Summary

XBench moves model evaluation from “smart” to economic value. It began as an internal Sequoia tool in 2023 to track frontier-model progress month by month, then shifted to open source and paper publication in 2025; 龚源 argues that benchmarks ultimately determine what the industry optimizes for: “If you only test whether a model is smart enough, it will only become smarter.” The question now is whether a model “has economic value and can actually get work done.”
HongShan X is trying to institutionalize Sam Altman’s function. This is not about investing first and providing resources afterward; it means working with scientists before a company is formed to define its technical and commercial strategy, team, financing and equity structure, and to write the institution into the company’s DNA. 龚源 uses Sam Altman and Ilya’s partnership to explain the role: Ilya was critical to the technical breakthroughs, while Sam Altman helped define the ChatGPT product; the two were “indispensable to each other.” HongShan wants the institution—not a star investor—to perform that function.
The creation model is not pursuing assembly-line incubation, but certainty around scarce opportunities. 龚源 estimates that the market produces roughly 5 unicorns a year, or perhaps 6-7 in a strong cycle; HongShan X has produced about 1 a year over the past 3 years and now aims to reach 2, lifting creation’s market share from roughly 20% to 40%. “There are only so many ideas in a year,” so the model cannot pursue many projects; every one must become that year’s star company.
One move keeps recurring in 龚源’s career: when there is no seat, she sets up another table. After 2016, the mobile-internet table was full, so Huachuang used a process of elimination to move into new computing, autonomous driving, robotics and commercial space; by 2022, tech deals had become crowded club deals in which investors camped outside the door for a TS and received only 2%-3% in the first round, prompting her to leave the old playbook again for HongShan X. Her underlying judgment: “You cannot beat the people who own the original table while sitting at that same table.”
Early tech investing led her to an important revision: in hard tech, the sector often matters more than any single company. Cases including DeePhi Tech, Cambricon, LandSpace, WeRide and Hesai showed that hard tech usually lacks mobile internet’s winner-takes-all dynamic, so investors cannot be judged only by whether they picked the single winner. Her conviction at the time was not a fully underwritten return case, but a question prompted by seeing a group of top scientists: “If the future public companies aren’t built by them, then who will build them?”
Sequoia taught her that the goal is to become the earliest and largest shareholder—not simply to invest earlier. Using Mech-Mind as an example, she saw the company at a RMB50M valuation, then still had to keep underwriting it at $500M and $1B valuations; Pony told her to imagine what she could post when the company went public. True conviction may mean investing through 7 rounds, treating every financing—and every short-term jump in price—as a fresh war over whether to stay in or get out.
An IC is not meant to turn everyone into a 7; it is meant to make non-consensus views fully visible. A 6 means “I agree if others invest, but I don’t want to invest myself”; an 8 means “we have to invest”; a 9 means being willing to bang the table for it. Good deals often score an 8 from the deal team while everyone else lands at 4, 6 or 7. 龚源’s rule: “You don’t have to turn someone else’s 4 into a 7. Just make your own 8 clear.”
AI may give some AI scientists 100x or even 1,000x the productivity, spawning solo companies, one-man shops and new corporate forms; VC, as a factor of production, will necessarily change too. But she does not think the standard for judging investors will change. Asked in 2019 whether she preferred money or investing, she answered “both”; today, her answer is: “Of course I love money, because money is its own first principle.” The method can be rewritten, but long-term returns remain the final test.

Deep dive

1. XBench Takes Model Evaluation from “Smart” to “Can Get Work Done”

XBench began as an internal Sequoia tool established in 2023 to track what was happening in large models every month. By 2025, the team decided it should serve more than investment judgment, so it open-sourced the tool and co-published a paper with external institutions.
龚源 said Sequoia “should be the first VC in the world to publish a paper.” Her rationale was not brand marketing: a benchmark is an exam, and deciding what to test amounts to telling the model community what questions to answer and which direction to pursue.
The key shift is from asking whether a model is “smart enough” to asking whether it has economic value and can actually get work done. If an evaluation rewards intelligence alone, models will simply keep getting smarter; bringing the technology into the real economy requires testing whether it can complete real tasks.

2. What HongShan X Wants to Institutionalize Is “Sam Altman’s Function”

HongShan X was established about 3 years ago. It was not publicly promoted and had no fixed mandate. 龚源 compares it with Google X: an innovation base that looks for new paradigms from first principles and is “defined by demand, like X itself.”
It aims to be more than an investor: it wants to serve as the company’s earliest and most important strategic partner and co-founder, jointly determining the technical and commercial strategy, financing plan and team, while building R&D capabilities such as XBench and discussing the actual technical path with scientists.
龚源 uses OpenAI to explain the combination. Ilya was critical to the technical breakthroughs, while Sam Altman helped define the ChatGPT product; “they were indispensable to each other.” Sequoia’s experiment is to institutionalize Sam Altman’s function, distributing one person’s compound capabilities across an organization.
Institutionalization makes the model replicable, but not high-frequency. 龚源 estimates that the market produces roughly 5 unicorns a year, perhaps 6-7 in a strong cycle; HongShan X has completed about 1 creation a year for the past 3 years and now wants to reach 2, lifting creation’s share from 20% to 40%. But “there are only so many ideas in a year.”

3. Beijing’s “Black Map” Led 龚源 into VC

龚源 graduated from college in the US in 2011, then spent 4 years in Silicon Valley doing strategy consulting, diligence for VCs and an attempted startup. In 2015, she moved back to Beijing for family reasons, arriving somewhere for the first time without a purpose.
She compared the resulting identity crisis with being thrown into a completely black map in a video game. Chinese Americans in the US were a narrow “cross section” in terms of both identity and career; Chinese society suddenly appeared three-dimensional, and she had to figure out which layer she belonged in and what she was suited to do.
For 3-4 months, she sent out resumes indiscriminately and visited financial institutions and SOEs on Financial Street, internet companies in Shangdi and Wangjing, and institutions around Huamao and Guomao—“lighting up the map with her own feet.” A headhunter she met at 3W Coffee eventually decided that Huachuang matched her temperament and brought her into VC.

4. Lechun’s RMB7M Equipment Purchase Proved Venture Lending—and Exposed Its Limits

When 龚源 joined Huachuang, she was discussing not VC but the then-newer business of venture lending. She had seen startups borrow from Silicon Valley Bank in Silicon Valley, and Huachuang wanted to bring the financing model to China early.
The first project was Lechun yogurt. The company needed to import a centrifuge for producing Greek yogurt, costing about RMB7M. Equity financing would permanently dilute the company for that RMB7M; with debt, the company could expand capacity, lift its valuation in the next round and repay the loan.
IDG, sitting nearby, offered a VC role while Huachuang offered a VL role. Lechun ultimately chose debt, creating a powerful positive feedback loop. But after more than a year of business development, the venture-lending operation was halted because the cost of capital was too high, and the team shifted to traditional VC investing.

5. With No Seat at the TMT Table, Huachuang Opened a New Tech Table

After 2016, the projects that would later become $10B mobile-internet companies had largely been allocated. 刘旌 called 2016-2023 the “historical garbage time and wasteland for TMT investors”; 龚源’s firsthand observation was more direct: “The big names had already filled every seat at the table.”
Huachuang chose its strategy by elimination: it did not understand healthcare, could not compete with the incumbents in TMT, and consumer was not at the right point in the cycle. That left technology as the most worthwhile table to open. 龚源 said the decision was not difficult: “If you can’t do this or that, what else can you do?”
The table consisted of only 7 or 8 scattered people from firms such as Northern Light, Lightspeed and Shunwei. Many had just entered the industry and were eager to prove themselves. The market was busy with themes such as the sharing economy, but she did not feel lonely: “The excitement belongs to them. It has nothing to do with us.”

6. Her Scientist Network Grew from 4 Tech Trees and Tsinghua Alumni Ties

The first signal was that large numbers of young Chinese scientists had already proved themselves in global academia. 龚源 cited 韩松 as an example: a Best Paper at a major conference and the prospect of a faculty position at MIT showed her that this cohort occupied a genuinely world-class technical position.
She was not trained as an engineer, but still had a foundation in mathematics, physics and competitions from middle and high school. More importantly, she liked scientists’ purity and the “light in their eyes.” In her view, the best scientists can explain a problem in plain language; people who rely only on obscure jargon “may not actually understand the subject deeply enough themselves.”
Huachuang initially chose 4 explicit themes: new computing, autonomous driving, robotics and commercial space. The method was to identify the leading company on each technology tree, enter through an anchor project and learn from the founders; at the time, “being able to join a follow-on round was itself a skill.”
Once the investments were made, the network began connecting itself. Many founders came from Tsinghua’s electronic engineering department, with cohorts ranging from “97, 98, 00 and 02”; people would first ask, “Which class year are you?” Shared professors and friends became trust links. After opening up the network over 1-2 years, she found that scientists building companies were actually part of a relatively small circle.

7. Hard Tech Is Not Winner-Takes-All; Sector Judgment Matters More Than a Single Bet

Huachuang made multiple one-of-several choices: DeePhi Tech rather than Cambricon in chips, LandSpace in commercial space and WeRide in autonomous driving. It invested in Innovusion for lidar but missed Hesai. Looking back, 龚源 says the LandSpace choice was clearly right, but the misses drove further reflection.
Her revised conclusion was that the sector matters more than any single company. Hard tech does not have mobile internet’s winner-takes-all characteristics, so an investment should not be judged only by whether it picked the one ultimate winner.
The non-negotiable threshold was that founders first had to be top-tier in the scientific arena, because “tech investing starts with technology”; they had to be people holding advanced productive forces. She did not yet know whether these hidden corners could make money, but believed the group would move to center stage: “If our A-share listed companies 10 years from now aren’t built by them, then who will build them?”

8. Before Public Methodologies, “Jianghu Masters” Were the Investment School

龚源 invested in more than 20 companies at Huachuang. The decision process was not complicated; the team learned by doing and discussed by doing, much like a startup. 海燕 was important in bringing her into the industry, but 龚源 also described her time in the jianghu as a period when someone would suddenly say, “I’ll teach you 2 moves.”
There were no mature self-media channels, podcasts or YouTube interviews, and even 1 or 2 good books on VC were hard to find. Things now considered basic—user reviews and product evaluations—required someone to “open the top of your skull” and show you how to look.
A veteran who had angel-invested in DJI and later invested in Hesai taught her to read Amazon reviews and YouTube product evaluations. 龚源 considers herself lucky, but says luck came with a condition: she was willing to be helped and to work extra weekends and nights to complete the research the senior investor needed.

9. Joining Sequoia Was Not a Rejection of Huachuang; She Had to Look Again While Young

In 2019, as the market collectively migrated toward technology and To B, Sequoia approached 龚源 after 4 years at Huachuang. She hesitated for a long time, eventually landing on a reason similar to going to the US after high school: “I’m still young. I want to go see it.”
The move made sense professionally but was emotionally difficult. She and 海燕 cried in each other’s arms more than once. “老熊” even secretly had her fortune told, intending to stop her if the result was bad; the reading was excellent and said she was short on wood, which became the backing for letting her go.
During interviews with Sequoia’s frontline partners, she felt the difference between “the formal army’s martial arts” and those of the jianghu. 周逵 asked whether she “liked money or liked investing,” and she answered “both.” Neil asked about her career ceiling; she said she had always been an elementary-school teacher who produced good students but could not accompany them through high school, college and pre-IPO.
She asked Neil whether he worried about “the innovator’s dilemma,” because she believed success could only come from innovation and worried that a large organization could lose its ability to innovate. Neil did not deny the concern; he acknowledged it. 龚源 felt that candor itself answered many questions.

10. Sequoia Made Invisible Competition Visible—and Cut 3 Hours to 2.5

Her first assignment after joining was to rebuild the portfolio: reinvest in high-quality companies she had backed at Huachuang, including Mech-Mind, LandSpace and XREAL, while adding projects in robotics and improving the one-of-several choices she had made previously.
The adjustment was difficult because collaboration and competition became visible. Previously she only had to execute her own deals; at Sequoia she could see other people’s IC presentations, system activity and complete deal flow. She remained anxious until a senior colleague told her: “You are just seeing what is in other people’s bowls. What does what is in their bowls have to do with you?”
She later used the line to reassure younger colleagues. The strongest investors in the market had always been competing for the same deal; their competition was simply invisible elsewhere. At Sequoia, competitors appeared in the office, making it feel as though competition had suddenly increased. In reality, “it had nothing to do with you before, and it has nothing to do with you now.”
At a 2019 team-building event at Qingcheng Mountain, she spent more than 3 hours scouting the route; during the actual competition, everyone finished in 2.5 hours. She also had nightmares about every tech investor moving to Sequoia, but summarized the culture this way: “Having nightmares while climbing to the summit in 2.5 hours.”

11. The Hard Part Is Not Firing Early; It Is Making the Irrational Follow-On Bet

From 2019 to 2022, Sequoia taught her a new set of skills: investing in companies at Series C and beyond, sometimes even after Series B, and writing large checks. Early-stage failure is relatively common; later-stage failure is less tolerated, and the larger the check, the greater the required depth of diligence and conviction.
She came to understand that liking a company does not automatically produce returns. Only 2 things truly matter: whether you are the earliest investor, which determines the relationship with the founder, and whether you are the largest investor, which determines how much you ultimately make. Ideally you are both, which may mean investing through 7 rounds.
Mech-Mind was the clearest example of the time span involved. She saw the company at roughly a RMB50M valuation at Huachuang; Sequoia later led its round at a $100M valuation, after which the company quickly reached $500M and $1B. Having seen the low price and then continuing to accept higher prices is the most counterintuitive move.
In the $500M round, Pony asked her to imagine every investor posting on social media when the company went public: “What would you post? If yours looked like everyone else’s, you were just one of them.” The team therefore led the round again and increased its position. She came to believe that every financing must be a fresh war over whether to stay in or get out.

12. The IC’s Job Is Not Consensus; It Is an Accurate Display of Disagreement

Sequoia’s IC scores deals from 1 to 10, with no 5. A 6 means, “I agree if others invest, but I don’t want to invest myself”; a 7 means you want to invest; an 8 means you have to invest; and a 9 means you would bang the table for it. A 4 recommends not investing, while a 3 means banging the table to stop the deal.
龚源 believes that everyone scoring a 6 may mean nobody truly cares; everyone scoring a 7 often means the project has become consensus, the price is too high or the available allocation is too small. The deals most likely to make money are those where the deal team bangs out an 8 while the rest of the room is distributed across 4, 6 and 7.
The IC is therefore neither collective brainstorming nor debate and persuasion; it is “the expression of conviction.” Each person represents a real voice in the market. Put together, those voices show the institution the degree of consensus in the external market.
A partner’s job is to ensure that junior investors have both the courage and the channel to express a 4, 8 or even 9. “You don’t have to turn someone else’s 4 into a 7; make your own 8 clear.” She also believes Sequoia does not need a “silver bullet”: its scale and breadth of views are sufficient, and the IC itself works.

13. The 2%-3% Club Deal Made the Old Playbook Fail in 2022

By 2021-2022, 龚源 believed she had completed the first phase of her mission: building the portfolio and learning to generate good returns on a large platform. But after liquidity flooded the market, tech deals showed signs of over-investment in scientists. Everyone chased the same projects, creating crowded club deals.
Her direct question was: “If everyone owns only 2%-3% in the first round, how do you make money?” Continuing with the same method made it difficult to preserve the returns and performance she had previously pursued.
Combined with illness in her family, she took roughly 6 months off and treated it as a sabbatical. She never considered leaving investing and did not view the choice of institution as the primary issue. What mattered was first answering: “How should investing actually be done?”
Her first move was to force a stop and pull away from market momentum; the second was to discard the assumption that “this is simply how VC is supposed to work.” She asked again: if she did not take VC for granted, what should investing look like in that market environment?

14. HongShan X Began with One Question: What Do Tech Founders Need Besides Money?

Her starting point was that today’s founders and those of the next 10 years would be tech founders. The question was whether they needed only money, or also the people “standing at the door every day, demanding that the TS be signed today and then wiring money like crazy.”
To find the answer, she returned to scientists, both those she had backed and those she had not, and watched what they actually lacked. The conclusion gradually pointed to “understanding”—understanding the technology, the shape of the company and the path to success, rather than merely supplying capital.
The prototype of HongShan X emerged from that process. OpenAI already existed, and Sam Altman and Ilya had begun working together, but there was no result as visible as today’s to serve as a benchmark. The team was not copying an existing template; it started from demand and used an undefined X to preserve room for exploration.
龚源 believes this was not an isolated event in Sequoia’s history. The institution began with VC, then gradually added Growth, buyout, secondaries, infrastructure, seed, and finally Creation and X. “None of these things existed on day 1”; the organization has always expanded its boundaries through innovation.

15. AI Scientists with 1,000x Productivity Need a Different Partnership Structure

刘旌 offered the traditional VC view that good founders are difficult to train and difficult for outsiders to supplement. 龚源’s response was that this describes 2 different eras: in the internet and business-model innovation era, an exceptionally strong founder might be able to do everything; the structure of productivity is different in the AI era.
Many AI scientists were born after 95, some even after 1998, yet possess advanced productive forces that could reach “100x or 1,000x,” while their social experience may be close to zero. That mismatch makes a partnership structure more necessary, not less, though it cannot be imposed prematurely.
The best scientists will actively seek a partner who can raise their productivity and probability of success. 龚源 stresses: “I am definitely not interviewing him; he is interviewing me.” The investment firm is enlisting; the scientist is the general leading the troops to the front line.
She has no universal template for equity, the primary decision-maker or the allocation of roles. Every company is different, but there must be a central figure. The partnership must begin with the scientist recognizing that they need the best possible counterpart—not with an investor trying to prop up a weak founder.

16. Co-Creation Is Written into the Company’s DNA; the Difference Happens Before Investment

The central difference between HongShan X and traditional post-investment enablement is that one happens before investment and before the company is formed, while the other happens afterward. What the company should do, how its technical and commercial architecture should be built, how equity should be divided and how the team should be assembled are all part of co-creation.
龚源 says this allows HongShan to be “written into the company’s DNA.” The Sam Altman-style function is not performed by her alone; it is distributed across the institution, including Neil, 周逵, 龚源 and other team members, with different people combining around strategy, technology, management and capital operations.
She considers 无问芯穹 an early case that should count as a HongShan X project. She initially said it was founded in 2022, then corrected herself to “the first half of 2023.” The first phase focused on co-creating with university professors and scientists; the second shifted toward recently graduated “young prodigies” and younger scientists.
The first 3 months of each project generally require 100% commitment. 龚源 does the opportunity-cost math plainly: those 3 months could otherwise fund roughly 10 investments, so creation’s cost is 10 deals left undone. If the target is not sufficiently likely to become the market’s star company that year, it does not merit that allocation of time.

17. Every Project Makes Her Want to Join; Institutionalization Requires Her to Stay

龚源 admits that during the first 3 months of every creation project, the physical reaction becomes: “I want to join this company.” Without that total commitment, the project is probably not good enough.
She ultimately does not join because the force that makes the company work is not just her. Unless she can bring the entire team in with her, the institutionalized effect cannot be replicated. She admires 张予彤’s courage in joining 杨植麟’s team, but still chose the institutional path under Neil’s influence.
In her view, institutionalization turns the work she likes into something “evergreen.” Sequoia’s 20 years of accumulated experience are concentrated at one point and handed to a founding partner who is co-building the company. She believes this may be why the model needs to be realized at Sequoia.
What she truly likes is the build process of taking something from idea to implementation, and she wants to repeat the 0-to-1 process many times. She jokes that she came to Sequoia hoping to see college but discovered she “likes kindergarten better.” Yet she may still spend half her time on later-stage companies: without having seen college, it is difficult to design kindergarten well.

18. ChatGPT Shifted the Focus, Not the Method of Betting Along the Technology Tree

Before ChatGPT launched, 龚源 focused more heavily on hard tech in the Chinese context; afterward, HongShan X naturally devoted more attention to AI. What changed was the focus, not the way of looking at the problem: she still follows the full technology tree in search of paradigm shifts.
Her expectations for 2026 extend beyond applications. Models may develop new paradigms such as self-evolving, prediction models and world models, while the application layer still offers extensive room to build.
Farther down the stack, compute is pushing 3D stacking and new computing methods; on the energy side, teams are exploring both nuclear fusion and fission; other new computing methods include quantum computing. Multiple layers are evolving at once, making the technology world exponential rather than linear.
The investment challenge is therefore not “where are the fish”—today, fish are everywhere—but how violently a paradigm will arrive. “Only by predicting the intensity of the paradigm’s arrival” can investors know where to catch the biggest fish.

19. Missing the Hot Frontier-Model Deals Is the Visible Cost of a Certainty-First Strategy

刘旌 noted that China’s primary market became active again after ChatGPT, with Sequoia investing in leading companies including Zhipu and MiniMax, and asked whether 龚源’s personal participation was limited because so much of her time went to HongShan X. 龚源 acknowledged that this was a “question worth reflecting on” and admitted that not every objective can be optimized to the maximum at once.
If the market produces only about 5 unicorns a year, her first priority is to ensure that 1 comes out of creation, then try to reach 2. She then asks whether the remaining time and energy can be invested in the other 3 projects.
Leading projects that emerge from the public market contain substantial randomness: when you meet the company, when you encounter the opportunity and whether you can secure an allocation are not controllable by one person. Chasing only random outcomes may not produce a reliable method.
She chose to expand certainty first: capture the random opportunities when they appear, but if they do not, raise creation from 1 company to 2. “I have been lucky enough to receive sufficient feedback on certainty,” she says, but that luck does not erase the cost of missed opportunities.

20. What She Values Is Not Simply “Being Like Me,” but a General Worth Serving

龚源 describes her own taste as an extreme desire to succeed, a desire to win, curiosity and first-principles thinking. But what she values more is whether a founder is a general capable of leading troops.
She sees many peers interviewing CEOs from a client-side perspective; her own approach is from the service-provider side. She asks whether, if she worked for him, she would have “meat to eat” after success, whether he would fail or abandon her halfway through. She is not hiring a CEO; she is deciding whether he is worth serving.
Traits such as resilience and honesty can be polished in an interview. Behavior is more honest: if she left Sequoia today, would she follow him? If he called at 12 a.m., would she answer? The standard requires no investor ego, but it does require an extremely high bar.
She once worried that as she grew and met more people, her standards would become too high to function: at 36, could a 25-year-old still attract her enough to make her think, “I want to serve him”? Her answer today is: yes, still.

21. Sequoia’s Core Culture Is “Wanting to Win” and Conviction

龚源 received the call informing her that she had become a partner at 6 p.m. on Lunar New Year’s Eve, while her family was in Sanya preparing dinner. The surprise came from the timing and her family’s reaction, not the title itself. She says the work did not materially change; the real pressure began on the day she started HongShan X.
If Sequoia’s many values are reduced to their foundation, she chooses “wanting to win and conviction.” VC amplifies individuals, and every project needs an owner. One person who wants to win is not enough; it takes a group that wants to win and is willing to take responsibility for non-consensus views.
A many-to-many organization provides the structural support. A junior can work with any senior, and if they cannot find a sponsor, they can even email any senior directly. Seniors cannot retreat to their own mountains and become managers; they must remain in combat. Cross-disciplinary discoveries often emerge in the “valleys,” and a many-to-many structure also reduces losses in organizational gaps.
She warns younger colleagues that the greatest risk is over-interpreting business mechanisms as personal relationships and office politics. “The most effective approach is to forget about personal relationships.” Every organization can develop gaps between what the system rewards and what creates career value, but she believes Sequoia self-corrects and that such problems will surface within 3 years.

22. AI Will Rewrite the Shape of VC, but Not the “Make Money” Yardstick

Once AI raises productivity by 100x or 1,000x, solo companies and one-man shops will proliferate, and the company as an organization of production relations will necessarily change. VC, as a factor of production, will also be redesigned. 龚源 therefore expects AI-era VC to be very different.
The change will not touch the first-principles standard for evaluating investments. Re-answering 周逵’s 2019 question, she no longer says “both”: “Of course I love money, because money is its own first principle.” The method can change, but investing still has to return to the first principle of making money.
刘旌 suggested that the primary market should also publish a leaderboard. 龚源 believes an annual ranking would encourage short-term behavior and is considering whether a rolling 10-year ranking could work instead. VC is a 10-year industry; highly deferred cash returns are ultimately what validate attention, titles and methodologies.
She went to the US early in her career because she “wanted to do something different,” completing her undergraduate and graduate credits in 3 years. Her 2019 goal was to “survive at Sequoia and do something different,” which she believes she has achieved. Her mother sees becoming a partner as the best possible outcome; to 龚源, it is only the beginning: the next peak is to define the next destination herself.

🔬 "The Most Innovative Diffusion Research Is Happening in Drug Discovery, Not Image Generation"1000 Designs a Day: Neural Concept's Thomas von Tschammer on AI-Native Engineering