Summary

Valhalla (Odin) splits the company into two narratives: long term, build “the last scientific general AI”; short term, build “the most powerful AI drug-discovery platform.” Its core asset is a full-modality molecular world model that can simultaneously model and design small molecules, proteins, RNA, and DNA—not a single-modality hammer. Odin believes the industry’s focus on single modalities is path dependence: “Once people have a good tool, they want to use it to hammer their pipeline and find a good nut.”
The most important investment signal is his candid answer on timing: “The one thing 2026 made possible was that the capital markets came back to life, so we had the money to do this.” This was not a technology inflection point: the technology was already in place when ODI was completed; what was missing was “how to organize everything.” Total funding is in the tens of millions of dollars, with 5Y Capital leading the first round and Qi participating, ZhenFund moving quickly in the second round, and a small extension round afterward; the team has 30 people.
The company is explicitly not building a pipeline. The reasons are twofold: a pipeline would distract from the AI effort, and “once you build a pipeline, you’re kind of the enemy of pharma, not pharma’s friend.” The host added that platforms and pipelines are not permanently opposed: “I build the platform to generate more pipelines, and those pipelines in turn prove that the platform’s value is real.”
The valuation strategy is counterintuitive: build a good company that every Chinese investor can afford to back. The host summarized it as collective participation—“make no company in the world too hard to invest in.” On dilution and control, Odin’s answer is that investors become friends once they invest, and “later-stage executive incentive plans can gradually give some of that back.” The pricing process itself was highly random: the FA quoted $100M, $120M, and $90M without being able to explain why; Odin ultimately decided, “This is the price—don’t ask why.”
His view of the competitive landscape is blunt: this is a Warring States period, and there may be a Qin five years from now. The technology is slowly converging, and “this era is very much like the early days of LMs, or even earlier.” He encourages everyone to invest more—“the more people invest, the more prosperous the field becomes”—and frames AI for Science in US-China terms: China is a step behind in LLMs, physical intelligence has a chance to catch up, and AI4S is where he hopes China can lead outright.
The most verifiable near-term milestone is an LD-type model. It would design arbitrary combinations of modalities, such as small molecule to protein for enzyme design or protein to DNA for diagnostic reagents. The most tangible application is “spending another RMB20 when you get your blood drawn today to test whether you have P-tau 217,” enabling early Alzheimer’s screening. For ordinary people, the fastest achievable milestone is a personalized drug designed around their genotype and in-body protein-expression levels.
The founder’s risk profile is part of the story. He is 26, dropped out of high school in his senior year, self-studied his way into Zhejiang University, and left Baker Lab, a Nobel-winning laboratory, by choice because a title is “a shackle made of jade—beautiful, but ultimately limiting your ability to become great.” He knows that the top seat in hard tech “basically offers only one shot,” and admits he was pushed into it: “The night wind blew me here, so I might as well do it.”

Deep dive

1. Defining Scientific AGI: Compressing 300 Years from Newton to Quantum Mechanics into 10 Years

Odin uses Kepler to define scientific AGI. Kepler’s discovery of the 3 laws was essentially “a great deal of data analysis, the formulation of various hypotheses, and eventually finding the right path.” If that process could be fully automated, “you could keep compressing the course of human history.”
His quantitative anchor is straightforward: humanity took 300 or 400 years to move from Newtonian mechanics to quantum mechanics. If that process could be automated, “you might have the chance to complete the entire scientific journey in 10 years, or even less.” This is not a new foundation; it is “still built on this generation of AI,” more like “the last, hugely important application for humanity, composed of a different set of infrastructure.”
The host explicitly rejects the idea that AI will replace people. “AI ultimately isn’t replacing you,” but freeing humans from complicated engineering problems so they can focus more on “the philosophical, more fundamental questions.” The company’s positioning has two layers: long term, scientific AGI; short term, “the most powerful AI drug-discovery platform.”

2. The Full-Modality Molecular World Model: Single-Modality Work Is Path Dependence, Not Inability

The term “molecular world model” predates the recent popularity of world models. “Before world models became popular, we already called ours a molecular world model”; once the term caught on, “we might as well stick with it.” The framework divides biomolecules into 4 categories—DNA, proteins, small molecules, and RNA. “If you can simulate and design every molecular modality, then you are effectively a god of the microscopic world.”
Asked why even places like Baker Lab focus on a single modality, he points to historical inertia and the tool-and-nail effect. Small-molecule researchers come from medicinal chemistry and want AI as a better computational tool for R&D: “Once people have a good tool, they want to use it to hammer their pipeline and find a good nut,” with the nut usually being a pipeline. The same pattern followed Baker’s protein-design breakthrough: “Once that hammer appears, everyone wants to hit every task with it.”
He names other teams working on full-modality systems, framing the disagreement as an architectural one. He cites Punny Stark of the Bose team and Jasper of R G Three. Odin sees his own advantage in having worked across the modalities: “Our biggest advantage is that we’ve worked on small molecules, nucleic acids, and proteins. On small molecules, we were one of the more cutting-edge groups at the time, so we have a lot of accumulated expertise.”
The host asks whether others simply cannot do it. Odin draws the line between posterior and prior knowledge: “If you know this in hindsight, then of course you can do it if you want to. But ex ante, you need enough courage to commit to the direction,” because the direction may not work and a viable technical path may not be obvious.

3. The Prior Bet Came from Physical Intuition: Every Intermolecular Force Is a Taylor Expansion of Electromagnetism

He traces his conviction back to studying physics 5 years ago. Physics has 4 fundamental interactions, while the hydrogen bonds, hydrophobic interactions, van der Waals interactions, and other forces discussed in biology are merely “Taylor expansions of different orders of the electromagnetic interaction”—different manifestations of an underlying force.
That led to the question of whether a unified force could describe all intermolecular interactions. He acknowledges that physics and human engineering capabilities made the problem intractable in the early days: “It may be very difficult to find a unified physical equation to solve all the problems.” But “in the age of AI, you can make the problem much more of a black box.”
The tactical path is deliberately indirect: solve small molecules first, then proteins, then nucleic acids. “Eventually you find that there are commonalities in how they are modeled, and then you have the opportunity to connect everything in one system.” His prior remains explicit: “I have always believed that we would have a model that unified everything.”

4. ODesign’s Validation: 8 Targets, 1 Undergraduate, Sub-Nanomolar Activity Every Round

Before founding the company, Odin’s open-source ODesign project worked with the Lingang Laboratory on an early full-modality molecular-design system. “One model could do small molecules, RNA, DNA, and proteins,” supported by basic downstream wet-lab validation.
The strongest data point was a single all-in design round across 8 targets. “Basically every round produced a molecule with sub-nanomolar activity.” The team was tiny: “I only had one undergraduate with me, and that undergraduate was called 杜刚.”
The significance was the lowered labor threshold, not just the activity level. Previously, each of 8 targets might require a PhD student to follow the entire design and experimental process. Now, “one undergraduate working with you can solve 8 targets,” which he presents as direct evidence of why scientific AGI matters.

5. The Unlock Was Capital Markets, Not a Technology Inflection Point

Asked why 2026, Odin gave an answer few founders would volunteer: “The one thing 2026 made possible was that the capital markets became active again, so we had the money to do this. We had the ability to do it very early; we were simply waiting for the right time.”
He explicitly denies a newly arrived technology inflection point. “When we finished ODI, the technology had already arrived.” The missing piece was “how to organize everything”: experts, algorithm teams, engineering teams, the right CRO, and the right university partner. “It takes a very complex system to support this project.”
The funding and headcount are concrete. 5Y Capital led the first round, Qi participated, ZhenFund moved quickly in the second round, and the company later completed a small extension; total funding is “around several tens of millions of dollars.” The team “just reached a headcount of 30 yesterday.” Before incorporation, it was already a complete small research group; after incorporation, the task was to bring those former colleagues back together to pursue “more visionary work,” including things they could not afford to do when they were poor.

6. The Wind That Finally Reached the Field: AlphaFold Is a Transformer Variant, Diffusion Turns 1-in-10,000 Screening into 1-in-48

Odin connects the technology chain in a straight line. AlphaFold “is effectively a variant of the transformer,” which made protein-structure prediction scalable. Then came diffusion: “Professor 宋阳’s SDE paper; we actually started exploring diffusion models very early, including for conformational prediction.”
The screening efficiency shift captures the magnitude of the change. “Previously, you had to screen more than 10,000 proteins to find 1 or 2 active compounds. Now, using what is likely RFdiffusion, you may find 1 in 48.”
His broader view of AI for Science is a combination of waiting for the wind and acting when it arrives. Different technical breakthroughs have to blow the wind into the field’s own scientific problems, but waiting for that wind and actively testing different approaches “are both indispensable.”

7. Next Milestone: Arbitrary Modality Combinations, Down to a RMB20 Alzheimer’s Screen

The technical milestone is the release of an LD-type model, “closer to the final molecular world model.” He deliberately avoids describing the architecture because that would turn the discussion into “a very technical report.” Instead, he focuses on the capability boundary: designing and validating molecules across arbitrary modality combinations.
The mappings are concrete. Designing a protein from a small molecule is enzyme design. Designing a small molecule from a small molecule can be used to detect metabolic waste in the human body. Designing DNA from a protein enables diagnostic reagents—for example, using P-tau 217 as a biomarker and designing a DNA aptamer that binds to it.
The most intuitive application is simple: “When you go to the hospital to have your blood drawn today, you could spend another RMB20 to test whether you have P-tau 217,” providing an early Alzheimer’s screen. The highest-value and fastest achievable milestone for ordinary people is personalized medicine: a drug tailored to an individual’s genotype and in-body protein-expression levels.

8. No Pipeline: A Platform Should Be Pharma’s Friend, Not Its Enemy; the 30-Person Team Was Complete on Day One

In the short term, the company will not build a drug pipeline, for 2 reasons. First, “it would distract you from pushing AI forward.” Second, “it becomes difficult to provide your services to large MNC pharma companies once you build a pipeline, because you’re kind of the enemy of pharma, not pharma’s friend.”
The host adds that platforms and pipelines are not permanently opposed. Historically, “they are organically unified: I build the platform to generate more pipelines, and those pipelines in turn prove that the platform’s value is real.”
The team made an atypical choice by staffing the back office on Day One. “Most startups lack an administrative team on day one. We had a very strong COO on day one; he not only understands operations, he comes from biotech.” From the COO to administration, HR, and even a cashier, the support structure was ready.
The other 2 components are the research team pursuing “the stars and the sea,” mainly students and postdocs from CUHK, and the commercialization team. The research mandate is: “You don’t need to think about commercialization today; just do the best work.” As for whether AI researchers need CS or model-training backgrounds, his answer is that “those labels are not that important.” Everyone is simply someone who has studied a little longer than others; what matters is the willingness to learn and adapt to challenges.

9. Leaving Baker Lab: the Nobel Aura Was “A Shackle Made of Jade”

Joining Baker Lab was itself a bet, beginning with pessimism about small molecules. “The small-molecule problem is fundamentally an engineering problem. Human engineering capabilities could not produce small molecules quickly enough, so your AI and your molecules were disconnected.” The only path then was to build a pipeline and use chemical-synthesis models to predict synthesizability, with “all kinds of compromises.” He applied to 3 schools and chose Baker Lab: “I was betting that Baker would win the Nobel Prize; I just didn’t expect it to happen so quickly.” The conviction came down to intuition: “I thought you could win the Nobel, so I went all in.”
Why leave when the Nobel-winning lab was gaining momentum and resources? “You can become a senior engineer, but you rarely get the chance to implement your own technical path.” He reduced the decision to a question: “Do you want to make something great, or do you want to collect some titles and then make something great? I couldn’t wait.”
He remembers the moment he made up his mind: on a balcony in Seattle, with Mount Rainier in view, reciting a Buddhist verse: “In this life I cultivated no good fruit, only loved killing and setting fires; suddenly I broke the golden cord, here tore apart the jade lock; the Qiantang tide has come, and today I finally know who I am.”
His description of titles was the sharpest line in the conversation: “It is a shackle made of jade—beautiful, but it limits your ability to become great.” The structural reason is that a large lab inevitably becomes a compromise among factions. The final model becomes the greatest common denominator, just as science has very little room for outliers.

10. Entrepreneurship Is Born of Dissatisfaction with the World: Credits Do Not Come Back to You

He reduces the motive for entrepreneurship to a declarative judgment: “Entrepreneurship is fundamentally our dissatisfaction with the world.” The mechanism is a mismatch between authorship and reward: whether in a large company or academia, “you make the contribution, but the credits do not return to you.”
He gives a verifiable example from a book they wrote, Graph Neural Networks and Modern Drug Design. The publisher wanted the professor listed first because he was a full professor and the book would look more authoritative, adding that “when writing a book, we also have to consider the social benefit.” Odin’s objection is direct: “The book is a crystallization of your ideas, written through 2 and a half years of work, but in the end the credits were redistributed for so-called social benefit.” That is why he says he likes watching 姜维’s films and repeats: “I came to Goose City to do only 3 things: fairness, fairness, and fairness.”
The same dissatisfaction runs through every departure: dropping out in his senior year of high school while retaining his enrollment and self-studying at home, cycling 10 km to the library each day, printing every year’s national college-entrance-exam English papers and working through them despite “wanting to vomit whenever I saw English,” then ultimately scoring 150 out of 150; leaving Zhejiang University; and leaving Baker Lab. “Fundamentally, they were all dissatisfaction with the world—no matter how hard you try, it feels as if some mysterious force is always boxing you in.”
After returning to China, he recruited in a distinctly old-school way. Starting in January, he called PhD students he had worked with or mentored: “It was like Zhu Yuanzhang raising a rebellion—going to Pei County to tell your brothers, ‘I’m doing this thing now. Let’s do it.’” He attributes the response to “conviction, will, and some of your historical reputation.” Historically, “most of what we wanted to do has been done.”

11. From Orphan Drugs to the “Steam Engine of a New Era of Scientific Discovery”: the Biggest Constraint Is Insufficient Funding

The vision has 2 levels. “Small-scale greatness” means designing molecules across different modalities while continuously compounding model capability with biology infrastructure. The immediate beneficiary is orphan drugs: many diseases have few patients, and commercial considerations leave many conditions with “no available treatment.”
He cites 蔡总, an ALS patient, who mobilized broad efforts from scientists, pharmaceutical companies, and grant programs. The field has since expanded, with more than 10 drugs moving through clinical trials. “You realize that it wasn’t that people couldn’t do it.”
“True greatness” is framed through the spinning jenny. If a spinning machine allowed 1 child to match the textile-production output of 10 adults, “we hope to let an undergraduate produce scientific discoveries at the level of a PhD student or even a professor”—the “steam engine of a new era of scientific discovery.”
Asked about the biggest challenge, he does not hedge: “The biggest challenge is that there isn’t enough money. Your ideal may be grand, but if there is no rice in the pot, the project cannot be completed.” Ultimately, this achievement “will certainly be completed through a collaborative body of industry and academia.”

12. Fundraising and Alienation: Building for VCs, the Market, or Yourself

He describes the collision of ideas that comes with raising AI4S capital during a frothy period. “Most investors still buy into the logic of big tech.” There is also a split on the technology: some believe model capability does not matter and only the commercial loop matters; others believe that if you are doing it, you should do the best thing possible. His response is that “somehow these are all noise,” invoking a personal maxim: “Remove the noise from the outside world, remove the noise inside yourself, focus on something, and ultimately become one with the sword.”
He calls the greatest fundraising risk “alienation,” and admits he has had the thought: “VCs are all telling this story today, so I’ll tell it too and get the money first.” He knows the VC story menu by heart: build a pipeline, find a seasoned scientist for AI for Science, build a pipeline, then monetize it; or assemble a world model, assemble agents—“agents may already be outdated”—then add robotics, physical intelligence, and perhaps anti-aging.
His answer to the 3-way choice is “Yourself,” returning to Buddhist doctrine: “All sentient beings possess the wisdom and virtue of the Tathagata, but cannot realize it because of delusion and attachment.” He treats adding robotics to increase valuation as delusion and insisting on telling his own story as attachment. Both have to be removed so he can return to his original motivation, “rather than asking how to lift the valuation or raise more money.”
He acknowledges that this answer does not work for most people. “Many people’s problem is that they don’t know what they are, so when you tell them to be yourself, they don’t know what ‘myself’ is. Honestly, I don’t know exactly what myself is either, but I roughly have a profile of it in my mind.”

13. Valuation Is Decided by Fiat: the World Is Not a Slapdash Operation, Just a Very Random Place

The biggest recent decision was whether to take money, how much to take, and at what valuation. Opinions pulled in opposite directions: some said the company needed a cash reserve at this stage and should ignore valuation; others said the dilution was too high and control would be lost. His decision was: “Build a good company that every Chinese investor can afford to back.” The host summarized it as “make no company in the world too hard to invest in.”
His direct response on dilution and control is to redefine investors as friends. “Once an investor puts money into you, they are no longer your enemy but your friend. As your friend, they cannot truly allow you to lose control,” and “some later-stage executive incentive plans can gradually give some of it back.”
The pricing process was almost entirely unsentimental. He asked an FA, identified in the recording as 翟飞, for a quote: “$100M, $120M, $90M—why? He couldn’t say either. You’re thinking, what?” He ultimately relied on references—how much predecessors and peers had raised—and decided: “This is the price. Don’t ask why.”
That led to his broader judgment about the world: people often think it is a slapdash operation, but “the world is not a slapdash operation; it is a very random world. All choices are basically the same.” What matters is how you turn a choice into a result. The practical implication is not to make the dice land in the right direction, but to “make them collide quickly and converge to the next stage quickly.” His advice to PhDs and professors preparing to spin out is pragmatic: “Find a good FA; that is more important than finding me,” and get a good law firm while learning the terms yourself. He has no formula for identifying a good FA and does not claim to review or summarize patterns; he relies on chemistry. If the conversation goes wrong? “Then take the loss. Losses are blessings.”

14. AI4S’s Warring States Period: One Qin in 5 Years

He views the wave of PhDs and professors leaving academia to start companies positively: “It’s a good thing; everyone gets to eat.” He also sees the field as cyclical: “This sector has been quiet for a long time, ever since the Waterloo of the previous generation of AI drug discovery.” In both AI drug discovery and AI for materials, no one can see the full picture, so different people can only be approached with different pitches.
His landscape call is the most quotable one: “This is what I would call a Warring States period. There may be a Qin 5 years from now. The technology of unification is slowly converging, so this era is very much like the early days of LMs, or even earlier.”
He takes no side against other directions: AI drug discovery, or AI for biology; AI materials design; organoids; and Virtual Cell all “have their own value and rationale.” His message is to invest more: “The more people invest, the more prosperous the field becomes.” His implicit country thesis is that China may be a step behind the US in LMs, while physical intelligence offers a chance to catch up or even pull ahead, and “we hope to lead outright in AI for Science.”
His view of architecture also contrasts with the host’s question. Architecture matters because it “basically determines the moat of your company,” but it also does not matter because “architecture is impermanent and changes rapidly.” The priority is the problem being solved: “The problem is your destination. To solve it, I may try different architectures.”

15. Odin and Valhalla: Reprogramming Life, and “If God Exists, How Could I Tolerate Not Being God?”

The name began as a pronunciation joke. His Chinese name is 昊天, a natural deity in Taoism. A French friend could not pronounce the “h” and voiced the “t” as a “d,” making it sound like Odin, so he adopted the name because “the meaning is the same.” But the deeper meaning is the episode’s central theme: “If God exists, how could I tolerate not becoming God? So God does not exist. If you can create every molecule, then you are God.” Asked whether that creates pressure, he says his fate is strong enough to bear it—and that he shares a birthday with Sakyamuni.
The technical path to “playing God” is laid out by intervention level. Baker represents protein intervention, “kind of like epigenetics—you might inject once a day.” The ability to interfere with and design RNA means intervening at transcription: “One injection might last half a month.” Gene editing addresses the problem at its source. The underlying assumption is that all life processes are controlled by interactions among different molecules; intervene in those interactions and you are “reprogramming life.” The endpoint is to create the signaling pathway, then a cell, and ultimately a life.
He does not evade the ethical question but draws a clear timeline: “It will, but not at this stage. We are still too far from the endpoint.” The real difficulty in creating life is consciousness—“whether there is a chance of creating consciousness from a combination of inorganic matter.” The line that best captures the company’s positioning is: “Pharma and synthetic biology are byproducts of our journey toward the stars and the sea.”
The company’s name was also an aesthetic choice. The original name, “Vientiane Bio,” was “too conventional and not rock’n’roll,” so they chose Valhalla. The second layer is a point of no return: the dead warriors gather in Valhalla for Ragnarok; “AFS is like the battlefield of Ragnarok. There will be troughs and upswings, but eventually it will reach something called the endgame.” He does not consider himself combative: “I am someone forced to fight.” He was soft-spoken as a child and took up boxing; “only those who can fight can stop fighting.”

16. Monastic Practice as Entrepreneurship: Wait, Fast, Think, and “the Most Important Thing Is to Get Out There First”

He compresses his daily routine into a Buddhist teaching: “Wait, fast, and think.” Waiting means that material conditions may not yet be sufficient to support a more ambitious project, so you wait for the right time. Fasting means having your own rules—when to wake up and when to sleep. Thinking means constantly asking what comes next. A typical day is waking up, eating 4 eggs, spending most of the day researching, coding, and reading extensively; after founding the company, commercialization and team building were added.
The phrase “This is a good thing” reflects an entire mental framework. “I have abandoned the binary opposition between good and bad, so everything is good.” Human conflict comes from greed, anger, ignorance, arrogance, and doubt—from attachments that cloud the mind. Once the dust is removed, “it is simply something happening. There is no good or bad; the good or bad is in your own mind.” He corrects the host’s passive framing of this as a change in mood: “That is a very passive way to put it. When your mind changes, your circumstances turn with it”—circumstances follow the mind.
Asked about his most recent delusion, he answers candidly: “For example, recording this podcast—I wondered whether this meant I was becoming some kind of public figure.” After thinking it through, he attributed it to coincidence and found little room for ego: “I keep reminding myself to practice the precepts, concentration, and wisdom, and extinguish greed, anger, and ignorance.”
The difference between being a number 2 and being number 1 is responsibility, not power. He once wanted to be the stereotypical ivory-tower technician: “I provide the technology, you give me money, and I do the research.” But he is impatient and sets high standards for himself. He understands the nature of the bet: “In hard tech, the top seat basically offers only one shot. Once you fail in hard tech, you essentially have no opportunity afterward.” The standard research path is to serve as a number 2 or number 3, accumulate experience, get a faculty position, and make a decisive attempt around age 40. He wanted that path too, but “the night wind blew me here, so I might as well do it.”
The line he recently memorized is also his advice to everyone still waiting: “The most important thing about entrepreneurship is to get out there first.” Thinking through the commercial model and team structure is far less important than starting. “Take the first half-step,” “raise the torch,” and “open your eyes and get to work.” The only thing that changed after founding the company was his mindset: “I take it more seriously now.” From Day One, he had to think about commercialization and became more realistic. What did not change is that “poor brothers remain poor brothers”; having money in the bank only reminds him why he started the company.
The closing question points to the same issue. If he could ask anyone in the world one question, he would ask Zuckerberg, “What is the meaning of the Patriarch’s coming from the West?” (何为祖师西来意) In other words, when you become rich and successful, can you still remember your original purpose? “I think the original intention matters more than anything. It tells you whether you built a company that makes money or a great company.”

41. Are AI Giants Buying Up Workers’ Screen Recordings? — Michael | Final Round / Hellyeah The $44 Billion Company Automating Your Company's Finances