2026-11-17 16:06
Introduction: GPT-6 as the Kill Line, Valuation Logic Reverts to Manufacturing. From one individual’s choice, glimpse an industry’s transformation. Tencent Technology’s in-depth dialogue series 「Lingting theSignal」 aims to reveal the strategic thinking behind pivotal commercial decisions in the hard tech wave. This is episode 08 of the series, featuring Wang Sheng, Partner at Inno Innovation Capital and Chairman of the Beijing Frontier Institute for International AI.
Over the past three years, Wang Sheng has invested in over 50 companies across the embodied intelligence supply chain. In June this year, he decided to hit the brakes.
At that time, embodied intelligence was still one of the hottest sectors. The Inno Innovation Capital fund had ample capital, yet new deal submissions to the investment committee had dropped from several per week to just one every few weeks. He likened investing to a high-speed train: even after applying the brakes, it continues gliding for a long distance.
The primary consensus in the primary market gradually emerged: institutions still hold cash, but are no longer eager to deploy it.
Wang Sheng’s decision to slow down had roots in his early experience. At the end of 2021 and beginning of 2022, he was searching for a promising large model company. He noticed a signal from technical shifts within already invested firms: more teams were adopting Transformer architectures, models grew increasingly larger, and moved from recognition and judgment toward generation. He reached out to Qi Fan-chao from Tsinghua University, persuading him to launch a venture. DeepTalk was founded in June 2022—nearly half a year before ChatGPT’s release.
After Inno’s investment, Qi Fan-chao met with 50 institutional investors. Few knew what Transformer or GPT even meant. Eventually, Sequoia Capital made the investment. Following ChatGPT’s launch, DeepTalk’s valuation surged 20-fold.
DeepTalk had first-mover advantage but didn’t secure sufficient time window. After ChatGPT, top talent flooded into the sector. Companies led by Wang Huiwen, Wang Xiaochuan, and Yang Zhilin launched with initial valuations in the hundreds of millions. A newly graduated PhD team now faced direct competition in funding, talent, and industrial resources.
Wang Sheng correctly identified the direction but missed the biggest players in this wave of large models. Around the same time, he began building positions in embodied intelligence. By late 2022, he predicted that large language models would inevitably enter the physical world, and 2024 might mark the dawn of embodied intelligence. If correct, 2023 should have been the right time to invest in robot hardware. He backed Accelerated Evolution and Songyan Dynamics, then continued downstream along the supply chain.
Correctly identifying the trend doesn’t guarantee backing those who ultimately succeed. Two teams he admired initially aimed for autonomous driving. Wang Sheng judged this path too saturated for startups and advised them to pivot to robotics. They declined. He did not invest. Half a year later, these teams transitioned into embodied intelligence and quickly became billion-dollar companies. Once again, he foresaw the trend, but underestimated their speed in course correction.
He later summarized this experience in one sentence: “People are paramount—no emphasis is too strong.” Even while consistently advocating early-stage investing as ‘investing in people,’ he still felt he hadn’t placed enough weight on human capital.
Wang Sheng is a Partner at Inno Innovation Capital. A rock enthusiast in college, he entered the internet industry post-graduation and became an early employee at Sohu. He has experienced multiple entrepreneurial ventures and sold two companies. Approaching 40, he shifted to early-stage investing.
From entrepreneur to investor, Wang Sheng has witnessed the process of technology creating new markets. But this time, the relationship between technological advancement and entrepreneurial opportunity is less clear than before. Stronger models can help a startup build better products—but they may also directly supersede the core capabilities the startup depends on. For early-stage investors, predicting the right technology direction is merely the starting point of due diligence.
Recently, I had an extended conversation with Wang Sheng. Our discussion began with one question: As AI grows stronger, what remains constant—or changes—for startups and early-stage investors?
His assessment was sharp: “GPT-6 is a kill line.” Companies building robot brains will be eliminated if their robots perform worse than those powered by GPT-6. Yet, increased model capability does not automatically trigger a new wave of application prosperity. We also discussed why capital is concentrating faster than expected, why AI applications are a pseudo-proposition, and why higher robot capability doesn’t necessarily lead to higher company valuations.
We sought to understand how technological progress reshapes entrepreneurial space, and how an early-stage investor reevaluates judgment when the industry shifts from believing in the future to testing reality.
Below is my edited conversation with Wang Sheng.
Wang Sheng: Our current focus is primarily on helping our existing portfolio companies grow and scale. New project evaluations have drastically decreased. While we’re still scanning, very few make it to the investment committee. Previously, we’d discuss several deals per week, with one or two getting approved. Now, on average, there’s barely one per week—or even none for several weeks.
Gu Lingyu: What’s the most direct reason?
Wang Sheng: I believe it’s cyclical. This year’s capital markets have been extremely hot, with massive capital deployed. Some funds raised their first close and spent it all by mid-year.
First, investments were too aggressive; valuations rose too fast, creating evident market bubbles. Second, concerns intensified due to the downturn in secondary markets. Thus, hitting the brakes was inevitable.
Gu Lingyu: Has the primary market run out of ammunition?
Wang Sheng: Not at all. Many investors still hold substantial cash—possibly even more. We have ample firepower. But around June, we explicitly signaled a slowdown. By June, we sensed a turning point in the market.
It’s like a high-speed train: once moving fast, braking requires significant distance and time. By September, everyone realized the slowdown had already begun months prior.
My personal sense: actual investment volume in September was only 20%–30% of what it was during the peak heat.
Gu Lingyu: Was the cooling of embodied intelligence also something you anticipated? Did you foresee a downturn this year?
Wang Sheng: I had predicted it last year and early this year. It didn’t happen—so I must admit I was wrong. I believed the shift from storytelling to execution was inevitable, but the timing lagged. Perhaps because capital remained abundant, and strong performances by model and GPU companies in secondary markets led many to expect the next wave of IPO surges would come from robotics firms.
Gu Lingyu: You accurately foresaw the shift in this AI wave. What were your immediate impressions when engaging with companies?
Wang Sheng: In late 2021 and early 2022, we firmly believed in investing toward GPT-like trajectories. At that time, ChatGPT was nearly a year away. But we weren’t just reacting to its release—we had already seen earlier versions. Their capabilities were strong, but no single milestone was enough to justify full confidence.
Over the years, we invested in numerous AI firms. Since 2019 and 2020, we observed a consistent shift: training models increasingly adopted Transformer architectures. Across multiple directions, two traits emerged: AI evolved from discriminative to generative; models grew larger, data volumes increased, and quality improved. Seeing these signals, we knew the tipping point was near.
Gu Lingyu: After spotting the trend, what was your first move?
Wang Sheng: The first project aligned with the large model era was DeepTalk.
We wanted to back a team developing a large-scale, Transformer-based, autoregressive model. In 2021–2022, such startups were virtually nonexistent. So we sought a team to drive the venture. We found Qi Fan-chao, a PhD from Tsinghua’s Natural Language Processing Lab under Professor Sun Maosong, who had already trained several generations of models at the Beijing Academy of Artificial Intelligence. We encouraged him to start. After our investment, he approached 50 institutions—most didn’t know what Transformer, GPT, or LLM meant. Eventually, Sequoia Capital invested. Shortly after, ChatGPT launched.
ChatGPT’s release triggered frenzy among investors. DeepTalk soon raised two additional rounds. Such a dramatic transformation occurred.
Gu Lingyu: In AI, is first-mover advantage crucial?
Wang Sheng: Yes—but only if given sufficient time window. Today, the time window is critical in AI investing.
DeepTalk launched shortly before ChatGPT arrived—too short a window. Suddenly, everyone thought, “This is it!” Everyone rushed in. Top-tier talent began forming companies focused on large language foundation models. A freshly minted PhD team struggled to compete in fundraising, talent acquisition, and industrial resource access.
This domain demands massive capital. Investors prefer teams with strong academic influence or industry experience. Several companies formed in early 2023 raised enormous sums—not just due to founder pedigree, but also the powerful backing of investors and promoters.
Gu Lingyu: Did you engage with them?
Wang Sheng: Almost all of them.
Gu Lingyu: Why didn’t you invest?
Wang Sheng: We’ve internally reflected. First, we underestimated the concentration of capital and intensity of competition. We predicted the tech and industry would trend this way—but didn’t anticipate how compressed the timeline would become. We thought it would take four to five years, but it unfolded in one or two. Capital followed suit.
Second, Inno felt we held a strong card—our investment in DeepTalk—and weren’t desperate. Third, these startups’ first rounds were priced extremely high. For early-stage funds, typical pre-money valuations were tens of millions or hundreds of millions; seeing billions was unprecedented.
Gu Lingyu: Were these companies concentrated in early 2023?
Wang Sheng: Yes—January through May 2023.
Gu Lingyu: Why was the window so tight?
Wang Sheng: Because capital is finite—everyone poured in dozens of billions.
But I can tell you clearly: from late 2023 through 2024, and throughout 2025’s first half, investors who backed these model companies were deeply anxious and despondent. By late 2025, especially after Zhipu AI’s IPO, everyone suddenly felt brilliant: “I was right! I saw it coming!” They selectively forgot the intense anxiety of the previous year.
Gu Lingyu: What were you still looking at?
Wang Sheng: At the end of 2022, we forecasted two paths post-large language models: vertical industry applications and multimodal systems. But by mid-2023, we rejected both.
We realized both were areas where large language models would dominate—offering no fresh opportunities for startups. So we didn’t invest.
Additionally, we predicted 2024 would be the Year of Embodied Intelligence.
Gu Lingyu: Why could you predict this as early as 2022?
Wang Sheng: Large language models had already emerged, with strong capabilities, and would inevitably spill into other domains. Beyond verticals and multimodal, they would overflow into embodied intelligence. We estimated a reasonable timeline: wait another year—from late 2022 to early 2024.
If 2024 is the Year of Embodied Intelligence, 2023 should have been the time to invest in hardware. Hence, we backed Accelerated Evolution and Songyan Dynamics.
Gu Lingyu: Can the hardware gain traction before the brain is ready?
Wang Sheng: Exactly. The brain ultimately needs to control the body. Push the body first, then refine the brain.
Gu Lingyu: Have you missed any robot companies you regret?
Wang Sheng: Numerous.
Gu Lingyu: Pick one you regret most.
Wang Sheng: You want to name names?
Many founders are close friends. Guo Yandong of ZhiPing Fang, for instance—he came to us early. We pushed internally, but ultimately didn’t invest due to various reasons.
Also Hu Yuhang of Shouxing Tech. Before graduation, he returned to China, persistently working on robotics. He even joined me on visits to different enterprises to experience the industry firsthand. From his initial low valuation to several hundred million, I invited him to our company four times. I really wanted to invest—but didn’t.
Gu Lingyu: Have you ever misjudged?
Wang Sheng: Among the companies we’ve invested in, none stand out as clearly doomed—requiring urgent divestment. Market bubbles keep inflating. These investments are relatively recent—longest about three years—just differing in pace.
But I believe this is due to past heat and excessive speculation. People cared less about survival, more about storytelling and technical claims—without demanding profitability or revenue. Challenges lie ahead. Some companies will fail.
Gu Lingyu: Is early-stage investing fundamentally about investing in people? That’s your consistent principle?
Wang Sheng: People matter most—no exaggeration is too strong. We’ve invested in teams before companies even existed, let alone products. If we admire the team—stable, complementary skills—we invest.
Even so, I still think we haven’t valued people enough.
Gu Lingyu: How do you define “not valuing people enough”?
Wang Sheng: A classic case: a founder is brilliant and likable, but pitches a flawed story and wrong business direction. We see no value. He must pivot—either significantly lower valuation, which we might accept; or we pass.
Other investors may lack ability to judge direction. They invest anyway, pushing up valuation. The founder quickly realizes the mistake, pivots, but by then we find the price too high—past our investment threshold.
Two founders I trust deeply contacted me repeatedly in early 2023–2024, seeking investment for a new autonomous driving model.
Gu Lingyu: At that time, autonomous driving was already saturated.
Wang Sheng: Correct. But they saw potential in end-to-end systems—shifting from rule-driven to vision-input/action-output. That’s a major leap. But the industry landscape differs: you’re not Tesla. Replicating that won’t work. We refused investment or demanded steep valuation cuts. At the time, valuations were only several hundred million. Today, both companies are worth 20 billion.
Gu Lingyu: Did they pivot to embodied intelligence?
Wang Sheng: Within half a year. We told them: “This won’t work. Focus on robotics.” They replied: “No, we’ll stick with this”—as they were all from autonomous driving backgrounds.
We believed their initial direction was wrong—but these individuals are exceptionally smart. After floundering briefly, they quickly pivoted, adjusted swiftly, and gained rapid attention.
Gu Lingyu: From the internet era to today, what’s your consistent standard for judging a great company?
Wang Sheng: Simplest criterion: I like the founder.
Many reasons. His words align with mine—though naturally, we all believe our own thoughts are correct. If his thinking matches or even surpasses mine, he’s worth backing.
Also, we need intellectual rapport. Investing is a marathon—we must stay together. Likeability starts with shared values, technical depth, professional competence, and vision. Also, we must resonate with his character, integrity, values, and work ethic.
Gu Lingyu: You’ve discussed many foundational models. What applications have you targeted?
Wang Sheng: Not exactly a strategy. We’ve invested in a few applications. From late 2024 to 2025, applications were hot—many top funds poured in. But almost all suffered total losses. Today, results are nearly catastrophic.
Gu Lingyu: Is Agent considered an application?
Wang Sheng: If forced to classify, yes—but we usually treat Agents separately.
Gu Lingyu: What exactly do you mean by “applications”?
Wang Sheng: More along mobile internet lines: building an App, product, or tool enhanced by AI. Agents, we view fundamentally as humans: specialized in a domain. You instruct them, treating them like people.
Gu Lingyu: As models evolve, will application space shrink?
Wang Sheng: Extremely. The unpredictability of this AI boom is overwhelming—its foundational capabilities remain uncharted.
In the mobile internet era, e-commerce, ride-hailing, O2O apps were viable. People debated only penetration rates: 80% in ten years vs. three. Market size was calculable, relatively clear. Today, we don’t know what AI can do. Existing tools may be rendered obsolete.
Not zero market—just shrinking rapidly.
Gu Lingyu: Do you think model capability growth is slowing?
Wang Sheng: On the contrary—it’s accelerating. No one knows the true limits of iteration or boundaries.
Gu Lingyu: Capability limits depend on compute. If costs spiral, big players can’t sustain infinite investment.
Wang Sheng: That’s precisely the current dilemma. Buying chips, building data centers—even power plants—bubbles continue to inflate. When it bursts, we don’t know.
Returning to applications: second reason. Apps in software, internet, and mobile internet eras emerged due to expanded device scenarios.
Computers in offices enabled productivity software: design, architecture, databases, Office suites. PC era: home internet use for information, shopping, entertainment. Mobile era: always-on devices—hailing rides, taking photos, ordering meals.
Apps require context. Solving problems depends on the scenario. New devices expand contexts. Today, homes have PCs, backpacks carry tablets, pockets hold phones. No new contexts created—why new apps?
Gu Lingyu: So you consider AI applications a pseudo-proposition?
Wang Sheng: Pseudo-proposition. Only Agents exist in AI.
Gu Lingyu: Do Agents offer startup opportunities?
Wang Sheng: Very limited. But non-generalized Agents are possible. I believe generalized Agents are fundamentally a big player’s domain.
Gu Lingyu: Where are the best opportunities for non-generalized Agents?
Wang Sheng: Sales, for example. Imagine a salesperson using a corporate version of Plaud, recording client negotiations, integrating enterprise workflows and knowledge bases, helping develop clients, predict purchase likelihood, and provide prompts. A few companies have already achieved success.
Gu Lingyu: Is this opportunity driven by data?
Wang Sheng: You may have access to abundant data and information, but can you fully digest it? With so much input, how do you extract insights and write reports? Human learning and work methods vary—leading to inconsistencies. AI can continuously learn and improve strategies within this environment.
We’ve backed a project building social world models. When launching a new product, you conduct market research: Will users buy it? Which color do they prefer? Are features functional? How to price? If consumer behavior, habits, and values can be simulated by AI, it can act as an Agent, providing feedback on these questions.
Naturally, collecting high-quality data and constructing models representing individuals across multiple dimensions is challenging.
Gu Lingyu: Aren’t you worried that the next model upgrade will overwrite these capabilities?
Wang Sheng: Not at all. Foundation models can be powerful, but data constraints limit them. High-quality, hard-to-obtain data accumulation and closed-loop operations within business are essential.
Using existing data, real-world user participation for validation, gathering high-quality feedback, refining models, and revalidating—once this flywheel spins, you cannot beat it.
Gu Lingyu: As models grow stronger, do founders face more or fewer opportunities?
Wang Sheng: This wave filters out only a tiny fraction. In mobile internet and consumer entrepreneurship, though success rates are low, people still feel they can try. A product manager rents cloud services, builds a website, develops an App, submits to app stores.
This wave demands vast capital and exceptional talent. Chip development, inference, model training, robotics—all highly demanding. People assumed the application era was their moment, but applications prove difficult.
Perhaps two things are needed. First, new devices create new scenarios, spawning new apps. E.g., AI glasses, AR glasses—perhaps entrepreneurs can find opportunities here.
Second, I’ve always believed that productivity-oriented solutions are hard for average app founders to crack. AI’s coding prowess makes replication easy. Big players have traffic and users. True value creation may return to consumer-facing offerings with unique emotional value.
Make films: your creativity may surpass Hollywood, production cost drops dramatically. Make games: leverage AI for low cost, high efficiency, appealing themes and mechanics.
This is my view—maybe wrong. I genuinely believe AI application startups only thrive when delivering unique emotional value.
Gu Lingyu: Is this why AI investing has cooled?
Wang Sheng: More accurately, AI applications have cooled sharply. It’s not solely due to capital cycles.
Gu Lingyu: Without mentioning embodied intelligence, what good startup opportunities remain in AI?
Wang Sheng: Strictly speaking, we’re nearing passive mode. Most energy goes toward supporting our existing portfolio.
Gu Lingyu: It’s a garbage time.
Wang Sheng: Exactly. Most top-tier VCs operate this way. As you said, perhaps until the end of this year or even spring next year, it will remain garbage time for AI investing—meaning not that no one invests or looks, but frequency drops to 1/10th. Only truly rare opportunities get funded.
Currently, two areas still attract interest: recursive self-improvement (RSI), and AI for Science—materials, healthcare, biotech, and other scientific research fields.
Gu Lingyu: Aside from AI programming, which paid scenario do you favor most?
Wang Sheng: Personal and small business use—content creation. Programming is content creation: writing code. Video, 3D modeling too. Our investment VAST sees rapid monthly subscription growth.
New growth drivers? Research. Scientists are high-IQ, scarce, expensive. Historically, insufficient numbers and funding. If AI fills gaps, immense value emerges.
Gu Lingyu: Could this spawn massive companies?
Wang Sheng: Absolutely.
Gu Lingyu: When?
Wang Sheng: Likely 3–5 years. Some optimistic types say 1–2 years. But from theoretical feasibility and algorithm research to training robust models, establishing data loops, finding effective methods—I believe it still takes this time.
Gu Lingyu: Do you still follow robotics?
Wang Sheng: We still watch—but won’t invest.
Gu Lingyu: Any standout performers recently?
Wang Sheng: If specifically referring to robotics and embodied intelligence, I’d say they’re all struggling. Not due to lack of capital or talent, but difficulty securing new funding rounds.
Gu Lingyu: Due to high valuations?
Wang Sheng: First, bubbles inflated too fast—valuations soared, making follow-on rounds difficult. Second, fundamental shifts in valuation frameworks and value assessment standards have occurred.
When an industry is immature and hasn’t created real value, people talk about future potential. Once value creation begins, people start calculating profits and problem-solving impact. If you can’t deliver—or need time—valuation drops.
Gu Lingyu: Why now?
Wang Sheng: Technology matures over time—must shift from narrative to tangible delivery. Refer to the Technology Readiness Curve: before the peak, trends, consensus, and expectations of future value sustain rising premiums.
Once consensus shifts, people assess actual profit potential and industry impact. If you can’t deliver—or need time—you fall. Then, if focused on product, tech, and business, revenues and profits grow steadily—healthy industries rise again.
I believe AI, including embodied intelligence and world models, is roughly at this juncture.
Gu Lingyu: At the peak?
Wang Sheng: At the peak of desire, preparing to fall.
Gu Lingyu: A good time to exit.
Wang Sheng: If you can exit, the earlier period was better.
Gu Lingyu: You sensed the downturn in June—was that a good exit moment?
Wang Sheng: Yes.
Gu Lingyu: From an industry perspective, recent discussions center on whether GPT-6 undermines independent robotics firms’ value. What’s your take?
Wang Sheng: Personally, it’s not weakening—it’s screening. We can treat GPT-6 as a kill line. If you claim to build robot brains or world models, but your robot performs worse than one powered by GPT-6, you risk being eliminated. Your efforts become meaningless. You must pivot—e.g., target specific scenarios, avoid general-purpose brains.
Any claim of generality must clear this line. It’s not impossible—time and data loops remain vital. To me, today’s GPT-6 has some spatial understanding, but lacks physical comprehension. Our world model investments emphasize physical and causal reasoning from day one. Of course, GPT-6 may incorporate physical data later.
Cost and latency matter too. Many robot scenarios demand rapid response—sometimes edge-side operation—requiring efficient understanding, planning, reaction, and execution. You can’t ask a robot to ponder for minutes.
Integration with hardware and depth of scenario deployment are critical. Foundation model companies can do this well—worth noting, but not surprising.
Gu Lingyu: Did you hold this view earlier?
Wang Sheng: Naturally. We always believed in leveraging large foundation models. If startups need pretraining capabilities, the strongest and largest are already established foundation model companies. We’d rather back people exiting those companies to do pretraining.
For foundation model companies, it’s about priority, investment intensity, and execution efficiency. But embodied intelligence is complex—entering factories, service hubs, with diverse hardware forms and connections. Far more complex and harder to deploy than software.
Is it inherently a big player’s domain? Not necessarily. Early experiments don’t imply immediate industrial readiness. But the kill line remains. Everyone races against it—becoming increasingly tense.
Gu Lingyu: Over the past half-year, attention shifted to brain-focused companies. Will this hype last?
Wang Sheng: If a company can surpass the kill line, it holds maximum value. But brain companies now face immense pressure. Smaller brain companies may feel relieved—big players likely won’t enter smaller brain domains tied to hardware anytime soon.
A reversal may occur: people worry about brain companies, while small brain firms see room to excel.
Gu Lingyu: Is this a new opportunity for hardware companies?
Wang Sheng: Undoubtedly. Faster emergence of intelligence means robots enter valuable scenarios quicker—increasing demand for hardware.
But valuation logic has changed. Previously, it was seen as frontier tech—insurmountable barriers—valued via “market cap of dreams,” 100x, 200x revenue multiples. Now, de-mystification sets in—robotics increasingly resembles manufacturing, driven by supply chains, scale, and capital expenditure. Hardware technical barriers aren’t as high.
Valuation models will inevitably drop—from hundreds of x revenue to 100x, 50x, 20x, even 5x. This seems inevitable in the coming years.
Gu Lingyu: Do brain companies use the same valuation model?
Wang Sheng: Not yet. Physical-world intelligence still has high barriers. Investors treat them as foundation model companies—requiring massive investment, with tech and data moats.
Gu Lingyu: What’s the most irreplaceable barrier?
Wang Sheng: First, people. Core team and founders making correct decisions and choices matter immensely. Second, capital. Data, compute, other talent—money can buy them.
Gu Lingyu: Why does the valuation logic still hold for brain companies? Couldn’t a combination of foundation model and hardware companies replicate the experiment?
Wang Sheng: Current demonstrations aren’t yet usable in industry. Suppose OpenAI prioritizes this—dedicate heavy resources for the next two years. Maybe it delivers—replacing many current brain companies.
Gu Lingyu: Wouldn’t that reduce industry scarcity?
Wang Sheng: It would generate uncertainty—and intensify capital concentration at the top.
Gu Lingyu: Should embodied intelligence company valuations decline?
Wang Sheng: Personally, capital will concentrate at the top. Mid-tier companies lacking excellence will struggle to raise funds—high pressure. But one caveat: truly excellent teams still have chances.
For capable, diligent teams with decent funding, market cooling isn’t bad.
Gu Lingyu: Now focusing on delivery—how to measure?
Wang Sheng: Not necessarily—it depends on choice. Some companies are pragmatic.
For example, we backed Cloud Deep. We suggested building general-purpose robotic dogs or humanoid robots for education, research, exhibitions—relatively simple, scalable, high demand.
But Zhu Qiu-guo insisted otherwise. He wanted to pursue inspection tasks in harsh environments—firefighting, electrical switch handling. These are fragmented, complex—hardware and AI both intricate, labor-intensive. But you realize: this creates real value.
Also Su Hao: I aim to achieve 99% usability in one scenario before moving to another. Today’s AI remains a black box—many things can’t be proven. Pursuing generality may never reach 99%, let alone 90%. This reflects different strategic choices.
Gu Lingyu: Focusing on specific scenarios—could this delay foundation model development?
Wang Sheng: That’s his choice. I’m focused on solving real problems—making one or a few connected scenarios work.
Others choose to build foundation models, raise hundreds of millions, immediately begin pretraining—physical AI, world models.
Gu Lingyu: Will such companies become rarer?
Wang Sheng: Certainly. Capital concentrates at the top.
Gu Lingyu: The market will become increasingly brutal.
Wang Sheng: Indeed. The best fundraising era is over—writing ten-page pitch decks, assembling star teams, telling stories, raising three rounds in six months—gone.
Gu Lingyu: Will these institutions face elimination?
Wang Sheng: This has been happening for years. Survivors are those with strong combat power and resilience—capable of tackling complex challenges.
Gu Lingyu: For VCs, what matters most now?
Wang Sheng: Post-investment support: help companies scale, refine technology and products, manage fundraising—these are key. For Inno, it feels better—we mostly lead first-round investments. Years of 10x–100x valuation growth give us wide safety margins. But for later-stage VCs, it’s painful. You might go from 5 billion to 10 billion in a month, then 20 billion the next.
Source: AI Observer
Disclaimer: Contains third-party opinions, does not constitute financial advice
The AlphaGo Moment in Cybersecurity, or Starting from These 466 Agent Matchups
12 days ago
HuggingFace's "AI Duck" Goes Viral, Boosting Chinese Chip Supplier Rockchip
09-02
Yushu Surges 629% on First Day of Trading, Humanoid Robots Sell 5,500 Units
08-19
14 Executives Depart, Horizon Unleashes a Robot Legion
08-11
Yushu's 61 Billion IPO: Who Will Make Money?
08-07
AI stocks down 40%-60%, yet NVIDIA's valuation is at a 10-year low
08-06
HBM Bottleneck Migration, NVIDIA Also Aims to Downgrade
08-06






