Summary

The core proposition is simple: intelligence is inflationary, while wisdom is scarce. Once 100-200 labs and application companies deliver intelligence at scale, what remains scarce is value judgment after context is added: “The most important shift from intelligence to wisdom is to add context in between.” 汪天凡 therefore believes the opportunity in AI applications may be hardware-defined “context machines.” BAI has already backed 4-5 such companies, including Looki, memory-centric glasses, and always-connected wearables; the differentiation is not recording audio or video, but an interaction layer that “injects the radiance of humanity.”
The old fixation on network effects is outdated. In the AI era, the growth driver is the product of 3 forces: the coding data flywheel, AI for AI, and the AI-native organization. “We have never seen all 3 come together before,” which explains the revenue takeoff at Anthropic and OpenAI and applies to every industry: companies strong in all 3 could vault into the top 5%, including VC firms themselves.
The investment framework is the “Three Non-Theories”: non-consensus, discontinuity, and nonlinearity. Non-consensus means cheap pricing; discontinuity means incumbents are unwilling, unable, or afraid to enter; nonlinearity means multiple growth drivers compound into exponential growth. In plain English: “Buy cheap, grow fast, face no competition.” The accompanying methodology is to stop asking early-stage companies about moats—a defensive question that is “very ridiculous”—and ask instead about competitiveness.
The current rush out of applications and into AI infrastructure looks exactly like the crypto-infrastructure trade 3-4 years ago. It grants “extreme valuation flexibility to things that cannot temporarily be proven or disproven, while assuming only the upstream matters,” sending founders the wrong reward signal; “in a few years, you will find that most of what the labs produce is genuinely useless.” Near- and medium-term funding risks include a marginal decline in Big Tech AI capex, take-profit pressure when model companies go public and shareholder lockups expire, and lower public-market benchmarks feeding through to private markets. The host pressed whether “things could get a little dangerous by year-end or early next year,” or whether capex could continue through end-2028; Will then pulled the discussion back to a 3,000-year horizon.
AI may be a civilization-scale revolution compressed from 3,000 years into a few decades. If “everything is only just beginning,” investors should back people who can respond to change rather than things that will change: “The starting point determines the endpoint because the intervening process has been compressed.” A founder’s real ambition and vision matter far more than “looking at AI through the numbers.” Meshy’s 渊明 is the example: in 5 years, he moved from the Taichi lab to “fully connecting the model to the user,” while the vision remained AI for fun.
As of July 2026, the posture should be to make trade-offs and stay calm. An outstanding investor is one whose actions match their beliefs: pricing 2 rounds into 1, taking materially more capital than needed, accepting severe dilution, and pricing at scale before PMF are all signs of a bubble everyone can recognize; the test is whether “you can actually stay calm when it is time to be calm.” Last year, when AI hardware was hottest, he was instead anxious that “the market was giving the founders we backed the wrong reward signal.” If the market cools this year, investors should spend more time with the founders who are still committed. His only advice to individuals is proactivity: use AI like a gym, because unused capabilities atrophy; on prompting, he separately recommends writing more than 300 words each time.
On the consumer side, the attention economy may have a hard ceiling. People open their phones 97 times a day and scroll short video for 2-3 hours “without feeling any happiness,” so investors should back products that help users live in the present and experience genuine joy—Looki’s automatically generated comics and the offline AI dinner Table for Six, now in 6 cities with high repeat rates. The program also suggested that WHOOP, inferred from the pronunciation and fitness context, is among OpenAI’s top 20 token-consuming customers, feeding training context into ChatGPT to prepare for HYROX.

Deep dive

1. Fifteen years in, still an apprentice: coverage has changed

汪天凡 is 37. He interned at BAI in 2011 and joined the firm in 2012, where he has spent 15 years; before that, he had internships in consulting, banking, and marketing at ExxonMobil, then the world’s largest company by market capitalization. He has always described himself as having “an apprentice’s mindset.” An investor’s maturity is measured by the number of cycles they have lived through, and he has now seen 3-4.
Early coverage was bottom-up and scattershot: “If 36Kr reported on someone, we made sure not to miss them.” His reflection now is that constantly sourcing companies bottom-up “wastes your time”—by the time an investor sees the deal, it is already half-consensus or full-consensus, the price is full, and there is no information or insight arbitrage left to capture.
The highest bar for investors today is to combine 2 capabilities: dissect the world and build a foundational macro view—“Without faith, it is very difficult to invest in this era”—while also using products firsthand and testing models. Otherwise, it is easy to be misled by concepts.

2. Intelligence is inflationary, wisdom is scarce: the opportunity in AI applications is a context machine

When asked why “nobody has been investing in AI applications in recent months,” he initially agreed. The turning point was realizing that “intelligence is inflationary, but wisdom is scarce.” Once 100-200 labs and application companies deliver intelligence at scale, what is actually scarce? He posed the question to 3 AIs and got the answer that wisdom equals intelligence plus feedback, experience, and reflection—a composite judgment about why you do something, what matters to you, and what you choose not to do. Pure intelligence without your context cannot make value judgments. “The most important shift from intelligence to wisdom is to add context in between.”
His resulting “observation and exposure” is that this opportunity may be hardware-defined. BAI has invested in 4-5 context-machine companies: Looki; a pair of glasses built around memory and launching in the second half of the year; an always-connected wearable that does not need a phone; and an AI accessory that records only the wearer’s own voice, with privacy as the priority. The test is simple: “Can it remember what I say, hear, see, and even do as a result? Current devices cannot carry all of that.”

3. What RMB80 from Huaqiangbei cannot replicate: the interaction layer must “inject the radiance of humanity”

The host’s challenge was direct: collecting context is ultimately just recording audio and video, which an RMB80 Huaqiangbei knockoff can replicate. His answer was that recording is a shallow layer; the real value lies in interaction design, which must “inject the radiance of humanity”—making users feel, “I cannot go without wearing it; when I wear it, I can genuinely reflect, understand myself better, and receive a degree of care.” That depth is unavailable from PC and mobile applications or text-based chat.
Koji revisited his earlier complaint about 汪天凡: over the past 2-3 years, the AI sector had not produced many truly outstanding product managers. But that is precisely what the moment requires to define these products. “It is not a pure hardware company.”

4. His own AI workflow: more than a dozen prompts stored in WeChat Input Method

AI has changed his investment process at 3 levels. Pre-meeting preparation is now “extremely thorough,” using fixed prompts that classify founders as academic or product-oriented and overseas or domestic. AI mining of recorded diligence interviews gives him “an entire additional layer of understanding.” Post-investment, he uses AI to track public-information clues across hundreds of portfolio companies.
One prompt asks AI to “listen to the raw audio” of the latest 3 episodes of a podcast, map the logic, identify supporting data, surface “obvious non-consensus views” and gossip, and decide whether the episodes are worth his time. A product-diligence prompt covers firsthand user feedback, repeat rates, NPS, classic positioning theory, and valuation methods, producing “roughly a 20,000-word report.” Another scrapes every member of the latest Tsinghua Yao Class cohort—without missing a single person—and ranks them by “how impressive” they are. These were analyst assignments in the past.
His advice to everyone is to make prompts more than 300 words long. Tell the model what you are worried about, what you fear, and what you want; do not worry that the prompt sounds foolish. “The more you give it, the better it can help you analyze and make trade-offs.”

5. Network effects are outdated: 3 growth drivers must multiply to produce exponential growth

In a post on Jike, he argued that “stubbornly pursuing network effects may be a fairly outdated objective.” The growth drivers for AI companies have changed: the data flywheel already visible in coding; AI for AI, or using models to develop models; and the model company as an AI-native organization, using internal context effectively and reducing coordination friction. “We have never seen all 3 capabilities come together before.”
Exponential growth comes from multiplication, not addition. That is why Anthropic and OpenAI achieved such a rapid “revenue takeoff.” The framework applies across sectors—even to a restaurant: a company strong in all 3 drivers could “turn the industry upside down” and enter the top 5%. The same logic applies to VC, where it will generate a new cohort of investment firms.

6. Why he still loves the industry: rewards, privilege, trust, and unity of knowledge and action

After 15 years, he “really likes” VC. It is an industry of “reward and privilege”: few jobs can continually satisfy curiosity. The feedback cycle is “neither fast nor slow”—boards meet once a quarter, but “the moment we finish an investment, the feedback has already begun.” It is also an industry built on trust: fiduciary responsibility to LPs and trust in founders and teams. To earn insight arbitrage, “we also cannot create too much cognitive or information arbitrage within the team.”
On the claim that VC is driven by luck, he says people who are always trying to hit one big winner will attribute too many factors to luck. He certainly wants to hit big, but the key is the posture—and there is a problem that can be solved through engineering: move as much as possible of “what you do not know you do not know” into the category of “what you know you do not know.”
The standard for an outstanding investor is unity of knowledge and action. Everyone can recognize a bubble: investing at the valuations of 2 rounds simultaneously, raising materially more money than needed, accepting severe dilution, lacking governance, or pricing a company at scale before it has PMF. “Whether you can actually stay calm when it is time to be calm is a severe test.” In July 2026, his view was that the market had reached a point where investors needed to choose, give some things up, and make trade-offs. Last year, when AI hardware was hottest, the team was more anxious because “the market was giving the founders we backed the wrong reward signal.” If the market cools this year, investors should spend more time with founders who are genuinely still in it.

7. The attention economy may be hitting a ceiling: back products that help people live in the present

The biggest effect of more than 10 years of mobile internet has been the “severe fragmentation exploitation” created by the attention economy: people open their phones 97 times a day, scroll short video for 2-3 hours “without feeling any happiness,” and watch concerts through their phone screens. He does not believe this state can continue to scale indefinitely. The underlying mission is to help users live in the present—to “stop regretting the past and stop feeling anxious about the future.”
That is now his standard for AI hardware. The question is not how portable the form factor is, but “can it liberate users from their phones and computers?” Wearing Looki while walking through Tokyo lets users take photos without pulling out a phone. An AI meeting recorder can make people more focused on building connections. An AI camera used while raising a child means parents do not have to keep making the child pose, and can spend the time on eye contact instead. “This is respect for life,” and genuine happiness is far more valuable than cheap, pacifying entertainment.
Table for Six is a positive example: an offline AI social dinner that personalizes question cards on the table based on the 6 participants’ prior answers. It has expanded to 6 cities, has strong repeat usage, and attracts especially many women because “the environment feels very safe.” “AI seems to have brought out some of the radiance of humanity as well.” The program also mentioned WHOOP—an inference based on pronunciation and the fitness context—as one of OpenAI’s top 20 token-consuming customers, feeding training context into ChatGPT to prepare for HYROX. In the AI era, everyone is watching cyber; physical and offline experiences are being undervalued.

8. A 3,000-year civilizational revolution compressed into decades: the only advice is proactivity

His framework changes with the time horizon. Over 30 years, AI is a medium-sized technological revolution; over 300 years, it is an Industrial Revolution-scale event; over 3,000 years, it is civilizational—a new species and a competition on the scale of “Homo sapiens versus Neanderthals.” What makes AI unusual is the compression: “A 3,000-year civilizational revolution may be compressed into a few decades.”
Only a minority will cross that divide: people with extreme proactivity who can build AI-native organizations. “Most people in this world will not be able to cross this threshold.” The 1%-5% who do will separate from those who do not. For everyone else, his only recommendation is the first habit in The 7 Habits of Highly Effective People: proactivity. “The frequency and intensity with which you spar with AI now will determine the quality of your thinking and reasoning in the future,” just as muscle mass and cardiorespiratory fitness in your 30s and 40s determine quality of life. Use AI as a gym; capabilities atrophy without use. Otherwise, “the machine may be exploiting you rather than you exploiting the machine.”
He stresses that this is not an efficiency-above-all philosophy. AI told him that one distinction between wisdom and intelligence is “whether it is altruistic.” People should amplify altruism, care for others, reflection on themselves, and genuine emotions and memories. “That is a very important part of creativity.”

9. Creation as immortality: from being pretrained into models to mind uploading

His logic of immortality is straightforward: if what you do can be pretrained into a model, “your contribution and your existence in this world will be passed down as model versions are released and iterated.” Creative work therefore matters especially in this era. “The only thing you may be able to do is create, and then be pretrained into the model.”
Through Side Pocket in the UK, BAI has invested in a mind-uploading project. Starting from neuroscience, the team is experimenting on mice with brain signals and motion video, exploring models that could eventually control their actions.

10. AI infrastructure looks like crypto infrastructure then; on-chain finance is “a transformation no less significant than AI”

He compares today’s rotation out of applications and into AI infrastructure with the crypto-infrastructure boom 3-4 years ago: “extreme valuation flexibility for things that cannot temporarily be proven or disproven, while assuming only the upstream matters.” The positive effect is to direct resources toward basic research and infrastructure—“no industry can develop downstream applications without investing in the upstream.” The danger is the wrong reward signal: focus only on building new labs and models, without asking whether anyone will use them. “In a few years, you will find that most of what those labs produce is genuinely useless,” at which point they will face a survival problem.
Looking back at crypto, he says BAI invested in blockchain-based financial innovation: stablecoins and the on-chain issuance of US Treasuries, US equities, and dollars, which “is now an established fact” but was still nascent 5 years ago. His conclusion is that this “may very well be a transformation no less significant than AI”: it gives anyone with a phone and computer the right to choose their currency, allows dollar assets to penetrate more deeply into every corner of the world, and is tightly linked to AI capex, US Treasury issuance, and currency appreciation and depreciation. “It may even serve as a hedge.”

11. Looki’s investment story: calling for a pendant top-down and finding an overlooked team

1.5-2 years ago, the market became enthusiastic about AI glasses after the sales of what the source calls Manna Ribbon. He saw 2 misunderstandings. Users might simply be buying a pair of sunglasses, while the supply chain for SoCs, communications modules, batteries, and other components was not ready, leaving the product in a “wear it, stop wearing it” state. BAI’s earlier investment, VITURE, was an AR headset that simplified the product around viewing content—part of a different era. Inspired by Limitless’s Pendant, he reasoned that glasses could not support continuous recording, while a pendant could record audio all day and carry more context. BAI therefore called out to the market top-down: “We want to invest in a company making a pendant, not a company making glasses.”
The deal’s origin was itself a non-consensus footnote. A friend’s fund rejected the company at its investment committee and referred it to him. He and the founder, 孙杨 (phonetic), “clicked immediately,” moving straight to how the 1st and 2nd generations could become smaller, require less user intervention, and raise fewer privacy concerns. What impressed him most was the mispricing: a highly relevant integrated hardware-software team from Pony.ai and Momenta’s autonomous-driving businesses had somehow been overlooked by the market and had not received a fair evaluation.
The answer to Huaqiangbei substitution lies in software. Looki broke out by automatically turning a user’s day into an AI comic. The understanding behind that interaction and content-production design is “a scarce capability.” Of all the AI hardware launched over the past 2 years, “very few” have delivered happiness beyond efficiency.

12. Stop asking about moats; ask about competitiveness: the Three Non-Theories and the 2026 vintage

He no longer asks early-stage companies about moats. Scale economies, network effects, intangible assets, and switching costs are defensive “outcomes”; “when you ask an early-stage startup a defensive question, it is very ridiculous.” The right question is competitiveness. When the founder of Xinhaitu (phonetic) was forced to name a moat and answered “the core moat is iteration speed,” he was describing competitiveness, not a moat. Early-stage companies use new technology to attack new markets; the quadrant identified by the innovator’s dilemma is precisely the one incumbents naturally ignore.
His “Three Non-Theories” are non-consensus—there is pricing room and the asset is cheap; discontinuity—large companies are unwilling, unable, or afraid to enter; and nonlinearity—multiple growth drivers multiply into exponential growth that can scale within a fund’s 7-10-year life. In plain English: “Buy cheap, grow fast, face no competition.” No competition means less dilution and a lower risk of death.
2026 “is certainly not the best vintage,” but the conclusion depends on the time horizon. OpenAI was founded 10 years ago, and the investors who backed it then are now harvesting the fruit. On a short horizon, there are concrete risks: Big Tech AI capex may decline at the margin over the next few years; shareholders may take profit after model companies list and lockups expire; and lower public-market benchmarks may feed through to private markets. “Things could get a little dangerous by year-end or early next year,” or capex could continue through end-2028. On the compressed 3,000-year horizon, however, “everything is only just beginning.” The question is whether you are investing in things that will change or in people who can respond to change. “The starting point determines the endpoint because the intervening process has been compressed.” It may not take 3,000 years; the thesis could be proven or disproven in 30 years, or even 3.
Meshy’s 渊明 is his model founder. 5-6 years ago, BAI backed the Taichi language and engine. 渊明 spent 5 years “looking for nails with a hammer,” moving from a pure lab to “fully connecting the model to the user,” serving game developers and 3D-printing users. The vision has always been AI for fun: “If AI cannot make people have more fun, and AGI arrives 300 days from now and everything is handled by the efficiency guys, what will people actually do?” He invested in the vision, not “AI 3D.”

Verification Notes

The exact identity of WHOOP and the speaker in that segment remain unclear in the original captions. The transcript marks the speaker as [Speaker?]; the digest uses neutral wording and preserves the inference.

Management interviews: the most underdeveloped skill in investing | Ross O'Toole Lessons From Backing The Best Founders In Fintech | Micky Malka