coexists with almost every chatbot.” The negative reading is that penetration remains far too low: “We still have a long way to go on marketing.”

49. Anthropomorphizing Agents Is Human Narcissism

He rejects the anthropomorphizing of models: “We are merely using familiar human words to describe them. Models are nothing like humans.” Assigning designer, programmer, and manager roles in a multi-agent system is an inherited mistake. Humans divide work because no human can do everything; models are more capable than humans. “Don’t transfer the limitations of being born human onto an agent.” Information loss and friction from human organizational structures should not be copied.
The correct approach is to recognize that building an agent means building 2 products: “one for humans and one for models.” The model-facing product is not a personality. It is context engineering—constrained decoding and action-space design—to reduce the chance of confusion without sacrificing generality.

50. Pure-Blood Agent vs. Workflow: One Vision Capability Beats a Pile of Guardrails

He divides the market into agentic workflows, which use human rules to maintain stability, and pure-blood agents, where “the entire process and method of completing a task are determined by intelligence itself.” The latter has a much higher ceiling and follows the bitter lesson: use more compute and a general method rather than adding more human knowledge.
Data visualization is the clearest example. The product instinct is to write a pile of guardrail prompts, but “every added constraint reduces the model’s diversity.” Manus simply added the ability to inspect an image. The agent then discovered that a font error was causing Chinese rendering to fail and fixed it, while also noticing overlapping chart elements. The approach moves from patching known holes to letting generalization solve problems no one has discovered yet. “My job is to stop everyone from doing this.”

51. Conservative on MCP, Measured Weak-to-Strong: Do 1,000 Small Things Right

When MCP was surging, Manus was unusually conservative. Dynamically discovering and unloading tools pollutes the action space and reduces KV-cache hit rates; “a lower cache hit rate seriously affects cost.” The team therefore developed a way to call MCP outside the native action space. Anthropic later wrote about the same method in a blog post.
How can an agent framework keep up with models improving on a weekly basis? Lock the current framework, run weak and strong versions from the same model family on the same benchmark, and keep tuning the framework to maximize the delta between them. The goal is to capture the largest possible gain when the next-generation model improves.
“Building agents is like training models: it is more important to do 1,000 small things right than 3 big things right.” His own mistake was trusting small models too much—betting that parametric knowledge did not matter and dynamic tools could compensate. The experiments showed otherwise. Knowledge, memory, and generalization cannot be separated; “large parameter counts are still useful.”

52. Organization: Sandbox Teams for Agent Products, and a Decision Table That Does Not Vote

The organization has a new species of team. A sandbox team maintains the operating system agents use—“like teaching someone who cannot use a computer how to use one.” An agent team handles frameworks, evaluations, and small-scale research, doing work for models rather than people. When outsiders said Manus had only Pick as a researcher and could not keep up with Frontier Labs, he rejected the premise: “The basic assumption is wrong. We hire researchers normally, and our goal was never to keep up with every model project at every Frontier Lab.”
The company almost never votes. “Voting alienates a team. People with different views start serving their own positions.” If the goal is shared, consensus can always be reached; the value of discussion is not necessarily the final answer but the additional alternatives it produces. 小红 makes the final call, while the technical lead has veto power over dangerous shortcuts.
The company has close to 100 people and reimburses every third-party AI product employees use. “As an AI company, we have to let employees know what the frontier looks like.” On AI replacing jobs, he uses the textile-machine analogy: many people stopped being textile workers but gained entirely new work. “The core of people’s fear of AI is that they don’t use enough AI.”

53. The Coopetition Universe: Value Lies on the Edges of the Network, Not in the Nodes

“We aren’t in competition with anyone” has a mechanism behind it. Manus has launched joint products with Notion and Microsoft; Manus appears in Microsoft Agent 365 and the Windows right-click menu. Google Cloud CEO Thomas announced before a room of developers that a feature previously reserved for Manus would now be opened to everyone. Pick’s own workflow is Manus connecting Notion, Granola, and Ashby. “A lot of the time, value lies on the edges of the network, not in its nodes.” He avoids calling Manus an agent OS: users’ files and software are not inside Manus, so it should not claim to be an OS. There will be no world-changing agent OS; every operating system is simply becoming more agentic.
His advice to vertical-agent startups is zero adaptation: “Don’t optimize for Manus. We should adapt to you.” If there is MCP, use MCP; if not, read the API documentation and call it directly; failing that, use a browser to simulate a human. He expects the agent ecosystem to have a star structure: general agents act as dispatchers, while vertical agents hold differentiated data, making a fully connected network difficult.
Overseas vertical B2B agents should bloom. “American founders have lost the courage and appetite to build to C—the exit mechanism is too mature. Chinese people still have dreams.” He likes Sierra, La Gora as pronounced, Replit, Lovable, Claude Code, and Cognition. On the Cursor-versus-Claude-Code battle: Cursor’s talent reserve is no weaker than that of the leading companies, and it has more options. “This war isn’t over.” There has never been competition with Kimi’s agent-model path because Manus has no domestic business. On the overseas model-company mainline, “I don’t think we’ll lose”; in the first fight with ChatGPT, Manus won on performance.

54. Four Lab Takes: Differentiated Strengths, No One Has Fallen Off

He expected one of the top 3 labs to fall behind, based on the simple view that whoever controlled hardware and scaling would win. In reality, each lab is using its strength to pull up the average. Anthropic optimizes for “high-economic-value tasks,” sets the pace through MCP and Cloud Scale, and needs to invest more in compute after its 2 large financings. In agent coding, Anthropic is still the best.
Gemini 3’s biggest signal is that it “strongly proves pre-training can continue.” Its multimodal inputs are ahead by a wide margin, and Google’s index is a durable advantage others will struggle to catch through technology alone. OpenAI has invested most heavily in pure reasoning—use OpenAI models if you want to climb benchmarks. Talent departures have hurt, including several key members of the ChatGPT Agent team and rumors of a shift toward B2B, but “as long as the bottom-up innovation culture remains, OpenAI is still one of the companies most likely to produce a new paradigm.”
xAI is betting on pixel in, pixel out. Elon Musk understands that the chatbot war is over and is betting on all modalities converging into pixels in and pixels out, with deliberately differentiated infrastructure. Meta’s line takes on new meaning after this acquisition: “杨乐坤 leaving may be a positive signal. Perhaps they will invest in more straightforward work with faster results.” The Llama team has already changed several times; its lagging position has “many factors outside the arena.”

55. 田渊栋, the Essence of RLVR, and Thinking Machines: “Success or Failure Mainly Depends on Qwen”

He is highly impressed by 田渊栋’s coconut, or latent reasoning. Recent work from Shanghai Jiao Tong and Tsinghua, likely, suggests that RLVR is fundamentally about improving the stability of pass@1. Non-RLVR models can often sample a correct trajectory with enough attempts; “you are solving the problem in token space through something close to search.” Latent reasoning removes that sampling step and avoids collapse, considering multiple possibilities in nearly parallel dimensions while enabling long-to-short generalization. He finds 田渊栋’s direction more practical—“No, I shouldn’t presume to comment.” He also likes work on attention sink and StreamingLLM.
Thinking Machines’ Tinker API has an excellent abstraction level, built around 4 key APIs, and is well suited to small and mid-sized research teams, though expensive. Its fate depends on the continued progress of open-source models: “Thinking Machines’ success or failure mainly depends on Qwen,” because Qwen has the broadest spectrum of same-source models with different parameter counts. He has not spoken directly with Mira, but people he knows rate her very highly.
His ranking is purely technical: he admires D Man Davis most. “伊利亚 is equally excellent; everyone is waiting for his next assignment.” The more mysterious company, called “Eden” in the original, is “too mysterious for me to say anything.” As for domestic figures, “I won’t rank them. It’s too easy to get hit.”

56. Three Things Commonly Confused Under “Online Learning”

He separates the buzzword into 3 ideas. Narrow online learning continuously updates model parameters. Mass personalization does not need parameterization: collaborative behavior patterns plus dynamically injected prompts and in-context learning may be enough. Multi-LoRA gives each user non-reusable parameters and can reduce inference efficiency; S-LoRA and another work whose name was unclear in the audio do not solve the fundamental issue. Continuous or lifelong learning is the old field he worked on with Maggie.
The test is whether the task distribution shifts over time. “Financial markets might qualify: the right answer today may not be the right answer tomorrow.” That is when genuine online learning is worth pursuing. A coding company optimizing tab completion through acceptance and rejection rates is merely on-policy data collection plus periodic training; the benchmark will saturate quickly. “I haven’t seen an application in agents worth betting on immediately.”

57. It Feels Like 2018: Scaling Has Not Stopped, and Every Scenario Is One Step Short

His description of the current moment is “a bit like 2018”: the frenzy after the Transformer and the first wave of applications enabled by BERT, with everyone waiting for the next leap. The narrow scaling law—the loss curve—can certainly continue declining. The broader claim that each multiple of compute unlocks a corresponding set of new scenarios is less certain.
The center of gravity has shifted from unlocking new scenarios to raising quality. “The biggest problem with agents today is not unmet demand; it is that completion quality still needs to improve. Every scenario is still one step short, and scaling will definitely help.” External factors matter too: Cloudflare’s human checks block agents, websites refuse agent access, and agent payments—Manus is working on them with Stripe—could make many workflows smoother.
“Agents have already exploded,” and 2026 will bring them to a much broader audience. The feedback on Manus 1.5 that made him happiest was, “This version crossed the productivity-tool threshold.” Users were genuinely using it as a primary application to make money. Simple tasks became faster, complex tasks received more inference-time compute, and the average perceived speed improved by 3–5x. One additional product lesson: “A new version number is the best way to help users understand that something has changed.” 1.5 was essentially a package of the iterations from the preceding months.

58. The Steady-State Question: When AI Does Not Consume User Time, the Douyin Constraint Breaks

Mobile internet reached steady state once users’ time had been fully divided up. “Whatever you do, you are fundamentally competing with Douyin.” Agents aim to create value without consuming user time, working asynchronously and concurrently in the background. Human-machine interaction time is no longer a finite constraint. He admits he has not solved the question: “I still haven’t figured out what the constraint producing steady state will be in the AI era.”
He also sees no real network-effect AI product. AI is merely an added value in networks based on other people’s output and networks based on user relationships. What he can identify is a network effect in atomic capabilities: add vision, and Manus learns to check whether the website it built actually works, causing agent capabilities to rise exponentially. Wide Research lets agents dispatch other agents. AI products may monopolize one category of mindshare and build a brand, but ChatGPT is only 3 years old. “If this is merely Google in 2002, it is still too early to say.”

59. The Next 3 Months: Proactiveness—Agency Is the Point of an Agent

Users themselves are part of the boundary. Entering prompts is annoying, and much of the relevant context never makes it into the prompt. The next step is proactiveness: “That should be the meaning of agent. Agent comes from agency.” The team already has a prototype that “we love using internally.”
The use case is concrete. After an interview, feedback must be entered into Ashby, while the notes sit in Notion and Granola. “Manus could review my Notion every morning before I wake up, fill in Ashby, and only ask whether I accept it.” He explicitly does not want a ChatGPT Pulse-style daily push, apparently: “That consumes the user’s time. Our users want productivity.” His 2 predictions for the next 3 months are shorter interaction time with higher output value and another substantial cost reduction. “Claude Opus 4.5 has already given everyone a little hope.”

60. A Design Agent for Non-Designers: Enhance People, Don’t Replace People

A teammate who had worked on editing software before Tencent acquired the company gave him a framework that stayed with him. Building an editing agent for professional editors is “extremely difficult”: professionals assess work through risk control, and the workflow is multiplicative—one bad step produces a zero. The right customer is someone who is not an editor but has editing needs, such as a creator. That is net new value.
Manus has therefore never approached the problem as replacing people. It wants to enhance people: “Increasing the output of your most effective employee is both more constructive and more realistic.” The same logic applies to vertical agents. Lovart serves designers; Manus serves non-designers with design needs. “Manus could watch Slack for you, then use Lovart to finish the design.”

61. Singapore, Air Conditioning Switched Off at 10 p.m., and the “Fleeing China” Accusation

Singapore was chosen for 2 reasons. Offices in Beijing and Wuhan had exposed weak cross-region collaboration, so the team wanted to work in one place. Serving a global market also required a compliant headquarters. SOC 2 Type 2, ISO 27701, ISO 27001, and GDPR have all been completed. “Calling it fleeing is wrong. We’ve always had a Singapore entity and have always served the global market. You go where your customers are.”
The office is a WeWork inside a shopping mall, “which can dynamically scale like cloud computing.” The mall’s air conditioning switches off at 10 p.m.; “how long everyone stays mainly depends on how long they can last without air conditioning.” He reaches the office at 10:30 every day. Singapore is boring, which is a major advantage for a startup. The China market is difficult in the short term: “We don’t have enough money to subsidize users, and agents are genuinely expensive. We have to make sure we survive first.” Many similar products have appeared domestically after Manus, which is good, but they will face the same questions: how long can subsidies continue, and how does commercialization work?

62. The Launch-Week Survival Log and 2 Major Worries

The 2 weeks after launch turned day into night. The team supported every time zone, the system went down daily, and people slept 3–4 broken hours at a time. The bookshelf filled with supplements. Everyone in Wuhan who could fly to Beijing did so. Meetings with investors were held standing up. During the day, most of what they saw was criticism; positive feedback from investors was the warmest part of the day. “People who understand still understand.” Virality did not make them complacent: negative feedback still outweighed positive feedback. “We’re all old hands with something stuck in our throats. Hard work still has to prove itself.”
The 2 worries are external and internal. Externally, they fear losing their distinctiveness. Internally, they fear Manus becoming complicated. “The easiest way to grow a product is to add more features, but every addition dilutes everything else.” Is the threat being cut off by a competitor? If you are already considering supply disruption, the product must be very good. The greater risk is losing the unique value that made it matter.
The pessimistic forecast is that Manus dies next month. “For most startups, including us, there is no inherent right to survive. The right to survive is something you earn by continuing to run.” The optimistic forecast is to give every high-value white-collar worker a 7×24 “love partner” that keeps reasoning without pause, as Pick put it. If Manus died next month, “I’d take a break. I’m exhausted.”

63. Quick-Fire Questions and the Final Foreshadowing

His closing views were equally direct. The AI bubble objectively exists, but that does not mean this AI wave is useless: “Human history has done things far crazier than this.” He is no longer buying NVIDIA and is not trading stocks at the moment. Asked 黄仁勋 what could surprise him over the next year, he answered, “Nothing will surprise me,” which Pick sees as the archetype of AI optimism. He agrees completely with 姚顺宇, especially the views that large companies are now copying startups and that the second half is beginning. He agrees with 杨志林 that any problem that can be defined can be solved, but strongly rejects the claim that building agents without training models is reverse engineering: “We are not doing reverse engineering. We are setting the pace for everyone else.” On world models: “I don’t understand them, so I won’t comment.”
The personal answers were mac and cheese, Beijing, and the fact that kelp is not an animal—important because he is allergic to seafood and often has to explain it. He is reading the picture book 《线条小狗》. For the paper that most influenced AI, he chose one whose name he could not remember and Flan-T5: “Everyone will say attention is all you need. I don’t want to say that one.” The most important fact from his current vantage point is this: “The next advances require users to participate.” Looking back at the line about writing memoirs when he was old, Meta’s full acquisition a month later supplied the final footnote.

Your Brain Doesn't Command Your Body. It Predicts It. [Max Bennett]Market Overview — December 30, 2026