John Yang (@johnjyang) header
John Yang (@johnjyang) profile picture

learning. prev cofounder @milliondotjs

San Francisco, CA
Joined November 2019
Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)Photo by John Yang (@johnjyang)
Hello,我是二虎🐯,也可以叫我 Will。 2021 年毕业后一直在创业公司做 SaaS 出海。做过 3 年产品经理,现在在 Manus 负责广告和 SEO,也兼职干点商业化产品经理的活。 3 年前立的 flag 还在:希望自己能在 AI 这件事上热爱 10 年。 重新介绍一下自己。先聊聊我现在最关心的 Agent、广告和 SEO,再讲这些想法是怎么一路长出来的。 01 / Agent:我的新职业,可能是制造员和驾驶者 我越来越相信:以后一个负责经营和增长的组织里,Agent 的数量会比人多。 我想象中的样子是:投放调价、SEO 维护、素材分析、数据监控、竞品追踪,每一件重复的事都有 Agent 在跑。人不多,但每个人身后都带着一队 Agent。 这几年,我已经习惯把竞品研究、公开数据收集、广告素材和 SEO 页面分析先交给自己的 Agent 跑一遍。以前只能看几个样本,现在能看更多。 按我的体感,以前做增长,花在创新上的时间大概只有 5%;现在倒过来了,95% 的时间都在创新。效果投放里的各种优化、反复改定价页这类工作,正在很快被自动化。 所以,我想象中自己的工作会变成两件事: Agent 制造员:把一件重复的增长工作拆清楚,做成能稳定干活的 Agent。 Agent 驾驶者:决定它往哪跑,看数据判断它跑得对不对,跑偏了自己接管。 这让我想起 3 年前给自己设想的能力模型: • 半吊子的工程师,不分前后端运维 • 出海增长专家,主要是 SEO / SEM • 不专精的产品舵手 • 合格的数据分析师 当时写“半吊子的工程师”“不专精的产品舵手”,多少有点心虚,觉得是自己没进过大厂、不够专精。 现在回头看,什么都会一点、能自己把一件事从头做完,可能正在变成 AI 时代每个人的标配。近一年,我也是第一次开始自己往线上提交代码。 那四个能力没有过时,只是换了个用法:工程和产品用来制造 Agent,增长经验和算账能力用来驾驶 Agent。但很多过去的打法,确实要重新学。 Agent 再多,人也不会没用。数据抓回来,不等于结论就是对的。问题怎么问、结果怎么核对、最后要不要做,还是得人来定。 这套能力模型,也正在改变我做广告和 SEO 的方式。 02 / 广告投放:我越来越觉得,它本质上是信号学 我理解的广告,有两套系统要打通: 系统 1:把产品能力翻译成市场容易看懂的广告素材,让用户明白它能帮自己做什么。 系统 2:把商业化结果翻译成广告系统容易理解的信号,保证信号的密集度和准确度,让系统知道什么样的用户值得继续找。 广告不能只看前端的点击和注册,还得看这批用户进产品后有没有留下来、有没有付费、多久能回本。 最早写这篇自我介绍时,还是“最近接手了相当多的 SEM 工作”。现在已经是 Google / Meta / TikTok 几个渠道一起投了。 目前正在跟团队一起打磨这两套系统。 03 / SEO:它是基础盘,但未必是以后让你赢的理由 SEO 一直是我比较擅长的方向,尤其是 SaaS 工具类产品。 我现在的看法是:它还是得做,是增长的基础盘。但大家找信息的习惯正在被 AI 改变,它以后大概率不再是让你赢的那个理由。 所以更适合交给一个人 + 一个 Agent 稳稳维护,把人省出来去做新东西。 这个想法,是从过去做 SEO 的经历里慢慢长出来的。 之前给 3 个网站做过程序化海量页面,大多拿到了很大的流量。但我一直不觉得,页面多、流量大,就等于增长成功。 在 Monica,接手时 SEO 月流量不到 10 万。技术基础有不少问题,做页面也很依赖多人协作。 后来跟团队一起,先修技术问题,再挑真正有需求的工具和页面,最后把页面生产改成低代码系统。 大概 8 个月,月流量最高做到了约 400 万。 到了 Manus,SEO 已经是一套不需要我天天盯着的维护体系。现在 Ahrefs 上的自然流量估算约 100 万,由一个人 + 一个 Agent 打理。 这是我前面说的“每个人身后都有 Agent”的一个小例子。我们团队正在往这个方向猛冲。 这里面有很多关于地区、设备和具体需求的 insight,欢迎找我讨论。 04 / 商业化:不管从哪个渠道来,最后都要回到经营的账本 无论广告还是 SEO,找来用户都只是开始。做过商业化之后,我看增长的方式也变了。 在 Monica 的 15 个月里,除了流量,也跟团队一起做了 onboarding、定价、退款和渠道。同一阶段,日收入翻了好几倍。还做过 Affiliate 渠道,做到了整体收入的 10%–15%。 再看一个增长项目,会多问几句:找来的是什么用户?为什么愿意付费?付费后有没有继续用?获客的钱多久能收回来? 也停掉过看起来很热闹、但 ROI 算不过来的项目。 现在还是喜欢搞野路子,但最后要能回到经营的账本上。 05 / 产品:能把人找来,不等于能让人留下来 账能算过来,也不代表产品就做好了。 快速做一个满足基础需求的 MVP,再通过用户反馈慢慢把它做大,还是我比较擅长的事情。 短板也还在:没体验过花一个月只做一个功能版本,对单个需求想得不够深。 这几年一直在对用户体验要求很高的团队里,每天都在补这一块。能把人找来是一件事,让人真的用起来、持续用下去,是另一件事。还在练。 而且我觉得这块短板以后更值得补:功能会越来越容易被做出来,用户愿意留下来,更多是因为喜欢这个产品,也信得过背后的团队。 06 / 往回看:这些想法,是从野路子里长出来的 简单过一下经历: • 大学时,在一家互联网小厂做过半年产品实习。 • 2021 年毕业后加入广大大(SocialPeta),做产品经理,从这里开始做出海。 • 2023 年底加入 Monica,继续做产品经理。 • 2024 年上半年转去负责产品自然流量增长,主要是 SEO;年中开始接触海外广告投放。 • Manus 发布后,开始负责它的广告(Google / Meta / TikTok)和 SEO,也做一些商业化产品工作。 刚进入出海时,完全是野路子玩家。不知道怎么系统性地做增长,也没搭好增长的数据体系。 但跟团队一起,依靠插件市场内的自然流量,做过超过 40 万用户的产品矩阵,没花一分钱买量,也没借助 SEO、SEM 或红人。 到现在还是很佩服当时在广大大的团队,硬生生挖出了一条行之有效的野路子。 这一路都是创业公司,做产品和找用户很难分开。我慢慢养成了一个习惯:从第一天就想用户从哪来、为什么愿意付费、这件事能不能持续经营,把自己当成产品的 CEO。 —— 从这些经历走到现在,我还有一个越来越强的感受:增长正在变成一个信任问题。 Agent 能复制,打法也能复制,但用户信不信一个团队,没法直接复制。 我觉得以后,人会花更多时间在“被信任”这件事上:把真实的经历、判断,甚至踩过的坑公开讲出来。这也是我开始在社交媒体上认真写东西的原因。 我也先拿自己试: Will + Agent = Wigent。 想把自己的经验、判断和工作方式,慢慢沉淀成一个能跟着我一起长的个人知识分身。我注册了 wigent.ai 这个域名,网站先搭好了,分身还在慢慢造。 会在这里慢慢写下我在出海 AI 里的实践:Agent、广告投放、SEO、商业化经营,还有产品思考。 还是喜欢跟真正动手做事的人交朋友。我超级 E 人,不用担心和我聊天会尴尬。 如果你也在做 AI / 出海 SaaS / 增长,或者也在用 Agent 做增长和经营,特别想认识你。
173
62
33
615
599,862
John Yang retweeted
We asked two isolated copies of the same AI to exchange a secret word with only a robot and a camera. They succeeded by using air-writing, Morse code, flag semaphore and a 5-bit binary alphabet, without agreeing on a protocol beforehand.
30
60
26
475
85,161
great thread here
Replying to @shdu11546816
Modern ad systems have always been run on machine learning algos, but in the past 5 years the trend has accelerated with AI now doing all the targeting for you. In the past you would hire a media buyer to tweak things. You could adjust the demographic you want to target, the exact locations, the time of day, etc... You used to have hundreds of these levers you could a/b test and optimize. In the past 5 years this has shifted to AI, so you don't need a media buyer anymore. You just set a conversion target and the AI targeting systems will grind on it. When you're selling something cheap like yoga pants, all you have to do is create good ads and set your conversion target to gross margin. The AI will find you sales, and it will even optimize for the exact customer profile that you'll earn the most margin on. However, when you're selling $20,000 items, you'll never hit enough of volume to properly train the ads. You need at least 50 conversions per week so unless you're doing $1 mil revenue weekly, you can't target margin. Instead you target something higher up the funnel like a phone call or a lead form submission. You can even put additional qualifiers- such as the phone call needs to last at least 5 min. What we did at first was to ask a follow up question after a customer submitted a quote request on our website-: How were you planning to pay? Option 1- Pay cash Option 2- Finance with FICO over 650 Option 3- Finance with FICO under 650. And then we simply targeted option 1 and option 2 as the reward function. This filtered out a bunch of the poors, but we still captured the lead even if they refused to answer. If they didn't answer the question, we would just pass them along to their quote automatically after 6 seconds, plus we had automated text and email notifications sent in case they closed out before they seen it. This had amazing results, so I realized that the biggest lever I could pull is setting a better target for the AI to grind on.
195
John Yang retweeted
I'm hiring researchers who want this to be their life's best work. You'll see the future of AI before most people do, inform the public and governments, and get to play with more kinds of robots than almost anywhere else.
We're hiring Members of Technical Staff. You'll measure what AI can do with real robots and show the public and governments how fast it's improving. We're a Public Benefit Corporation. We want researchers who understand the stakes of powerful AI. No robotics experience needed.
16
19
4
240
15,066
let’s go!! “hey instinct” goes hard
Instinct on the Action Button You can now make Instinct your iPhone’s Action Button. Press and hold, say what you need, and Instinct will take it from there, without you needing to leave the app you’re on. To set it up: 1. Go to app.instinct.com/voice-short… and tap Add Shortcut 2. Go to Settings → Action Button → Shortcut → Instinct. This is available to all Instinct users on an iPhone 15 Pro and newer.
3
425
John Yang retweeted
struggle together i learned programming and design by trying, failing, and trying again. when i hit a wall, i figured out how to climb it. read something i barely understood. took things apart. asked someone for help. started over. when i began my career, i kept doing things i had never done before. trying different ways to make something work. when i stopped learning, i moved on to the next struggle. many things i made weren’t very good. but after enough rounds, something stuck. what once took so much effort became part of how i thought. i could feel something was wrong before i knew how to explain it. it became like water. i think about what went into that. the bug i couldn’t fix. the design that looked right until someone used it. the conversation that made me realize my flaws. now we can ask a machine. i’ve made things with these tools that i might never have made otherwise. i don’t miss every tedious task. but i wonder what happens when we remove the difficulty before we understand what it was teaching us. we ask. something appears. it works well enough. we move on. we learn how to get results without necessarily understanding how it works. and because the result looks complete, it becomes easy to miss the questions we never learned to ask. there is another side to this. working with people is uncomfortable. everyone wants something different. everyone has a history you don’t know. someone cares about a detail you think is unimportant. someone pushes back when you want to move. someone needs time when you feel certain. it takes a while to find each other. but sometimes, after struggling together, you become a band. you know who notices what you miss. who can pull an idea apart without killing it. when to step in, when to slow down, when to let someone else carry the rhythm. you share something unfinished. admit you don’t know. hurt each other, apologize, try again. eventually you make something none of you could have made alone. some of my closest friendships came from working this way. smart, caring humans who saw me confused, wrong, tired, excited. people i could be vulnerable with. people whose ideas changed mine. i worry we are making it easier to avoid that beginning. an agent is always there. patient. ready to keep going. talking to it can feel easier than reaching toward someone who might misunderstand you, disagree, or have somewhere else to be. sometimes that ease helps. but if we keep choosing it, where do the friendships begin? i picture each of us with our own agents. our own explanations. our own little worlds. making more things, with fewer reasons to leave the screen or work through a difference with another person. and these worlds aren’t entirely ours. they run on machines in server farms, owned by companies, shaped by interests we may not see. we bring them our questions, our work, our loneliness. they become part of how we understand what is real. is that freedom? or a comfortable kind of confinement, where everything responds to us but fewer things reach us? i don’t think pain makes something good. there are struggles nobody should have to endure. but i don’t want to lose the difficulty of learning something unfamiliar, or getting to know someone who isn’t like me. maybe we need tools that help us spend more time with the world, instead of giving us fewer reasons to enter it. tools that make it easier to find people, share something unfinished, and work through what we couldn’t figure out alone. sometimes i want to close the laptop. sit with people i love. play something we haven’t figured out yet. miss a note, lose the rhythm, listen until we find each other again. like a jazz band. everyone bringing something different. nobody knowing where we’ll end up. i don’t want a life without struggle. i want people to struggle with. something we can feel, change, and make our own. music we couldn’t have made alone, instead of the silent static of the server farm.
75
125
40
1,560
75,158
Lots of Security incidents today, here's some practical tips for those who just wanna protect yourself: 1) UPDATE EVERYTHING - main targets are: a) Browsers - Chrome, Safari, Firefox b) OS: : Settings -> General -> software update; Linux: sudo apt update && sudo apt upgrade; Windows: I think there's a button. Do this for any remote home servers & Openclaw/hermes boxes c) Phones: Update it all d) Desktop apps (Slack, Cursor, Discord, etc) 2) Limit web browsing this week: Stay within walled gardens (twitter, youtube, etc); don't click random links this week. Just wait until the patches go through. Do step 1 again after a week. 3) Avoid untrusted Wifi networks until your devices are updated 4) Stay vigilant about phishing emails offering "certificate fixes" or fraud updates 5) Have an iPhone 11-16? TURN BLUETOOTH OFF for the time being until apple patches come through. 6) Code/Dev: UPDATE ALL YOUR REPOS - @charliermarsh has already updated uv 0.13.4 which reacts to the eMudra incident for uv (Python), Typesript according to Claude has equivelant. Tell your agents to go and update all your active repos to update package manager across uv, npm, bun, cargo, etc etc What advice did I miss?
101
1,898
102
10,137
849,775
duet is a good name
1
208
"The problem is the shape of a moat is often conflated with the mechanism" (a very personal mistake!)
I think I've gotten to the bottom of the issue with the "network effects" discourse. The problem is the shape of a moat is often conflated with the mechanism. And there is remarkably little consensus on what a network effect is, or a flywheel. Here's a very simple, very rough vocabulary that I think works with current working definitions. Seriously, learn these few rules so your boardroom conversation about defensibility isn't a waste of time: Shape of Moat: This is what happens to the advantage as the company scales Static : Doesn't change Sublinear : Advantage compounds but at a diminishing rate as the company scales (this is not a long term durable moat against challengers) Linear: advantage is fixed relative to a competitor as company scales Superlinear : Advantage compounds at an increasing rate as the company scale (this is the white wale everyone pretends they have, but almost no one does) Mechanism of Moat: The property, business process that produces the advantage. Network : more users creates more interaction between users. Hopefully superlinear. Scale : greater volume produces better unit economics. normally sublinear Data-driven : new users means more data for new users (or to get new users). normally sublinear Platform : growth on one side increases value to the other. sublinear to superlinear depending on size Distribution : scale improves reach or channel economics. normally sublinear ... etc. Flywheels: Flywheels are also conflated with the above. The most useful working definition of a flywheel is the aggregate advantage when multiple mechanisms are self reinforcing, something in a chain (A->B->C). Imagine multiple forces pushing a flywheel to make it go faster. The hope is this complex interaction produces a superlinear advantage. They are incredibly hard to reason about. And most discussion around them are based on vibes.
228
John Yang retweeted
🏀 We simulated tonight's OKC @ Spurs 5,000 times Every possession is built from real 2025-26 play-by-play: where each player shoots from, how often he makes it, who he passes to, and who's guarding him. Here's the most typical game ⏬
85
228
156
9,464
5,761,323
John Yang retweeted
one day you wrote console.log("here") for the last time and didn't even know it
213
1,728
268
17,884
607,007
John Yang retweeted
Anthropic released their 2021 research roadmap earlier this year (because of copyright litigation); I think some people might find it interesting: i.sm4.ca/concise-research-ro…
9
99
28
868
319,165
John Yang retweeted
What if the same game could feel native on every machine you love? PSP. PS Vita. Nintendo 3DS. iPod touch. Old Android. PocketJS is going 3D. Introducing Pocket3D.
24
150
14
1,581
62,825
John Yang retweeted
Proximal builds infrastructure that enables AI to improve from real-world experience. In the past year, we have grown to more than $200M in annualized revenue helping frontier labs and enterprises improve coding agents. Now, we are expanding beyond software engineering. Recent progress in mathematics demonstrates what happens when models are trained in domains with massive amounts of public data and easy verifiability, making it easy to find weaknesses, generate training tasks, and iterate quickly. We build infrastructure that enables this feedback loop in other domains: we are excited about a world in which AI systems design targeted drugs, accelerate chip design, and rewrite legacy software that still runs hospitals, power grids, governments, and other critical infrastructure. Since starting, we have assembled a small team of researchers and engineers and built world-class infrastructure for post-training and synthetic data research to overcome the challenges of scaling manual data curation. Our team comes from Cursor, Google DeepMind, Meta Superintelligence, Prime Intellect, Citadel, and Jane Street. More than half are former founders. We’re fortunate to be backed by investors who share that vision, including @generalcatalyst who led our $15M Seed Round at a $300M valuation, as well as @chemistry, @svangel, @dvlamoen and individuals like @LiamFedus, @kevinweil, and @bernhardsson
113
42
53
388
262,906
My conversation with Noah Shinn (@noahrshinn), founder of Instinct. Noah is building a personal AI assistant. It's still invite only, has spent nothing on marketing, and is growing roughly 10% A DAY. This is his first long conversation about the company. We discuss: - Why Instinct doesn't have an app - Buying compute months ahead of exponential demand - How users learn to trust it with a credit card - Safety and security - Agents coordinating with other people's agents - Instinct's business model - Apps built on consumer inertia - and more Enjoy! Timestamps: 0:00 Intro 4:11 What people are using AI agents for 15:07 Rethinking travel, reservations, and the internet 22:43 Trust, privacy, and personal data 27:50 The business model behind Instinct 38:04 How existing businesses will adapt 47:55 Designing a personal assistant people love 53:15 Growth, compute, and competing with Big Tech 1:11:44 What’s next for Instinct and personal AI
150
226
213
2,365
1,826,286
John Yang retweeted
We (@PantheonInc) have been working on a robotics data quality pipeline that uncovered a series of major problems in public robotics datasets, especially for world modeling. To improve the quality of data available to open-source robotics, we're publishing annotations for four of the most popular datasets. Some examples of issues, and our report 🧵
156
113
49
675
183,029
John Yang retweeted
On what made pg great, c. 2009-2016: - every person tends toward particular kinds of ideas. Someone might tilt toward ideas that are entertaining, or elegant, or useful. Pg tended toward ideas that are surprising. He would look for surprising observations, generate surprising hypotheses, select surprising inferences. He'd never put it this way, but he seemed to find a lot of alpha in surprise - pg was averse to abstractions. He almost always preferred a concrete way of doing things. Investors might ask about products or markets or business models. Pg would ask "what are you building?", "who is it for?", "how much will you charge?" It's the same questions, but he'd be as concrete and plain as possible - being a hacker, he made things. His ideas were always practical and embodied in books, essays, software, YC, etc. They were meant to be used. It's very difficult to imagine pg writing a treatise on ethics or the nature of time or smtg - this ethos is very much from Unix/C school of thought. Ironically, Pg was a Common Lisp hacker who embodied the worse-is-better philosophy. He would find the minimal possible thing that solves the problem, and it would be unmistakingly pg-esque. (If you haven't read The Rise of Worse is Better essay you should. Will the thesis be relitigated as AIs get much better?) - that said, I'm not entirely sure he himself cared about making something people want. His output seems like an exhaust product of his curiosity that turned out to be useful by accident. Then again, he just couldn't help himself but stumble upon such accidents. Perhaps his way of thinking produces things people want, so for pg the score takes care of itself. The rest of us have to work at it deliberately - he delighted in his ideas. The word "arrogance" has negative connotations, but pg was arrogant in the best possible way. If he thought an idea or an argument is good, it filled him with delight. It would then go into canon and he had no interest in revisiting it. There was no moving him once that happened - the slop version of this is he "trusted his process" but again he would never put it this way. I don't think he was capable of navel gazing. It's impossible to imagine pg considering whether he'd be a favorite of fortune, or what posterity will think of him, or what he thinks of himself. In pg, the brain cycles most people spend on this would go to object level ideas - I never saw the slightest sign he cares about what people think of him, nor did I ever see him try to impress anyone. I don't think he suppressed this impulse by force of will; as best as I could tell he has no such impulse. There is a Conversations with Tyler episode where pg keeps answering "I don't know" to a series of erudite questions without any embarrassment whatsoever. This is very pg - he was interested in you if he thought you're smart and ambitious. If you did something to disabuse him of this, he'd lose interest completely. But this was never permanent; if you showed evidence of being smart and ambitious again his interest would rekindle. (You only got so many strikes, though) This brought out greatness in ppl because they would try very hard to impress him - at the same time he genuinely cared about people's wellbeing. There are many stories where ppl had trouble in their personal life and he'd extend himself to help. I experienced some of this myself. Some values were much more important to him than ideas or ambition - Thomas Jefferson once wrote "I am of a sect by myself", which I think sums up pg's character really well. If you read enough founding-father bios, you start noticing that most Silicon Valley figures resemble one founder more than the others. Imo pg is very much Jeffersonian
7
11
4
307
18,878
How a robot arm is controlled, explained for ML people new to robot learning. First of the explainers from my own speedrun. Next one will be on ACT. Follow me to catch it.
19
128
19
982
338,337
John Yang retweeted
Introducing our Trusted Person network powered by our new Instinct-to-Instinct communication protocol. Your Instinct can now talk to other Instincts to coordinate plans on your behalf. Getting people together often involves a lot of back-and-forth: finding a time, working out the details, and following up when plans change. Instincts can now work with each other to help handle that coordination. For example: - Coordinating plans with your spouse - Planning a weekend trip with friends, including dates and arrival times - Scheduling time with colleagues and other collaborators - Hosting events with 10s or 100s of people - Organizing recurring plans like tennis, book clubs, or weekly dinners Disclaimer: Your Instinct will only be able to communicate with those in your Trusted Person network. This is the group of people who are closest to you: - Spouse - Older children and parents - Colleagues and trusted professional relationships - Small local businesses The Trusted Person network is currently available to our early access group and will be slowly rolled out across the user base over the next 24-48 hours.
208
99
230
2,469
835,674
John Yang retweeted
PocketJS can run on wildly different hardware because very little UI is delegated to the OS. Even the keyboard is rendered by ourselves. Like on this $10 iPod touch 4. Can you tell?
4
5
1
152
5,909