In the AI Era, Taste May Be Venture Capital's Last Line of Defense

“I thought the match was already over. Now they’re saying that was the end of the first half, it’s half time, and we come back for the second.” — Liang Jie, on why he’s back at the table

Guest: Liang Jie｜founding partner, Impa Ventures; previously Walden International, Sequoia China, China Growth Capital

In 2015, a young VC spotted Pinduoduo at the edge of the stage, recommended it, and didn’t get the deal. Ten years later an essay dug the episode back up and handed him a label: the man who missed Pinduoduo. His own first reaction was surprise — what is there to write about here, it’s a thing in the past, and not a thing that produced a result.

In this episode Raymond talks with Liang Jie. A VC veteran who entered the industry in 2011, he came up through Walden, Sequoia and China Growth Capital, then went out on his own in 2021 to run Skyline; in 2024 he founded Impa Ventures with two former colleagues, investing only in AI, with seven or eight deals done. Zhu Xiaohu says he wants to sit 15 degrees off industry consensus; Liang Jie says he wants 20: he doesn’t touch big-tech alumni pitching a story, on AI hardware he says flatly “we don’t understand it,” and on video and animation he looks but moves cautiously. Ask him about the one that got away over the past year and his answer isn’t a deal he missed but an entire company — Anthropic. “We certainly underestimated it enormously.”

So in the AI era, what’s left for private markets? His answer is taste: the scarcest data is always successful data, and the most extraordinary kind is the scarcest of all. Someone who has seen the most extraordinary companies simply has different taste.

What follows is the full conversation, edited and condensed.

1. “If I’d missed a Mobike, nobody would talk about it”

Raymond: Today we have Liang Jie. We met in the mobile internet era, when he was looking at Chinese internet companies going global. But he’s better known as the investor who missed Pinduoduo very early. In the serialized history of Chinese venture capital that the journalist Liu wrote, there’s a young man who spotted Pinduoduo very, very early and, through a series of circumstances, didn’t get the deal; his firm Sequoia doubled down in later rounds and it became a hugely important case at the time, while the person who first recommended it later left Sequoia. I’ve known Liang Jie a long time, and it was only recently, when Liu posted another piece on Xiaohongshu, that I learned Liang Jie was the man who missed Pinduoduo.

If I took Liang Jie to meet a new acquaintance today, I’d say: Liang Jie, missed Pinduoduo ten years ago. But I hope that next time I introduce you, there’s a different title.

Liang Jie: Thank you for having me. A new title next time is what I’m hoping for myself.

I’ve known Liu a long time; the earliest was when he was a reporter at 36Kr and we did an interview. Pinduoduo was still fairly new then; it IPO’d in 2018. Later, when he was writing that piece, he mentioned once asking whether he should write about me, and my first feeling was mild surprise: what is there to write about here? It’s a thing in the past, and it isn’t a thing that produced a result.

The reason people are willing to discuss it is that Pinduoduo’s success is on too large a scale to avoid. If what I’d missed were a HotMaxx or a Mobike, nobody would talk about it. Too ordinary.

The origin that day was actually that we’re building a new firm and have done eight or nine deals. But more importantly: that we still get to participate in a technological shift as large as AI is in itself a privilege. Everyone’s vintage differs. A person’s effective working life runs from 20 to 60, forty years, and hitting two large technology waves in that span is an extraordinary thing. That’s already very good luck.

Raymond: So how did you come to start Impa Ventures?

Liang Jie: My path is simple: grew up in Jiangxi, went to university at sixteen, studied science and engineering, worked in industry for five years after graduating, and started doing VC in 2011. Walden, Sequoia, China Growth Capital one after another, then went out on my own in 2021 to do Skyline. In 2024 I met the other two founders of Impa Ventures — one a former colleague from China Growth Capital, the other his colleague at a subsequent fund — and they said, why don’t we happily start a fund, invest in early-stage technology centered on AI, and do something significant.

The reason Pinduoduo makes such good material is that after 2015 it became genuinely hard for China’s internet VC world to get into anything. Excluding EVs and hardware, there are only two internet-related examples: Pinduoduo and ByteDance. Both started at the edge of the stage; they certainly weren’t the deals everyone crowded around at birth, they climbed slowly, slowly from the edge to the center. Which is why I’ve always thought this is impressive: being able to spot that tree at the edge of the stage — whether or not you manage to climb it — spotting the thing at all is still remarkable.

2. “Essence plus extreme”: Pinduoduo didn’t miss AI

Raymond: Impa Ventures is a new fund born after GPT, looking mainly at AI. So following on from Pinduoduo: over the past two or three years Pinduoduo has very clearly invested zero in AI — zero as far as anyone outside can see, and zero as far as I can feel as a shareholder in the US-listed stock. Do you think it missed this one, or do you see it differently?

Liang Jie: Pinduoduo’s attitude toward AI today is consistent with how this company has always been positioned. Or rather, Pinduoduo’s ultimate style is Colin Huang’s personal style. If I had to summarize Colin Huang’s style, it’s two words: essence — he thinks about everything at a very fundamental level; and then, having thought about it that fundamentally, he executes to an extreme. Essence plus extreme.

So in Pinduoduo’s view, what it needs to do now is be the world’s best value-for-money e-commerce, that battlefield is big enough, and there’s a long time left to work in it. It sees AI as a tool it will use when appropriate — and it may well be using it perfectly well today. But do I want to build a consumer AI app? That’s certainly not what it’s thinking about now. Pinduoduo is already a company whose daily and monthly actives in the US are roughly level with Amazon’s, and there’s still a lot of room.

Raymond: So how have you positioned in AI over the past few years?

Liang Jie: We’ve done seven or eight deals so far, concentrated in three directions.

The first relates to AI infrastructure. Infrastructure is an extremely broad word — models are infrastructure, and so are chips; on-device chips, inference chips, plenty of people doing those. The relatively hard, capital-intensive ones certainly aren’t a strength of a fund our size. What we look at is the softer, middle-layer things — a piece of inference middleware, a connection protocol. What’s it for? Pushing AI’s capability further up.

The second is model capability moving downward. Suppose model capability froze right here today. In some industries it’s already entirely usable; in others it isn’t enough. But either way it has to go into every vertical industry. There’s a word, deployment, which I find quite accurate. And what does that deployment require? It requires you to get inside that industry’s setting, understand its workflow, and understand in what product form the thing should appear at which step of that workflow. Companies like that I’d call vertical applications. We invested in a healthcare company in Australia doing all the admin steps in the medical workflow. It’s highly bound to the setting; it isn’t simply stuffing a large model in, it’s wrapping the large model so it can interface with that industry’s particular interfaces and APIs and genuinely take root in the industry’s setting.

On consumer we invested in a language learning company, called Talk AI domestically; we invested in its overseas entity. Domestically it’s already number one in one-on-one spoken practice, with revenue and profit, but the ceiling on domestic willingness to pay isn’t high, so the center of gravity has to be overseas. We encouraged the founder to go a hundred percent all in on overseas, and then we invested in the overseas entity — overseas starts from zero.

Raymond: A bit like YY spinning out Bigo when it went global. So is Talk AI consumer, or vertical, in your book?

Liang Jie: I’d probably put it more on the vertical side. Because in my mind, consumer in China right now means Doubao and Qwen — the ones that go buy advertising on Focus Media. That’s consumer.

Raymond: That definition is close to ours. Frankly we don’t dare invest in consumer: either big tech builds an identical feature, or the large models disrupt it. Animation is hot lately, with what feels like twenty companies — how am I supposed to judge? Things with industry barriers or industry interfaces, like education, recruiting, healthcare, finance and insurance, are harder for big tech to enter. For other consumer companies, tagging along at the C round is safer for me; leading an A round is instant cannon fodder.

Liang Jie: Of course if we were in Sequoia’s position we’d have to position discretely across all of it; potentially it’s a big direction.

There’s a third direction to add: the first is AI’s capability, the second is AI remaking the old world, and the third is the new world AI creates — we call it the agent economy. In future everyone will have many general-purpose agents and vertical agents, and these agents will interact, with protocols and connections between them. And after they interact, is there some new thing on top? Don’t know, but there’s a lot of room to imagine. I think there will be, and we may do some exploring.

3. “Big-tech alumni pitching a story — we basically don’t touch that”

Raymond: This doesn’t sound especially different from most AI VCs today. Liu’s piece mentions you want to sit 20 degrees off industry consensus. Some background: on Zhang Xiaojun’s podcast, Zhu Xiaohu said he wants to sit 15 degrees off consensus. So three questions: what is the industry consensus; what is 15 degrees off; and what is your 20 degrees off?

Liang Jie: Start with the state of play. Leave aside the RMB-leaning funds, whose weight is in chips and embodied-intelligence robots — we certainly don’t participate there. Among the pure-dollar styles, the most heavily invested categories are a few.

One is things that scale extremely fast: video, animation, fast-exploding, with active investing and relatively large amounts.

Another is a fairly young clever person out of a big tech company pitching a story. Whatever it is, the first round has to be invested at some price; the product isn’t out yet, plenty of people missed the second round so there’s another; still people who missed it, so a third. That category, we basically don’t touch.

The third is so-called AI hardware, which we also looked at actively, a whole sweep of it. Why do I say 20 degrees? I read Zhu Xiaohu’s interview and resonated with the vast majority of it, but he also invested in a piece of hardware. All I can say is we don’t understand it: a little toy-like thing hanging on your body that keeps you company — that kind of thing, we don’t understand.

Raymond: So the current VC consensus is roughly video and animation, the things that scale easily on Douyin and can spike in a short window.

Liang Jie: The direction and the demand are at least right; the problem is competition — maybe one in ten survives, or big tech comes in and does it. And then sustainability: if model capability covers it, does the thing simply cease to exist? But the demand is real, so we look. Just cautiously.

Raymond: And the big-tech-talent-monetization category, the pure story pitches?

Liang Jie: Certainly don’t touch it.

Raymond: I hadn’t thought about this before. A lot of VCs now camp outside the big brands — I won’t name them — waiting outside ByteDance, DJI, maybe DAMO Academy to see who walks out, and whoever walks out gets money. It has the feeling of picking kids up from school. I completely agree: coming out of a big tech company doesn’t imply success. These people are better at talking and have a better shot at the next round, and for a VC they certainly last longer, but whether they ultimately build anything —

My own view is somewhat non-consensus too: lately I like companies where Chinese people have an advantage, but not where every Chinese person has an advantage. EVs used to be a place Chinese people had an advantage, and then it became a place every Chinese person had an advantage — the whole supply chain got built, and anyone at all can build a car. A European LP told me you Chinese are so formidable at EVs, and I said go look at the share prices, they’re all at all-time lows.

Fake-demand AI hardware for companionship will die fast; no matter. But what about genuine demand? Some companies selling well globally now will also hit fierce competition later — like robots at CES, twenty brands and only two suppliers, both in Dongguan. Not absolute, of course; power banks have no technical barrier and Anker still broke out. I just think the probability is low, and I have no ability to tell who the next Anker is.

Liang Jie: When I was running Skyline, 36Kr asked me at CES what I liked for going global next, and my view was that hardware is China’s advantage. I still see it that way today; it’s just that so far few have hit genuine demand. But I’m still hopeful.

What worries me is that everyone was educated by the mobile internet as a whole, so ways of thinking are more or less similar. Call back to Pinduoduo — in 2015 many people judged that e-commerce was finished, that to do it you needed a particular hook, that Dewu or Weipaitang types still had a chance and nothing else did. And then Pinduoduo tore an opening in a way nobody at all had imagined. That was an extremely, extremely non-consensus thing. I don’t know what today’s version of that Pinduoduo looks like.

Raymond: Doing VC for a dozen years, you inevitably form your own framework. A framework helps, of course, but as you say, when you meet something genuinely paradigm-breaking, too deep a framework blocks the new thing instead.

Liang Jie: And a lot of it is luck, whether you happen to run into it. If I hadn’t met that person, I’d also have concluded e-commerce had no opportunity left. But you meet a person whose work and angle of thought crack the framework open all at once — if you never meet a person like that, you may stay inside that framework forever. If Colin Huang hadn’t done this, I think most likely it wouldn’t exist.

Raymond: How many founders do you meet a year now?

Liang Jie: Three to five deals a week, around three hundred a year.

Raymond: And the kind that makes you think about it before you sleep, that this person seems to have a point, seems not to be entirely a fraud — what’s the rate?

Liang Jie: Very few. On average one a month, maybe fewer.

4. “You couldn’t even break through in Canada — what are you doing?”

Raymond: You invested in an Australian company, so you look at people globally?

Liang Jie: Yes. This AI wave’s global opportunity is of course enormous. But the core American white founders focusing on the US market — that core battlefield we certainly can’t enter, and it’s very difficult for Sequoia China too. Predominantly ethnic-Chinese teams doing AI all over the world, whether remaking the old world or creating a new one, is where we have some opportunity. What we can build connections with is Chinese people: whether you’re in Australia, Singapore or the US, if you’re Chinese, I may have a slight edge over the big and small American firms — Chinese people have certain connections, or certain qualities that top-tier American firms have started to look down on, that we might recognize and see as a dark horse.

Raymond: Is that a bit like ZhenFund’s style? Investing in outstanding Chinese founders worldwide.

Liang Jie: ZhenFund’s framing probably also started after the AI era. Previously it liked backing returnees; later a lot of AI founders did a PhD at Stanford and started a company in Silicon Valley right out the door, so ZhenFund simply invested in Silicon Valley. ZhenFund really has done very well on this.

Raymond: As with that Australian company you mentioned, it serves every hospital and clinic in Australia, not one clinic run by a Chinese person in Australia, so the ceiling is naturally large.

Liang Jie: Whether the team is Chinese doesn’t matter. Conversely, you mentioned food delivery companies like Foodpanda — Chinese people doing it is fine, and breaking through in the US is fine. But some of the food delivery companies Chinese people ran overseas back then — why did I take one look and not want to look again? It’s obvious: a company that can break through in one place may have a chance; if you can’t even break through in a small place, there’s nothing to look at.

Raymond: So you’d looked at people doing food delivery in Canada?

Liang Jie: Yes, HungryPanda, Fantuan, we looked at all of them then. Those companies’ ceilings are at most one or two billion dollars; there’s no possibility of ten billion. You couldn’t even break through in Canada — what are you doing? You can say I can’t beat the US, but I dominate Canada, and DoorDash and Uber Eats can’t beat me even once — that’s hopeful. If you can’t manage that, the company has no value and the market has no value either, because Uber Eats and DoorDash may not even bother entering that market.

Raymond: Precisely because nobody bothers pinning you down. It’s a scrap market. So do you think AI today is much better than that?

Liang Jie: AI is first of all certainly a global play. But coming back to specific industry settings, there’s still some local character — at least the deployment part is a bit like ride-hailing and food delivery.

Raymond: Vertical industries have more defined settings.

Liang Jie: The Australian company we invested in feels a bit like ride-hailing and food delivery: there’s BD onboarding, there’s field sales, I have to tell a clinic what your workflow is and then bring them in. Overall it’s still product-driven, but at the end you have to adjust which modules the product needs, and once the framework is roughly right it can roll out across Australia. The US is broadly similar, the medical system differs, adjust a bit.

Raymond: On healthcare, in the US it’s now just a few companies: OpenEvidence, Abridge. My sense is twenty or thirty percent share of all major hospitals has already been taken. Extraordinary.

Liang Jie: OpenEvidence is already over 40%.

Raymond: 40%? That’s extraordinary penetration. And a division of labor has emerged: some do case notes, some like your company do admin, some do drugs and insurance. This is really an enormous revolution against SaaS. No wonder SaaS stocks fell so much.

Liang Jie: OpenEvidence penetrates much faster than we do, because it’s a pure knowledge product. A doctor sees a case in consultation and needs to look up literature. They built a medical model and spent a lot of money buying literature from day one, so every answer has a source — large models often hallucinate, and it cites which paper. Add genuine medical experts training the model, accumulating continuously, and the knowledge base is authoritative. So it has no field sales and no onboarding and grows very fast; ours is slow.

Raymond: It’s more like one person selling Wikipedia, no switching cost, so it’s fast.

5. “We enormously underestimated Anthropic; MCP is Android-scale”

Raymond: Over the past year, was there a company you encountered and badly wanted to invest in and didn’t get? Any one-that-got-away moment in 2025?

Liang Jie: Ones we saw and had the chance to invest in and didn’t — probably not many; the overwhelming majority we simply didn’t see. But I think we certainly underestimated one company enormously, which is Anthropic. Enormously, enormously underestimated. And you only see it when the year-end results come in: it turns out this company’s position in the industry, its growth rate, its position, may all be that of a platform-level company. At the start of the year the feeling wasn’t nearly that clear.

Raymond: What’s your understanding of Anthropic today?

Liang Jie: Probably what we’d call an OS for work, with a lot of room to imagine. Does that mean killing Microsoft? Don’t know. But do you still need Word and Excel today? In future you may not.

The other thing: the MCP protocol is extremely impressive, and people may be enormously underestimating its value. It defined the whole industry, and you can already see it becoming the industry standard. The value of that may be very, very large.

Raymond: MCP is Android-scale, surely, if you stretch the time horizon.

Liang Jie: I see it that way too. So at the start of 2025 I certainly enormously underestimated this company.

Raymond: And at that scale, its revenue growth exceeds 200%, faster than OpenAI’s.

Liang Jie: Revenue growth I’m not worried about; whether it’s 100 or 200 doesn’t really matter. Because it’s clear: Gemini, OpenAI and Anthropic together account for a negligible, invisible share of the whole US knowledge-economy GDP. So they’re all very large companies.

Raymond: I underestimated Anthropic somewhat too, but not as much as you. At the start of the year I’d have said OpenAI was the dominant player, seventy-thirty against Anthropic, or even eighty-twenty; by late last year it was clearly sixty-forty; lately I even feel it’s fifty-fifty, or the reverse. The core is still that it caught whether to go B2B or B2C. Claude Code I think is it picking someone else’s fruit — someone else used it to build the prototype, and it then took Cursor out. As of today, OpenAI is no longer the one holding everyone else down.

Liang Jie: Possibly two things. One is that OpenAI’s focus is consumer, and long term the consumer ceiling is of course larger, but it needs time; today the efficiency gains in B2B workflow are fairly direct, so those happened first, and consumer will depend on richer interaction later, a picture we can’t see yet. And part of it relates to Google — Google is back, and a consumer market that was one-company dominance has become seventy-thirty or sixty-forty.

Raymond: They’re getting more and more dangerous. We’ll see; we’ll see whether OpenAI lists.

6. “Someone who has seen the most extraordinary companies simply has different taste”

Raymond: Last time we discussed a ranking I’ve been studying: the 2015 Chinese unicorn list ITJuzi compiled, which I’ll put in the show notes. You look at history hoping to draw lessons from it. We didn’t live through the PC internet, but we at least lived through the mobile internet and now this AI wave, and the experience of the previous revolution ought to help with the next. Most of the founders on that list you’ve met, talked to, wanted to invest in and didn’t. Reviewing it, what’s your feeling?

Liang Jie: The feeling from reviewing is: when you’re inside it, you don’t know. You only know after you step out.

Why does a VC still need experience? Just as AI needs a lot of data to train on, and training more makes it smarter, a VC likewise needs data, and the overwhelming majority of it is failure data. Is it useful? Of course — you learn that all of these failed. But what’s scarcer is always successful data, and the most extraordinary data is the scarcest of all. Which is exactly why someone who has seen the most extraordinary companies simply has different taste. Because they lived through it. Not to advertise for a former employer, but one good thing about Sequoia is that you saw the best things in all of China, and in the later stretch of the mobile internet, in the world.

Raymond: The internet’s center in the first wave was certainly America, from Yahoo to Google, and from the late nineties to around 2000 China was following. But in the later stage of the mobile internet, I think it was China, not America.

Liang Jie: Yes. Over the mobile internet’s decade, the Uber model and the DoorDash model did come first, but the ones that ultimately ate the large-country markets were these companies. The most native and most representative: the first product is WeChat, certainly the most impressive product in the world; the other two are ByteDance and Pinduoduo. One short video, one e-commerce on top of WeChat — that degree of flourishing isn’t something the American toolkit could have grown. Food delivery was started by the Americans, but Meituan’s organizational structure, infrastructure and understanding of the business model today, I think, exceed DoorDash’s. Taobao and the e-commerce companies were also started earlier by Americans, but the overall operating level is far better than the Americans’.

Raymond: Even earlier it was actually GrubHub. When we were doing banking work for Meituan we had to benchmark it against an American company — American investors were then used to the “you’re the Chinese Uber, the Chinese eBay” framing, but GrubHub was one percent of Meituan; it simply couldn’t be benchmarked. Including the whole recommendation-algorithm layer, which America still doesn’t understand today. There’s a generational gap in there. So you witnessed the founders of ByteDance, Pinduoduo, Meituan —

Liang Jie: Some at closer range, some further. Further means the firm invested and you hear about that founder’s changes every day, hear colleagues discuss what problems he’s hit, from an initial few hundred million dollars to a billion to ten billion, and finally the merger with Dianping and what to do about Ele.me. Having been through that whole sequence, your taste inevitably changes.

Including what the market calls VC 2.0 and VC 3.0: in 2.0, Gaorong and Source Code are extremely impressive; in 3.0, say Monolith with Cao Xi as the representative, or Jinqiu as the most active early-stage player, or the ones doing secondaries and primaries at later stages and taking allocation in top American companies — all of them came off that platform, and I think there’s a reason: that platform had a great many people who had seen good things. Cao Xi once said, when you see lightning, you know it’s lightning.

Raymond: Honestly, when I look at that 2015 list, I keep thinking that I still can’t make a clear judgement on them. I think repeatedly about ByteDance: when would I have invested? At any given point, basically not in any round. The only possibility would have been during the pandemic — the company had fully grown but the valuation was discounted, in 2022, with a market valuation available at two hundred billion. Zhang Yiming is certainly an excellent founder, but he isn’t the kind of person I’d have recognized at a glance.

Conversely, SIG discovered the usefulness of recommendation algorithms in America and elsewhere. Their thinking was: I know recommendation algorithms are a good technology, I don’t know who will use it, and then they happened to meet this person, this person is Zhang Yiming, he’ll use it, so go use it, and I’ll take a shot. So they went from the technology to the person, which is very different. In my period I didn’t have that good an understanding of recommendation algorithms, so at no point would I likely have invested in Zhang Yiming. And plenty of companies in between had founders I thought were excellent who just happened to be in the wrong sectors — used cars, education, P2P.

So my summary of that list is two things. First, strip out the personal relationships as far as possible and look at what the industry’s underlying technology is, what delta variable that substrate has revealed, and which person — surnamed Zhang or Li, doesn’t matter — will use that variable to build something. Second, the sector matters vastly, vastly more than the company. Consumer internet and SaaS are entirely different lives. Look at Bai Ya doing Youzan all these years; China’s soil is different too.

7. “ChatGPT is roughly the iPhone 4”

Raymond: Last time we also discussed what concrete historical moment AI is in today. Are we in some year of the iPhone, or some year of the PC? The year the BlackBerry appeared, or 2007, 2008, 2009, 2010 of the mobile internet? Working backward like that, you can also derive the social form of the time and what investing should chase. So two questions: do you think we’re more like the PC internet or the mobile internet — PC was obviously bigger than mobile; and second, which year of that respective cycle are we in, and what does that suggest?

Liang Jie: On the first, I lean toward ChatGPT’s arrival being like the iPhone 4. AI is first of all a technology, and there had already been many applications, but when ChatGPT arrived everyone felt an aha moment; it was fairly disruptive for people. On the surface it looks like a better search, but I think it’s nothing of the kind — it’s a new brain, and it can drive everything that comes after, the agents we talk about. So ChatGPT is a bit like the moment the iPhone 4 arrived: the iPhone may already have existed, but early on it wasn’t that good, like the earlier models that hadn’t yet reached the point of making you go aha.

Then this chat box may be something like WeChat, a fairly large origin point. In mobile, WeChat did an enormous number of things for the ecosystem; whether there’ll be a ByteDance and a Pinduoduo on top of this thing, I of course hope so, but today certainly isn’t that moment yet. Pinduoduo was founded in 2015 and ByteDance in 2012, but short video is what actually took it to a hundred billion dollars — Toutiao may have been a several-tens-of-billions thing.

Raymond: So mapping onto the mobile internet, we’re roughly between 2010 and 2015? If we were a time machine, what should we invest in today? Ride-hailing and food delivery?

Liang Jie: Ride-hailing and food delivery, which is to say these verticals. The capability exists; you take that capability and do all sorts of things with it. In the mobile internet everyone went mobile and had a smartphone, so you needed a lot of services; today the brain exists, GPT or Gemini, and I take it and deploy it into every industry.

Raymond: And what’s the ByteDance and Pinduoduo — the Douyin-and-Pinduoduo position?

Liang Jie: That I don’t know. Maybe something related to agents. Everyone uses a great many agents, these agents can do some quite magical things for us, and then some interaction is produced.

Raymond: Back to verticals. They’re relatively easier to get running, but which one exactly —

Liang Jie: They have relatively more certainty, but the room to imagine isn’t that vast. They’re certainly not the next Anthropic; not that magnitude. Is there something at that magnitude? I very much hope so. And it’s possible that thing has already appeared — for instance it may be Anthropic itself. Ten years from now we find the biggest riser was Anthropic; entirely possible, it’s underestimated.

8. “I thought the match was over; it turned out to be half time”

Raymond: Suppose we’re in an elevator and you have two minutes to know me, and I’m an AI company. What would you ask to figure out whether this person is investable?

Liang Jie: First, what problem does it solve, and what role does AI play in that process — why couldn’t it be done before, and why can it be done now that there’s AI. Also how big that problem is, and whether it’s sustained: as things evolve, will its advantage as a participant in this industry amplify or shrink, or will the industry ultimately fragment?

Raymond: That’s a question of dynamic competition. So do you prefer to invest in companies in more winner-take-all industries, or more polymorphic ones? Communication tools are extremely winner-take-all; banks, P2P and games are not. In a winner-take-all industry, backing the number one makes you rich; investing in a polymorphic industry, you don’t easily die. Two different ways of investing.

Liang Jie: My preference is certainly the first.

Raymond: Definitely the first, I understand. So that’s really the five degrees between you and Zhu Xiaohu — he’s 15 off consensus, you’re 20.

Liang Jie: But that’s only preference; actual action is different. When I was running Skyline doing going-global, as an investor I couldn’t see ten-billion-dollar opportunities — so do I just not invest? I have to participate too, and the companies I invest in then have less upside. This is my profession; I can only pick from what I can currently see that’s still reasonably good. I can’t say there’s no technological revolution right now so I’ll retire; I still have to invest. With no platform opportunity, do you invest in mid-size branded opportunities? We invest in some. You know these companies certainly won’t lose money, but how much they’ll make is unknown. From the standpoint of being obliged to show up for work, you can still invest a bit — at minimum it keeps you connected to the industry, so that when the real platform opportunity comes, you’re in a position to get it.

Raymond: That’s what’s hard about private-market investing. We recently wrote a piece called “Guarding Common Sense”: why common sense is actually hard.

Liang Jie: Late 2021, early 2022, no new technological revolution, the setting sun of the mobile internet — did we have that judgement? I did. So do I not participate in the setting sun? If you don’t participate, that’s saying goodbye to this and going off to do something else.

Without ChatGPT, without AI, this industry might have disappeared. If China’s dollar VC capital base at the mobile internet’s peak was 100, it later fell to roughly 10; without this AI wave it would now be down to 5. It wouldn’t be zero, but it’d be 5. And at 5, 95% of people should leave the field. Having done it for a while and felt I couldn’t participate, I’d actually already thought about it; I might have left. But now, because of AI, it’s back to 20 or 30. Which is why I say I’m lucky — I thought it was only 5, and it’s 30, so I came back.

Raymond: Let’s talk about life then, two old farts talking about life. At every moment a person is investing their own time; every day is my youth. Every career or life choice I make, I try to optimize: whichever industry, place or thing I think has the highest investing efficiency today, that’s where I go. So I don’t think I’ll retire — but I also don’t mind leaving the field. Suppose I ran a dollar VC, it fell to freezing and GPT never appeared; then I’d switch tracks. I might go open, I don’t know, a hedge fund; I might go open a coffee chain; I might do anything else — anything with an increment for my intellect and possibly for my wealth. Just like that.

Liang Jie: What you’re describing is really a question of attitude toward life, and on that I’m the same. So what’s the difference? The difference is that I’m a VC practitioner by nature, and I love this. I have something unresolved, because I don’t feel I’ve caught a truly extraordinary company or built something very large. Today at least gives me the chance to be back on the field — I’ll do this my whole life. I thought the match was already over. Now they’re saying that was the end of the first half, it’s half time, and we come back for the second.

9. “What ultimately decides whether you land Pinduoduo or a knockoff is emotion”

Raymond: Why do you like the VC industry? If you deconstruct it, what’s the most attractive part?

Liang Jie: First, certainly, through this industry you get to participate in some of the world’s most important technological changes and are constantly predicting which direction the world goes. In that process you keep raising your understanding and learning from the smartest people, which is interesting in itself.

The second is that in that process you’re a fairly important participant. You can of course buy stock, but that return is more on paper. Some people with intellectual vanity say I buy stocks to validate whether my hypothesis is right — that part exists. But the sense of participation in VC is more important: I think this company is good, I talk with the founder, drink with him, accompany him through fundraising. That sense of participation is different from buying stock, where all you need is a strategy. If you don’t enjoy that thing, then go do public markets; you don’t need to be in private.

I introduced a very successful public-markets friend to my two partners, and the two of them had entirely opposite preferences: one said anything that makes money interests me; the other said I’m not interested — simply making money no longer drives him.

Raymond: So the key is what drives you. One is chasing technological waves, the other is that you want the feeling of being in the room, witnessing history firsthand.

Liang Jie: Right, it’s a kind of vividness, or sense of story. You buy Nvidia stock and make a lifetime’s money, but you know Nvidia has nothing to do with you, and Jensen Huang has nothing to do with you. Whereas the cases you participate in may be only one or two Series B companies, but it’s zero to one, and at certain key moments you went through things alongside the founder, and then you produced that process. The satisfaction that experience gives you is more interesting than merely the size of the number.

Raymond: It’s like how I could buy SpaceX stock and make ten or twenty times, but I could never go to a small island in the Pacific and launch rockets with them, have the rocket come down, and go months without a shower. So a final question: with AI here, if you deconstruct VC, what can’t be replaced by AI? What means Claude can never generate a Liang Jie, and Cowork can never have a Liang Jie to look at deals?

Let me give my observation first. There’s a blog arguing that the difference between people and machines is really the ordering of values. There’s a man with a brain injury who can do all the reasoning — one plus one is two, no problem — but ask him to choose one of two pairs of socks to wear and he can’t. He wants to add that value-ordering part to AGI. And there’s a WeChat piece arguing: AI does company research extremely well, but at the end, when you tell it to pull the trigger and say which stock to buy today, it’s often uncertain — because it can collect data for both the bull and the bear side. So a lot of the time, on that final step, on the 49-versus-51 risk judgement, a person may have a little bit extra.

Liang Jie: As a tool, AI is of course extremely efficient, and from a rational-analysis standpoint it can already do most of it, even exceed us.

The first is connection between people. It’s a machine; it has no connection with people. In investing today the principal is certainly still a person, and in the foreseeable future you won’t hand it to a machine; that sense of trust isn’t there.

But extending the topic: connection between people must exist because we’re the same species. Which is the preference you mentioned. We meet many founders, everyone at the same level, and from a rational standpoint these people are all the same, but I particularly like a certain one — that’s the difference. Some motivations are personal preference; some are a sense of the big picture beyond reason. On pure calculation everyone is the same, but once you’ve seen many large things, large companies, say hundreds-of-billions-dollar companies, and the flavor of those founders, you get a new and different understanding of the big picture. At that point, relying only on rational calculation, these few people look identical; combined with that sense of the big picture, the judgement differs.

Raymond: That’s still on the rational side of ordering. This kind of prioritization does look hard for AI right now — it doesn’t much have that ability to step outside. It’s a matter of taste.

Liang Jie: Taste may just be: its data contains both sides, and when it comes to the final ordering, some things you can’t quite articulate make the ordering slightly different. Call it a sense of the big picture, call it an ordering of values; in the end it may be a person’s emotional inclination.

Raymond: And that emotion may ultimately decide whether you land Pinduoduo, or a knockoff called Pin Shao Shao.

Many thanks for Liang Jie’s time today. VC is an industry with a particular preference for risk, and Liang Jie is someone who strikes me as having his own thinking while also giving me a strong sense of trust. Some people are more easily replaced by AI, because there isn’t much human flavor about them. Liang Jie may be among the most human-flavored people I’ve met. So if we all eventually get replaced by AI, my feeling is Liang Jie is somewhat later in the queue.

Liang Jie: That’s high praise. I hope I’m not replaced by AI in the short term either. I think in the end VC is a business about people, and in the end it all comes down to everyone’s taste, risk appetite and sense of the big picture — those preferences may decide whether we can find the next Pinduoduo.

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.