Odyss AI Necklace: A Health Companion with 潘宇扬 & CreekStone's 李一豪

Summary

• 潘宇扬 is betting not on the novelty of an “AI necklace,” but on the idea that once glasses hit limits on comfort, battery life, appearance and prescription-lens costs, a neck-worn device may be the more first-principles multimodal entry point. In his view, Ray-Ban Meta’s several-million-unit annual sales rely mainly on the overseas sunglasses market and Ray-Ban’s distribution and brand, rather than proving that AI glasses have become a mass-market terminal; a necklace sits at the front of the body, can see and hear, and places roughly 50 grams on the part of the body best suited to carry weight. “Products should serve people,” rather than forcing users to wear glasses so AI can “hear what I hear and see what I see.”
• Odyss’s core function is to automatically record “every bite” of food and exercise, turning the full path from energy entering the body to being expended into data that can be planned around. It does not seek human-viewing image quality, but uses a low-frame-rate, low-resolution, high-contrast “AI camera”; the team found that 10 FPS delivered no recognition-benchmark gain over 30 FPS, then combines T0-to-T1 load prediction, on-device compression, small cloud models and asynchronous queues to trade limited bandwidth for a target battery life of 18–20 hours. “Every bite you eat is on the record”; even drinking instant-noodle broth can be inferred from changes in the liquid level and IMU motion.
• The team ruled out a general-purpose recorder from day one because hardware needs to give consumers a clear reason to buy, while context can only be a byproduct of a good experience. 潘宇扬 believes general AI is often just a “general-purpose illusion” assembled from multiple optimized vertical agents; meeting notes ultimately belong to the phone OS, and an ordinary person’s life may not be worth archiving all day. More importantly, many products summarize personal information into a system prompt without being able to verify, user by user, whether the agent actually improved or worsened, so context will ultimately prove itself in verticals such as health and coding.
• Odyss’s overseas entry point is not simply “weight loss,” but a mix of biohackers, GLP-1 users and people already at risk of chronic disease. 潘宇扬 cites market observations that roughly 190 million Americans have at least one chronic condition, while the global smart-wearables market has nearly 1 billion users; GLP-1 users eat less but may favor high-fat, fried or roasted foods when choosing a small amount of food, while people with chronic-disease risks need to identify hidden ingredients such as gluten, lactose and purines. The US defines obesity at a BMI of 30 versus 28 in China, yet some US interviewees with BMIs of 30 or even 40 still considered themselves “very strong,” so the product is selling health status and work efficiency, not appearance anxiety.
• Compared with photo-based calorie-counting apps, the necklace’s incremental value is not guessing the dish name one more time, but removing the friction of frequent logging and recovering the time information that a single photo discards. A single image cannot reveal a bowl’s depth, food underneath or how much of a shared dish someone actually ate; it also cannot show whether vegetables or carbohydrates came first, or how quickly someone ate. The team’s consensus is that ChatGPT may currently be the most accurate recognition app, but the limits are still inherent in the input. When the host wore a continuous glucose monitor, a bowl of beef noodles delayed until 3 p.m. drove blood glucose to 17–18 and left him asleep for an hour, showing that the valuable object is the behavioral chain, not just the calorie total.
• The criticism that the product is “anti-human” instead pushed the team to remove real-time interventions, retain only data presentation and future planning, and reduce privacy risk through product boundaries. 潘宇扬 says the harshest criticism came mainly from investors rather than users; his response is that humanity’s descent from the trees was itself “anti-ape,” and that being anti-human only limits the ceiling of the user base—it does not eliminate commercial value. Odyss has no gallery, users cannot view complete video or audio, and see only a health summary; most of the time, the underlying data is discarded after the model produces the summary. The hardware will also include an easy way to disconnect. The team will not interrupt users halfway through a meal to say, “You’ve eaten too much.”
• Consumer health hardware also succeeds or fails on whether it becomes an identity symbol, which is why Odyss wants to pursue a brand position that is “more Oura than Oura.” The name draws on the Odyssey-like idea of adventure and self-transcendence; jewelry qualities, product quality and social currency take priority over medical seriousness or Duolingo-style pure gamification. The team has produced nearly 10 industrial-design concepts and used scoring by US and UK target users, landing pages and ad tests to narrow them down. The company now has roughly 10 people, has completed a first financing round worth several million dollars and is in discussions for a second; it plans to launch overseas as early as Q2 next year and no later than Q3. 潘宇扬 believes the core technology can be validated before launch; the biggest uncertainties remain demand for a new category, geopolitics and regulation.

Deep dive

1. Odyss Sets Its Boundaries Around Overseas Health, Not a General-Purpose AI Gateway

• 潘宇扬 is 28, graduated from the University of Electronic Science and Technology of China, and describes himself as an INTJ and Gemini; the team has “just 10 people,” has completed its first financing round worth several million dollars and is now engaging with investors for a second.
• The product is an AI necklace designed for all-day wear that recognizes food and exercise, organizes the resulting data and builds health plans; the plan is to launch overseas no earlier than Q2 next year and no later than Q3, with a possible China entry “in the future.”
• This is not a product already in mass production: the current prototype has not completed final power-consumption and structural optimization, and the masculine industrial design shown publicly is only one of nearly 10 concepts.

2. AI Glasses’ Hype Masks Four Physical Barriers That Will Evolve Slowly

• 潘宇扬 is not permanently bearish on glasses, but believes building AI glasses today is “like building VR in 2015”; if myopia or corrective optical surgery is still not widespread 10 years from now, glasses may become a terminal.
• In his view, Ray-Ban Meta’s several-million-unit annual sales are driven primarily by the huge overseas sunglasses market and Ray-Ban’s near-monopoly distribution and brand power; China lacks both conditions.
• A true mass-market product must be comfortable, require little charging, look good and fit into a dress code; standard glasses weighing roughly 48.6 grams already press on the ears and nose, while materials and batteries will not advance every 6 months like large models.
• The replacement cost of glasses also includes lenses, refraction and pupillary-distance measurements; the lenses may cost more than the frame. For people overseas with normal vision, the motivation to wear glasses solely to chat with AI is weaker still.

3. The Neck-Worn Form Emerges From the Intersection of “Invisible to Wear” and “Positioned at the Front”

• The most important condition for a multimodal gateway is not a long feature list, but that “you won’t feel uncomfortable after an hour”; 潘宇扬 believes the neck is the body’s best load-bearing position, and that even if a roughly 50-gram necklace slips into an odd position, users may not notice it.
• The second condition is being able to see and hear clearly, which requires positioning the device at the front of the body; fingers and wrists are poor locations. Combining the two constraints leads to a neck-worn form, rather than starting with jewelry and searching for a use case.
• His product philosophy is that “being close to the eyes and ears” is merely a way to serve AI; a new experience must exceed the old experience plus switching costs, and the final question is still how it serves people.

4. Food Is the Biggest Unfilled Gap in Health Data

• The core insight 潘宇扬 took from Up Live—known in Chinese as Beyond 100 Years—is that health should shift from treating illness after it appears to preventing future risk through lifestyle planning.
• Food, exercise, sleep and mood all affect health, but the latter three already have abundant products. Diet, despite its enormous impact, lacks dedicated hardware. “This is a completely blank market,” and it happens several times a day for everyone.
• The team tried putting cameras in spoons and placemats, but found that they solved only point problems and could not cross-check a piece of meat against information such as voice and location at the time of ordering. The ideas were eventually abandoned.

5. 潘宇扬’s Career Connects AI, OS, Agents and Hardware

• He joined Huawei through campus recruitment in 2019, first working on algorithms and strategy for Xiaoyi, the smart assistant, in the pre-large-model era, then moving to HarmonyOS, based on the early OpenHarmony framework. His experience spans both AI and operating systems.
• After joining ByteDance, he initially entered a confidential project that still has not fully launched; before Flow, the broader organization that housed Doubao, was established, a group of product managers were already exploring new internal topics.
• Coze originally aimed to turn industry knowledge into Agents, consolidate it into platform data assets and serve a larger AOS; that starting point was later shown to be “not that workable.” After the product’s first public launch in 2024, he moved into AI hardware.
• He later worked on forms including the Ola Friend earbuds and also explored AI glasses; those experiments led him to conclude that short-term mass-market terminals need a different carrier.

6. CreekStone Wants to Build a Chinese AI-Founder Network With Two People

• 李一豪 is 36, studied at Tongji University as an undergraduate and in France for a master’s degree, and has spent the past 8 years in VC. With partner 陆欢, he founded a two-person early-stage AI dollar fund that combines investing and incubation.
• The ambition is not merely to fund companies, but to form an entrepreneur community resembling YC and the Silicon Valley mafia, where founders across AI entertainment, productivity tools and Agent infra can continually collide.
• The name CreekStone was generated by Doubao from a prompt, while the logo was made with help from Figma AI; “water benefits all things, stone stands firm” represents patient capital and steadfast support at critical moments.

7. General-Purpose Hardware Lacks a Purchase Reason and Is Easily Absorbed by the Phone OS

• From day one of the startup, 潘宇扬 “completely prohibited doing anything general-purpose”; his view is that so-called general products are often just collections of vertically optimized Agents, while the real value remains in the optimized use cases.
• Software can download 100 apps in a day, but hardware needs a clear purchase motivation and must answer, “What am I buying this for?” The answer must be either emotional value or a concrete pain point.
• He cites the DJI drones he buys every generation and the convertible whose roof he opens only once a year as examples of emotional value: “It makes me feel powerful.” Even low usage can be enough to drive a purchase.
• A meeting-minutes peripheral may create short-term PMF through convenience, but system-level summaries across Feishu, Tencent Meeting and phone calls will ultimately belong to the phone OS. As for recording daily life all day, he says bluntly, “Ordinary people’s lives are actually very boring.”

8. The Product Does Not Photograph a Meal; It Recognizes Every Bite

• Odyss aims to record when a person puts meat, vegetables or other food of what size into their mouth, then estimate nutritional composition, calories, GI and the effect on the body.
• Exercise data will enter the same chain, with the goal of explaining “what happens to energy from the moment it flows into the body until it ultimately flows out,” rather than merely generating a food list.
• The host joked that hardware monitoring installed on a toilet would “close the loop.” The episode did not say Odyss covers this endpoint, but its data ambition is indeed to connect intake, activity and bodily outcomes.

9. The “AI Camera” Sacrifices Human-Eye Image Quality for All-Day Power Efficiency

• Because snacks, coffee and water happen in scattered moments, requiring users to tap to start recording each time is unrealistic, so the camera must be always on; the team also removed the gallery on day one.
• 潘宇扬 distinguishes between two types of imagery: humans like high resolution, high frame rates, saturated colors and realism, while AI needs high contrast, high dynamic range, sharpened edges and clear contours.
• “To build a perfect human camera, you’d have to make something like a DJI or Insta360 product,” but the size, power draw and cost would all be high, and that is not the team’s strength. Odyss therefore chooses low frame rates, low resolution and on-device feature enhancement and compression.
• The team’s benchmark showed that both 10 FPS and 30 FPS were already well above the current VLM sweet spot; raising the frame rate brought no gain. “So we’re going to build an AI camera.”

10. The Device Requests Data by Scenario Rather Than Mechanically Streaming Full Video

• While recording a podcast, the image barely changes and the system should listen more; while cooking, it needs faster visuals; for text, QR codes and barcodes, it needs clarity but not a high frame rate. The load must be allocated dynamically.
• On-device processing is limited to simple compression such as histogram matching: if two consecutive frames are highly similar, there is no need to upload both. Genuine scene understanding is handled by an independent cloud VLA system.
• The model predicts the load needed at T1 from the T0 frame and outputs only JSON parameters; 潘宇扬 says it could be 3B, 1B, 0.5B or smaller, with latency of roughly 1 second or several hundred milliseconds.
• The instant-noodle-broth corner case does not rely only on “seeing the liquid enter the mouth”: the system compares the liquid level before and after the bowl is lifted, then combines it with the duration of movement in the IMU to infer consumption. The host summarized it as, “The net of heaven is vast, but nothing slips through.”

11. Compression and Asynchrony Work Around Bandwidth; The Battery Target Is One Charge a Day

• 潘宇扬 acknowledges that the industry has no transmission method that is both fast and power-efficient, and he does not plan to solve that problem alone; his bandwidth estimates are roughly 30KB over BLE and up to about 80KB over Bluetooth, while Wi-Fi is faster but power-hungry.
• A T0-to-T1 judgment may require only a 10KB or 20KB image, well within Bluetooth’s range; only a small number of high-definition tasks would activate Wi-Fi, while queue redundancy allows them to be processed later.
• Health results do not need to appear in real time bite by bite; delays of several seconds or even several minutes are acceptable. As with sleep-ring data that users typically review after waking, asynchronous processing may better match user behavior.
• The product target is 18–20 hours, or one charge per day. Users would remove the necklace before bed and hang it on an “exquisite jewelry charging stand” before wearing it again the next day; the battery should be hidden behind the necklace or elsewhere to avoid disrupting the front-facing design.

12. The Core of the Privacy Design Is Leaving No Material Anyone Can Abuse

• In overseas interviews, the team found that users were more concerned about “the person across from them” misusing images than about the company’s abstract use of data; that distinction changed the product boundary.
• Odyss has no gallery, users cannot see complete video or audio, and see only a health summary; most of the time, the data is discarded after the model generates the summary, and the company will not emphasize camera specifications in its marketing.
• The hardware will include an easy manual disconnect so users can shut it off immediately in private settings; if an extremely sensitive environment requires even a phone to be left outside, the necklace should follow the same rule.
• 潘宇扬 acknowledges that social attitudes will take time to change. The product’s perceptible signals have been pushed as far as possible, but privacy settings will still present a broader social challenge.

13. More Context Is Better, but the Reason for Collecting It Must First Make Sense to Users

• 潘宇扬 has adjusted his view over the past few months: in the long run, more context is still better. GitHub was once merely a code repository; only after AI coding emerged did its value as training data become clear.
• But ordinary users will not buy hardware because it “helps collect your context.” “My mother doesn’t know what context means.” A good experience must come first; continuous data can only be a byproduct of immersive use.
• Many current products summarize personal data into a system prompt, yet no one can evaluate for each user whether the modified Agent actually improved or worsened; the rate of information injection is far faster than the ability to verify its value.
• Context will therefore develop like a barrel with one long stave first: GitHub first enabled coding, while food data may first form a closed loop in health before gradually filling in other dimensions.

14. “Anti-Human” Did Not Kill Demand, but It Made the Team Remove Real-Time Intervention

• The harshest criticism 潘宇扬 heard was, “You’re building an anti-human product,” and it came mainly from investors; none of the interviewed users described it that way.
• His rebuttal was deliberately sharp: short video follows human nature, but “apes coming down from trees and becoming human” was itself “anti-ape.” An anti-human product might not serve 7 billion people, but could still reach close to 1 billion global smart-wearables users.
• The team accepted the valid part of the criticism: the early concept would have alerted users when they were full or had reached their calorie limit, but all of that was removed. Once the meal has been made, saying “you ate too much” is not helpful.
• The first generation will only present data and plan future behavior. Whether to add real-time intervention in the next generation will depend on actual usage, rather than turning the product into a disciplinarian in advance.

15. Broad Demand and GLP-1 Users Share the Same “Nutritionist on Call”

• The broad value is turning “what you ate, why you chose it and how it relates to your body” from a messy process into scientific, plannable behavior; users do not need to learn extensive nutrition science or weigh every meal.
• 潘宇扬 compares the product to a professional nutritionist who is “always standing by”: it does not require users to actively photograph and report, and tries not to disrupt the rhythm of daily life.
• GLP-1 users shocked the team because they feel full and can eat only small amounts, yet may favor high-fat, fried or roasted foods when choosing that limited portion, worsening their nutritional balance.
• These users do not need to keep cutting their intake; they need to plan limited food around their daily rhythm and avoid hypoglycemia, low mood and poor concentration.

16. Chronic-Disease Risk and Biohackers Form the Two Ends of the Overseas Market

• The figure 潘宇扬 cites is that roughly 190 million Americans have at least one chronic condition; people over 35 often “wake up” after a family member falls ill or a doctor issues a warning and begin investing in their health.
• Lactose, gluten, purines and other dietary components are difficult to identify by intuition. The necklace can link allergies, high uric acid and other risks to actual intake, but the team is explicit that it will not package the product as serious medical equipment.
• At the other end are Oura Ring- and Whoop-style biohackers who may already be healthy but want a complete view of daily changes; some people have even bought camera-equipped food scales priced at $30–40, with annual shipments of roughly 1–2 million units.
• The two needs sit on one product line: biohackers consume the data presentation, while people with health concerns consume the recommendations converted from that data. They are simply using the same underlying dataset through different tabs.

17. US Users Care About Health Efficiency; The Product Need Not Begin With a Chinese Weight-Loss Narrative

• If the team started in China, it might make weight loss a prominent selling point; overseas, it will not, because local culture is more accepting of different body types and appearance anxiety may be a weak motivator.
• 潘宇扬’s comparison is that China’s obesity threshold is a BMI of 28, versus 30 in the US; in interviews, some people with BMIs of 30 or even 40 still did not consider themselves fat, but “very strong.”
• The product therefore asks whether being above a healthy range affects work efficiency, mental state and physical condition, rather than telling users, “You’re too fat.” Weight loss is only one possible outcome of health management.

18. Behavioral Data Matters More Than Calorie Counts

• Compared with photo-based calorie apps, the necklace removes the need to pull out a phone, take a picture and manually correct the result every time someone eats; snacks and extra meals are scattered throughout the day, making that friction a natural barrier to mass use.
• A single photo cannot reveal the depth of a bowl, food underneath, how much of a shared dish someone actually ate or what was left over; even the strongest model is constrained by incomplete input.
• It also misses eating order and speed: whether vegetables or carbohydrates came first affects GI, while eating too quickly delays satiety and can lead users to keep eating before their bodies respond.
• A continuous glucose monitor worn by the host showed the point clearly: beef noodles eaten at 3 p.m. pushed blood glucose to 17–18 and caused an hour of sleep. The valuable object is the behavior chain linking visual, motion, time and environmental data.

19. Dietary Blind Spots Are Often Everyday Misjudgments, Not Advanced Knowledge

• 潘宇扬 gives the example of a relative with diabetes who refused watermelon because it was too sweet but happily ate sour hawthorn candy; the latter may contain far more sugar. This kind of taste-based nutritional misjudgment happens every day.
• While wearing a continuous glucose monitor, the host ate beef noodles on an empty stomach at 3 p.m.; although he ate the beef first and only about one-third of the noodles, his blood glucose still rose to 17–18, and he fell asleep on the sofa for an hour.
• 潘宇扬 compares mixed drinks to the body’s dependence on patterns: the body prepares hormonal and metabolic responses for a fixed routine, so abruptly changing concentration, timing or food structure can trigger a stronger stress response.
• He also says people tend to store calories in uncertain environments; industrial food has evolved far faster than the body, and the impulse to finish whatever food is in front of us still reflects the old fear that there may not be another meal.

20. Odyss Wants to Be Jewelry and an Identity Marker, Not a Medical Instrument

• The team is explicitly avoiding a CGM-style medical route: products in that category took more than 20 years to move from specific diseases into general health, along with certification and compliance burdens.
• It also will not simply replicate Duolingo-style gamification. 潘宇扬 has maintained a streak of more than 700 days but is still afraid to speak for real; emotional value can retain users without necessarily delivering capability.
• The name Odyss draws from Odyssey and expresses a journey of adventure, self-transcendence and living longer with more vitality. 潘宇扬 extends that spirit to the identity narratives represented by Westerns, Marlboro, whiskey, Jeep and Hummer.
• For wearable jewelry, “the meaning it conveys and what other people think of it are the first priority.” Accuracy and functionality are secondary questions, because users must first be willing to keep wearing it.

21. The Goal Is Not to Copy Oura, but to Become “More Oura Than Oura”

• In 潘宇扬’s view, Oura Ring is highly abstract: the app uses different landscapes to carry metrics accumulated over time, creating the feeling that “the body is slowly recovering and continuously gaining energy.”
• He also notes that fingers have less blood flow and more joint interference, so some measurements may not be more accurate than those from the wrist. Yet the ring’s distinctive form and imaginative space allow it to transcend pure parameter competition.
• Whoop’s path starts with professional athletes, sports stars, fitness KOLs and a data-dense dashboard. It has a strong geek identity and is harder to take mainstream.
• Because Odyss can identify exercise visually, it does not need users to select an exercise type first and then match an IMU algorithm as Apple Watch does. Still, 潘宇扬 believes a complete story, concept and imaginative space matter more than adding one more data point.

22. Whether AI Hardware Survives Long Term Must Be Tested Against First Principles 10 Years Out

• 潘宇扬 does not want to predict which form factor will definitely survive a year from now, because model capabilities and hardware conventions are still changing rapidly; stretching the horizon is what reveals which products are merely transitional.
• Plaud’s card is a good product and can achieve short-term PMF, but he cannot imagine users still attaching a recording card to the back of their phones 10 years from now. Whatever happens, that function will enter the phone.
• Plaud once made a necklace version of NotePin, but sales lagged the card; its third-generation product, with a screen, moved back to a card form. In his view, the necklace failed to follow the first principles of location: recording and the phone can sit together.
• He frames a hardware purchase as A, the user; B, the decision logic; C, the product; and D, the problem. All four must be clear. If the form, motivation, product or problem is ambiguous, the story is difficult to make work.

23. Industrial Design Is a Selection Under Three-Way Constraints, Not a Single Genius Decision

• The necklace must fit internal components, so it cannot compete with pure jewelry on aesthetics alone. The team’s goal is to make it attractive and refined, then use technology to narrow the gap in decorative value.
• They toured luxury-goods stores with designers to find inspiration for structures that could accommodate the components, producing nearly 10 different industrial-design concepts. The team judged stacking feasibility, designers supplied the aesthetics, and target users decided whether the result was acceptable.
• Users in the US and UK first scored the concepts; 3 were selected for model fitting, feature explanations and different landing pages, followed by ad campaigns to observe a larger sample. The final choice has not yet been made.
• 潘宇扬 emphasizes that software is usually downloaded for its function and only then judged by its UI; hardware is judged first by appearance, then function, and finally use. This corresponds to P&G’s two “magic moments”: being noticed on the shelf and actually beginning to use the product.

24. CreekStone Is Betting on a Few People With “Huge Ambition and Tiny Ego”

• 李一豪 says he sees roughly 500 people a year, but those who meet the standard “have never exceeded 5, maybe 2 or 3.” The first requirement is huge ambition and a tiny ego, alongside deep self-awareness and a relatively complete worldview.
• Ego is tested not with personality questions, but by presenting contrarian views, asking fatal strategic questions and even commenting on team members. Only people who can discuss these issues openly and make the hard call are refusing to use self-esteem to defend prior choices.
• The second standard is the courage to reject consensus while operating above common sense: first understand that hardware must look good and purchase costs must be reasonable, then choose a field they can define and that matches their capabilities.
• Continuous learning and “being generous with relationships” matter too. 李一豪 saw 潘宇扬 carrying an AI device protruding from his pocket, a backpack full of hardware and detailed knowledge of every spec; he also saw him accept sharp advice and spark collaborations with founders across different tracks.

25. Investment Relationships Are Built Through High-Density Accompaniment, Not Priced in a Single Pitch

• When the two first met, 潘宇扬 was still at ByteDance. He knew he would eventually start a company but had not resigned to do so. 李一豪 then sent him long AI updates almost every day, making clear that this was not “spray and pray.”
• 潘宇扬 understands CreekStone’s AI-native approach as “having no path dependence”: young people use first principles and product innovation to solve vertical problems in AI’s main tracks.
• More importantly, the investor enters the founder’s life and breaks down company structure, hiring and partnership models one by one. 李一豪 supplies keywords; 潘宇扬 searches and builds his own understanding instead of copying answers.
• In another round, the Linear Capital team spoke with him from 9 a.m. until 9 or 10 p.m., questioning everything from product and human nature to underlying motivation. His reaction was, “It was so fucking exhausting,” immediately followed by, “It was so fucking exhilarating. This is entrepreneurship.”

26. The AI Era Took VC’s Map Away and Amplified ByteDance’s Talent Spillover

• 李一豪 sees AI as a new stochastic paradigm, with “no map” as the most important reality. GPs need to stand beside founders, encounter similar people and context, and provide support with a “weak-side mindset,” rather than over-modeling or over-directing.
• Funds are therefore becoming smaller and more agile, while partners step forward to express their views. Meanwhile, exceptional individuals can amplify the final 1% of execution into a chance to capture 99% of the value; passion, focus and taste matter more than credentials.
• 潘宇扬 believes ByteDance can never truly be “always entrepreneurial,” but it is still the most entrepreneurial large company: projects grow more from the bottom up, and products that do not yet exist in the market are allowed to emerge internally. Coze is one example.
• 李一豪 values the context ByteDance employees have occupied: they have seen products with 1 billion users and may have directly worked with trillion-parameter models, adjusting model performance through data. But he does not limit the opportunity to big tech; independent teams can also develop first-position insights through sharper contact with culture, people and social change.

27. Entrepreneurship Is Ultimately Decided by People, the Market and the Ability to Keep Adapting

• 潘宇扬’s deepest motivation is that after his life ends, someone will still be “better because of something I did.” Large-company products feel like “dancing in shackles,” with time consumed by resource competition and persuading people above and below, sometimes with people who have never met users directing the product around a PowerPoint.
• The shift from software to hardware is that hardware cannot be A/B-tested or rolled back at will; once molds are made and products enter mass production, changes are nearly impossible. The shift from employment to entrepreneurship is that every dollar feels as if it is coming from his own pocket, while every decision affects the entire team. He has not played a single game in the 6 months since starting the company—not because he has no time, but because even when he has time, he cannot bring himself to play.
• If Odyss succeeds, the short-term reason will be that product, marketing and supply chain were all executed correctly; the long-term reason will certainly be that “we picked the right people.” If it fails, the technology can probably be exhaustively validated before launch, and even if a large company builds the form, it may not go as deep into health as he does. The real unknowns are demand for a new category, geopolitics and regulation.
• The team is mostly made up of people born after 1995. Hiring prioritizes entrepreneurial spirit, first-principles thinking and exceptional learning ability, with conventional business skills ranked last. The company will not expand its full-time staff unless necessary; it hires only when 99% certain that someone will fit.

28. 潘宇扬’s Startup Playbook Is to Define Problems Rather Than Solve the Hardest Ones

• His 4 principles for founders at the same stage are: do not try to solve industry problems that even OpenAI cannot solve; use first principles to define a new category; focus on a vertical and use a rough MVP to validate PMF; and have the founder stay close to the market and users.
• His warning is that startups should not merely replace one module of an old product with AI, nor pursue complete, military-style standards from day one. “Even if it has holes everywhere, as long as it has PMF, we always have a chance to scale it.”
• If given $3M to invest as an angel, he would first invest in 李一豪, though he also says he definitely could not get into 李一豪’s fund because the amount is too small. The second slot would go to a young former ByteDance colleague who is still working there, helping him cross the river from employee to founder.
• The third choice would be a friend who makes coffee: despite a privileged family and strong educational background, he dared to turn a hobby into a business and lightly validate PMF with a single SKU. 潘宇扬 says he is willing to “send charcoal in a snowstorm,” and “willing to pay for idealists.”