Ryan Kim

Ryan Kim@0xryankim · 18 天前观点输出

Everything is happening all at once. We are in the world changing with the fastest speed. We will live at least 200~300 years or maybe we will not die. DNA was preserved for survival, but we don't need to worry too much about survival. I guess people will choose to die at the end. why do they choose to die? because it's not fun anymore. Human is animal of motivation. If we don't have motivation to live, no reason to live. So they have to convince you have reason to live, and that's from motivation. Motivation can be various, fun, mission, love, etc. You have to make motivation portfolio to live longer. If you find really good person to live together, marry. If you want to have mission, make company. If you find geniune passion for some appetite, learn and develop. and have fun. If you have thousands of reasons to live, you will live forever. https://t.co/rQbWFtOdHq

104232

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Ryan Kim@0xryankim · 2026/08/06观点输出

Welcome Gurnoor! Let's make the next chapter together :)

101878

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Ryan Kim@0xryankim · 2026/06/16政策影响

The internet is entering its "own" era. The most valuable thing to own might surprise you. a16z's latest piece, "The AI ownership rush," makes a sharp point: the internet has moved from read → write → own, and AI is the apex of that arc. The evidence is everywhere — the scramble for exposure to the big labs through SPVs, secondaries, even pre-IPO synthetic perpetuals. As they put it, "people want in… even when the ownership itself is synthetic or uncertain." They're right. The hunger to own a piece of AI is real, and it's only getting started. I'd just add one layer to it. Most of that rush is aimed at owning the model — equity in the labs, exposure to the weights. That's a rational trade. But weights are also the fastest-moving part of the stack: frontier capability gets matched, then commoditized, and inference costs fall toward zero. The model layer is where the value is created — but it may not be where it ultimately settles. So alongside "who owns the model," I think the more interesting question is: what does the user get to own? a16z actually points at the answer themselves. One of the first things they open-sourced was a companion stack — a way to build and host your own AI companion with memory. Their consumer thesis is moving from productivity toward emotional connection: AI that learns from your photos, your conversations, your routines, and reflects you back. The data backs the shift. 72% of US teens have tried an AI companion; over half use one regularly — and yet 80% still say they prefer real friends. They're not replacing relationships, they're rehearsing them. Either way, the thing that matters is that the AI knows them — six months of memory, yesterday's argument, the way they talk. That's the part that compounds. It survives a model swap from one lab to the next, and it creates friction when it moves. Models get better and cheaper; the relationship only gets deeper. One depreciates, the other accrues. This is why I think the next Disney and the next consumer-crypto breakout both live in this layer. Disney's moat was never the animation technique — it was Mickey, and the relationship people built with him. Netflix spinning up a GenAI studio is the same bet: generation becomes commodity, IP and relationship become the asset. And it's where own gets its fullest meaning. Owning lab equity is one form. The form I'd watch just as closely is the consumer owning their own character, memory, and relationship — portable, and theirs. That's an asset you can't buy through a pre-IPO perp, because it isn't issued. It's accumulated. a16z is right that the own era is here. I just think the biggest thing to own won't be the model — it'll be the relationship built on top of it. — Refs: • a16z, "The AI ownership rush": https://t.co/HYADeDHXPD • a16z's open-source companion stack (build your own AI companion with memory): https://t.co/dLtB7PtYaG

106625

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Ryan Kim@0xryankim · 2026/06/15赛道影响

The next AI bottleneck isn't payments. It's trust. And trust isn't built with features — it's built through relationship. For two years we've been building AI as the Assistant. It explodes productivity. But productivity has a shadow: AI sharpens competition between people, and the algorithms push our thinking to the extremes. More and more people end up lonely. And where does a lonely person go? YouTube, TikTok, Instagram — places where you only receive what someone else made. Consumption covers loneliness for a moment. It never fills it. You take in; you never create. Yet people still chat, constantly. Chat is nearly the last space of active creation humans have left. If its share of our screen time is shrinking, the reason is simple: we have fewer people to talk to than we used to. And of the few we have, how many can we open up to completely? You can count them on one hand — and even those narrow under knowledge gaps, politics, or simply having to watch your mouth. This is the opening for the Companion. Not the chatbot. Not the fantasy machine. A relationship you actually have to develop over time. Three things most people get wrong about it: 1. A Companion is not "the one who always agrees with you." That's just an echo chamber — the exact problem AI already created. A partner that flatters you cages you. A good companion has to be able to say "you're wrong about this." Not a validation machine. A growth partner. 2. It augments human relationships — it doesn't replace them. The right design is a practice ground and a cushion: a place to sort out your feelings, then walk back out into real relationships stronger. A companion that pushes family and friends away can win short-term retention, but it collides with society in the end. (And notice: early data shows people who use companions actually spend more time with human friends. That's not a threat to the thesis. It's proof of it.) 3. The real battleground is incentive alignment — not model quality. If "the friend you trust with everything" also "recommends great products," whose side is that recommendation on? A companion that lives on commerce commissions ends up working for the advertiser. Only a model where the user pays directly keeps the companion on the user's side all the way through. Here's the part most people miss: A Companion isn't a category. It's the on-ramp. For AI to take the center of our lives, capability was never the bottleneck. Emotional resistance is. People don't withhold trust because AI is incompetent — they withhold it because they can't yet give it their heart. The Companion is the first interface that dismantles that barrier. Daily conversation. Remembered context. A presence that doesn't change. And once that trust forms, it doesn't stay put — it transfers. From "the AI that listened to my loneliness" to "the AI I trust with my money," "the AI that watches my health," "the AI that decides alongside me." So this was never about capturing the loneliness market. It's about holding the first gateway through which humans cross into a real, companionate relationship with AI. Whoever holds that gateway becomes the default for every domain trust flows into next. And that's why the moat here isn't the model — it's the accumulated relationship itself. Memory and context are the switching cost. Own the companion's memory and identity, and you own the relationship. The next platform of the AI era won't be won on payments, or even on intelligence. It'll be won on trust. And trust is built only one way — through a relationship that remembers you.

103208

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Ryan Kim@0xryankim · 2026/06/14观点输出

RT @OpenRouter: Introducing the Fusion API, the smartest compound model in the market. Fusion achieves Fable-level intelligence at half th…

0178700

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Ryan Kim@0xryankim · 2026/06/10赛道影响

Spent a full day at SuperAI Singapore. ~11 sessions, from hyperscaler capex to AGI. The throughline was clearer than any single talk: the model layer is commoditizing, and value is migrating to deployment, context, and trust. A few things I can't stop thinking about. 1. Commoditization is real — but partial. Lower-tier, routine tasks collapse toward zero margin (the telco trap: mobile data grew 1,500–2,000x over 15 years while telco stocks went nowhere). High-value verticals don't. As basic video generation commoditized, BytePlus moved up to 4K/8K and real-voice inputs — and users pay more. The frame that stuck: scale problems (coding) get commoditized; scope problems (cancer research, materials, math — where the question keeps changing) are where durable value lives. 2. The agent infrastructure stack is forming fast. Agent-native search (Will Bryk @WilliamBryk, Exa: "search is infrastructure for AI"), agent governance and identity (Moe Abdula, Google Cloud: manage agents like employees — ID, registry, performance review; Google runs ~150k internal agents, ~100k built by non-engineers), trusted web data (Ariel Shulman, Bright Data), policy that lives outside the agent (Tencent), and agent payments/permissions (Abhi Tiwari, Stripe). Enterprise agent adoption went ~5% → ~40% in a year. One Fortune 500 projection: 100 agents per employee. 3. Context is the new moat — "harness is cheap, context is expensive." Oscar Hjelde of Norges Bank Investment Management — covering 7,000 companies with 700 people — said it best. Harnesses can be outsourced to the big labs; the expensive, slow-to-build asset is your own context: data formats, internal benchmarks, governance, documented process. Tencent's "durable assets that survive model swaps" is the same bet. 4. A verification economy is coming. The web now fights bots not by blocking them but by serving them plausible false data — a bot sees a $399 rate where you see $199, updates 5,000 records, sends 5,000 wrong emails. As fabrication gets trivially cheap, auditing and verification get expensive and valuable. 5. The subsidy era is ending, and compute is being financialized. $20 plans became $100–200; "unlimited" got capped. Token costs rose ~100x in a year, and every 100x drop in unit inference cost drives ~10,000x more consumption. The biggest idea of the day, from Evan Conrad (@evanjconrad, San Francisco Compute): a compute futures market — ~$600B spot today, a potential $6–8T derivatives market with CME-style contracts. "Gross Token Product" floated as a national metric; token export as a trade category. 6. Scarcity shifts from the average to the authentic. LLMs supply "the average" in infinite quantity, so the premium moves to what can't be downloaded — authenticity, individuality, physical presence, in-person community. Jevons explains both physical AI and the revival of offline, handmade experience at once. The one-line version, and the frame @balajis and @benedictevans closed the day on: AI lowers the floor, not the ceiling. More people can build — which also raises the noise floor. The edge moves to taste, judgment, distribution, trust, and pricing design. Full debrief, 10 sessions, no fluff: https://t.co/pYmHtVf7pL

315542

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Ryan Kim@0xryankim · 2026/06/09赛道影响

Over the past few months, I’ve spent more time refining my AI workflows than simply using AI tools. I use both OpenClaw and Hermes, each with 3-5 sub-agents underneath. Their memory systems are connected through Obsidian. I’ve also built a knowledge graph using Andrej Karpathy’s LLM Wiki as one of the foundations. A few months ago, I was mostly trying to “vibe code” websites, tools, and random prototypes. Now I’m spending more time defining reusable skills and workflows: repeatable loops that can run inside chat apps, remember context, coordinate agents, and produce outputs I can actually use again. The more I build this way, the more I realize that the frontier is not just the model. It is the harness around the model. The memory layer. The sub-agent structure. The workflow design. The feedback loop. The way intent becomes execution. But there is one area where I still struggle a lot: design. I don’t have great design taste. More importantly, I often don’t know how to describe what I want at a high enough resolution. When I try to create Instagram carousel posts, character images, or website UI/UX, the output often misses the mark. Sometimes the typography feels wrong. Sometimes the composition is off. Sometimes a character looks almost right, but not quite. Sometimes the fingers are wrong, the text is broken, or the whole thing feels like an AI-generated compromise. And then, once in a while, something works. When that happens, I feel an absurd amount of relief. I lock in the workflow, preserve the prompt structure, save the references, and try to turn that one good result into a repeatable system. This is slowly changing how I think about AI. I used to believe AI would flatten human capability. I thought developers and designers would become less important because everyone would have access to powerful tools. But the opposite seems to be happening. Great developers become even more powerful when they learn to build better harnesses. Great designers become even more powerful when they can translate taste, composition, motion, emotion, and intent into something models can execute. AI does not simply equalize people. It amplifies clarity. The more clearly you can imagine something, the more likely it becomes real. This is why I think the next wave of leverage will belong to individuals and small teams with unusually sharp intent. People who know exactly what they want to make, who they are making it for, and why it matters. A single person with a clear imagination and a strong AI harness may produce better results than a bureaucratic organization of 10,000 people with unclear intent. Capital also starts to move differently in this world. A pitch deck is slower than a working demo. A memo is slower than a usable product. A market thesis is slower than a loop people can touch, test, and share. As the cost of making things falls, investment decisions will accelerate. Money will flow faster toward visible, usable, emotionally resonant outputs. And this brings us back to taste. What are we trying to solve? Who are we trying to help? Who are we trying to make less lonely? Who are we trying to make more capable, more comfortable, more inspired, or more alive? AI may help with the “how.” But the “who” and the “why” still matter more than ever. That is why harnesses matter. They reduce the distance between intent and execution. They turn imagination into repeatable production. Models will keep improving, and they will absorb more of what we currently call workflow. But the harness layer will not disappear. It will move up. Nvidia can keep building better chips, but chips do not replace operating systems. Operating systems can keep improving, but they do not replace email, design tools, social networks, games, or enterprise workflows. Every layer creates the conditions for the next layer. The same will happen with AI. Models will become more powerful. But around them, thousands of specialized harnesses will emerge: for design, research, education, investing, healthcare, entertainment, operations, and personal productivity. Some will disappear quickly. Some will become platforms. A few will capture enormous value because they become the loop where intent, data, execution, and capital keep circulating. Today, everyone is looking at the models. But I think the real question is becoming: Who will build the harnesses that turn intelligence into compounding workflows? Because in the end, intelligence alone is not enough. You still need taste. You still need intent. You still need a loop. And you still need a reason to build.

7112643

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Ryan Kim@0xryankim · 2026/06/08观点输出

@mushumik_a 😇

000438

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Ryan Kim@0xryankim · 2026/06/08政策影响

Asia’s internet did not evolve like America’s. In the U.S., the internet economy was built on top of infrastructure that already existed: credit cards, mature banking rails, addresses, email, search engines, and independent software markets. Because those foundations were already in place, the efficient model was an ecosystem of specialized services connected through APIs. Asia was different. In China, Korea, and much of Southeast Asia, many pieces of digital infrastructure were either missing, inconvenient, or still immature when the consumer internet took off. So the most efficient model was not “many services connected through APIs.” It was one app that could solve everything. WeChat bundled messaging, payments, shopping, mobility, games, and government services into one interface. Korea had a different path, but a similar outcome. Naver became the portal that had to capture more and more services to build enterprise value. Kakao followed the same logic in mobile. Even though mobile was less portal-like than PC, the culture of super apps remained. That created convenience. But it also created closed ecosystems. Regulation played a role too. In the U.S., antitrust pressure made platforms more cautious about owning every layer directly. In Korea and China, the priority was often to grow first. After fierce competition, the winner captured the ecosystem. There was also a cultural layer. Asia has generally had less resistance to centralization when it delivers convenience. The “fast and convenient” mindset made dependency on a few dominant platforms feel natural. Each country ended up with its own super app. And each super app became its own closed world. But the AI era changes the logic again. Everything is being rebundled around agents. Users will no longer only navigate inside platforms. AI agents will search, choose, compare, call APIs, transact, and execute on behalf of users. That means the most important users of platform functions may no longer be humans clicking through apps. They may be agents calling services directly. In that world, open platforms have a structural advantage. If your service is easy for AI agents to access, understand, and execute through APIs, it can become one of the default choices in the agent economy. Once agents start choosing you, distribution can compound very quickly. But if a super app only connects its own internal services and stays closed to outside agents, it risks becoming invisible to the new execution layer of the internet. If innovation outside the platform moves faster than innovation inside it, isolation becomes dangerous. This is the context behind https://t.co/eNoato0mdk. Hashed started OBA, the Open Builders Alliance, because we believe Korea and Asia need more open APIs, more open-source resources, and more builder-accessible infrastructure. On May 30–31, we held OBA Weekendthon at Kakao AI Campus. 48 founders and more than 30 teams came together for two days and one night to build real products. Nexon opened up game IP and source code. LG U+ shared its own AI model. MyRealTrip, RocketPunch, Swing, CryptoQuant, GS Neotek, and others opened APIs, open-source resources, and technical assets. OpenAI joined as the main sponsor. This was not just another hackathon. It was a signal. If Asia wants to compete in the AI era, we cannot only build better apps inside closed gardens. We need to build an ecosystem where builders, startups, large companies, and AI agents can connect directly. Open up, and build together. That is the direction Hashed is backing.

022350

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Ryan Kim@0xryankim · 2026/06/08市场影响

I think the roots are old, but the modern form is new. Koreans historically had a lot of `heung` — expressive energy, emotion, rhythm, performance. But the modern turning point was when Korea started treating culture as a national industry, especially from the Kim Dae-jung era. That gave the old temperament an institutional base. Then freedom of expression turned it into brands, films, music, spaces, and content. So: old temperament, new system.

001140

AI：偏平台经营与生态扩张，强调交易平台和应用入口的长期位置。