你大概见过这几种程序员。
有人发了一段用 AI 写出来的应用，评论区第一条是："一股 AI 味。"点赞最多。
有人刚被裁，转了一篇《AI 就是资本的又一场骗局》，配文："早就看穿了。"
还有一种最常见，也最让人心疼：下班地铁上刷到"普通人如何用 AI 做副业"，顺手收藏；周末买了一门 Agent 课，又买了一门提示词课；收藏夹里躺着几十个教程，真正打开过的没几个。每天都觉得自己又落后了，每天都在"准备开始"。
先交代立场，免得你以为我是来卖课的：我自己就有一堆"活着、能用、没赚钱"的项目——跑在免费套餐上，没人付过钱。所以这篇不是成功学，是把所有数字现查一遍之后的判断：在当下这个节点，普通开发者最优的解法，是用 AI 拿到商业结果，完成自己的原始资本积累。不是给别人的作品打分，不是被焦虑推着从 a 学到 n，然后在收藏夹里竹篮打水。
一、红利不是风口，是一个时间差
很多人把红利理解成"风口"，好像站上去就能飞。这个理解害人：它让人要么盲目冲进去，要么因为看到有人摔下来就断定风口是假的。
用商业语言说，红利是一个时间差：生产成本下降的速度，快过市场价格下降的速度。 在这段时间里，客户还按旧成本给结果定价，而你已经在用新成本生产。中间那块差额，就是利润。
毛利窗口 = 客户按旧成本锚定的价格 − 你的新生产成本
这一轮成本下降有多快？a16z 统计过：同等能力的大模型，推理成本大约每年降 10 倍。2021 年 11 月，GPT-3 水平的能力（MMLU 42 分）要 60 美元一百万 token；到 2024 年底，同等水平只要 0.06 美元，三年降了 1000 倍。Epoch AI 后来的测量更激进：只看 2024 年 1 月之后的数据，中位年降幅从 50 倍跳到了 200 倍。
而价格呢？客户心里的价格锚，还停在"要请人做"的年代。价格是由认知决定的，认知的迭代速度远远跟不上 token 价格。
对写代码的人，这个时间差还有一个更直观的版本：五年前做同样一个产品，你需要一个前端、一个后端、一个设计师，几个月工资；现在一个人加一套 AI 工具，年成本能压到几千块。单次尝试的成本降了几个数量级——这是这轮红利里对程序员最友好的部分：你不需要赌上全部身家，你只需要多试几次。
a16z 的 LLMflation 原文，结论就一句——同等能力，每年便宜 10 倍。
二、窗口是真的：别看情绪，看谁在付钱
判断一个东西是不是骗局，最简单的办法是看有没有人真金白银付钱。情绪可以造假，付款记录很难。
Stripe 分析了平台上收入最高的 100 家 AI 公司：达到 100 万美元年化收入的中位时间是 11.5 个月，比 SaaS 黄金年代最快的那批公司还快 4 个月；达到 500 万美元年化收入用了 24 个月，而 2018 年头部 SaaS 公司用了 37 个月。个案更夸张：Lovable 3 个月做到 1700 万美元 ARR，Bolt 2 个月做到 2000 万美元 ARR，Cursor 三年做到 1 亿美元 ARR。
一人公司的数字也不逊色。Danny Postma 一个人做 AI 职业照 HeadshotPro，年经常性收入约 360 万美元；Maor Shlomo 的 Base44 上线六个月被 Wix 以 8000 万美元现金收购；Pieter Levels 没有员工、没有融资，一年收入 300 万美元以上。
国内一样有样本。Manus 2025 年 3 月上线，年底年化收入突破 1 亿美元；2025 年 12 月 Meta 宣布以 20 亿美元上下的价格收购，2026 年 4 月 27 日被国家发改委叫停。可灵的节奏更戏剧：2026 年 3 月 25 日，OpenAI 宣布关停 Sora，同一天快手发布财报——可灵 2025 年第四季度收入 3.4 亿元，12 月单月收入突破 2000 万美元，年化收入运行率 2.4 亿美元，全年收入约 10.4 亿元。同一个赛道，同一天，一家在退场，一家在收钱。
这些都是公司，离个人远。但注意一个共同点：它们赚的都是"旧价格锚"和"新生产成本"之间的差价。 HeadshotPro 和 2023 年爆过的妙鸭相机做的是同一件事，区别在于它卖给职场人做 LinkedIn 头像，场景更窄、更刚需，客户愿意为"看起来专业"多付钱。
Stripe 官方博客原话——AI 公司正在改写增长规则。
证券时报的报道标题就是全部剧情——Sora 退场同日，可灵立下收入翻倍军令状。
三、先泼冷水：一半以上的产品没赚到钱
看到这里如果你开始热血，该泼冷水了。上面全是幸存者，而幸存者偏差是这个行业最大的骗局，比"AI 是骗局"这句话本身还大。
更诚实的数字来自 TrustMRR。这个平台直接对接 Stripe 等支付渠道，展示的是经过验证的真实收入。2026 年 9 月，有研究者抓取了平台上全部 10,150 个产品：
两个细节值得盯着看。第一，AI 产品和非 AI 产品有收入的概率几乎一样（49.3% 对 49.0%），月收入中位数一个 \(161、一个\)175——AI 改变的是做产品的成本，没有改变"有没有人买"这个基本问题。 第二，另一个数据源 8 月抓的 8,817 个产品里，开发者工具类是零收入重灾区：69% 一分钱没收到。做工具的人最多，收到钱的人最少。
国内是同一张表的放大版。AI 漫剧 2025 年市场规模 168 亿元，听着是风口；但 2026 年上半年抖音新上线 22.19 万部，98.7% 没到盈亏线。成本降了九成，门槛降了，所有人就都进来了。
妙鸭相机是窗口关闭的标本。2023 年 7 月，它靠"9.9 元实现写真自由"一夜爆火，高峰期 4000 多人排队、等十几个小时，日活突破 60 万；然后同类产品一拥而上，用户的价格锚从"照相馆几百块"变成"AI 写真不就几块钱"。团队 2025 年 9 月底解散，从爆款到解散不到两年。窗口是真的，而且关得比你想象的快。 那年夏天在朋友圈嘲笑"AI 写真千人一面"的人，一分钱没拿到。
Tech Startups 对 TrustMRR 全量数据的分析，标题就是答案。
央广网：168 亿元的市场，和每分钟 500 元的成本，写在同一篇报道里。
重庆日报转载红星资本局：团队 2025 年 9 月底解散，只维持最低运营。
四、为什么我们宁愿囤课，也不下场
窗口是真的，失败率也是真的。理性动作应该是增加尝试次数、压低单次成本。但大多数人选的是另一条路：评论，和焦虑地学习。
评论的成本是零。说一句"AI 味太浓"不需要任何技能，还能立刻获得一种审美上的优越感，点赞即时到账。下场的成本是真实的：要花时间，要暴露作品，要面对大概率的失败和别人的嘲笑。在"零成本、即时回报"和"有成本、延迟回报、大概率失败"之间，惰性会替你做选择。
"AI 味"这句话作为审美判断完全成立，我自己也烦那种一眼生成的腔调。但它作为商业判断几乎没有意义：客户付钱买的是结果——问题解决没有，时间省下没有，钱赚回来没有。批评"AI 味"的人在评估作品，付钱的人在评估结果，两拨人根本不在一个市场里。 而且"AI 味"是一个工程问题，模型每一代都在变好，靠它守的护城河，每年都在变浅。
比评论更值得聊的，是焦虑地学习。这种行为经济学有经典样本：DellaVigna 和 Malmendier 追踪了美国三家健身房 7,752 名会员三年，选月费 70 美元以上的会员，平均每月只去 4.3 次，摊下来每次 17 美元以上——而他们本可以买 10 次卡，每次 10 美元。平均每人因此多花约 600 美元。他们高估了未来的自己。
买课是同一种心理：付款那一刻，你买到的不是能力，是"我已经在努力了"的安心感。MOOC 的数据更直白：Reich 和 Ruipérez-Valiente 2019 年发表在《Science》的研究，分析了 MIT 和哈佛在 edX 上的 565 门课、1267 万次注册，拿到证书的只有 3.13%；而付费学员的完课率是 44% 到 56%。差距不在课程质量，在有没有目标、有没有代价。
再算一笔时间账。《中国网络视听发展研究报告（2025）》显示，短视频人均单日使用时长 156 分钟，一年约 949 小时，差不多是半个全职工作的年工时。你说没时间做副业，时间都在，只是用在了刷"普通人如何靠 AI 月入过万"的视频上。
对程序员还有更扎心的一条：这个时代的工具每隔几个月就换一轮。 你花三个月学完的框架，可能学完那天就过时了。没有客户需求牵引的学习，等于在一条加速向前的传送带上倒着跑。真正好用的学法是按需学：客户要一个功能，我去学怎么做；产品卡在分发，我去学怎么投放。每一次学习，都绑定一个具体的、有人付钱的问题。
顺带拆掉两种最常见的场边言论。
"AI 抢走了我的工作。"这件事是真的，不该轻描淡写。斯坦福 Brynjolfsson 团队 2026 年 8 月更新的研究：AI 高暴露职业里，22 到 25 岁年轻人的就业比低暴露职业的同龄人低 19%，一年前这个差距还是 13%。但同一份研究还说：整体经济没有出现大面积失业，有经验的人没有出现类似缺口——冲击集中在依赖"可编码知识"的入门岗位，而且主要通过少招人发生，不是裁人。翻译成商业逻辑：AI 压低的是"按岗位出卖标准化劳动"的价格，判断力和对结果负责的能力，价格没有被压低。 理性的应对，是从"被定价的劳动力"挪到"给结果定价的人"。
"AI 是骗局。"截至 2025 年 12 月，中国生成式 AI 用户 6.02 亿，普及率 42.8%，一年增长 141.7%。估值可以有泡沫，很多公司会死，很多叙事会破灭，这些我都同意。但泡沫和骗局是两回事：2000 年互联网泡沫破裂，死了一大批公司，亚马逊没有消失。把泡沫说成骗局，逻辑上不成立，行动上的效果只有一个：给自己不下场找一个体面的理由。
Pew 今年 6 月的调查给这节收了个尾：52% 的美国成年人表示对 AI"担忧多于兴奋"，2021 年这个比例是 37%；而在 AI 专家那里，47% 的人"兴奋多于担忧"，普通公众只有 11%。越靠近这项技术、越在用它干活的人越乐观，越远的人越焦虑。焦虑和距离相关，和真相关系不大。
斯坦福数字经济实验室 8 月更新的论文，副标题问号前面是事实。
Pew 6 月调查的年龄分组数据——五年前最乐观的 18-29 岁群体，如今担忧比例最高。
五、不下场的人，成功率是精确的零
把前面的数据放进一个简单模型。假设每做一次产品尝试，做到年化 10 万美元收入的概率是 p。按 TrustMRR 的数据，p 大约是 2.4%——这已经偏乐观，愿意公开收入的人本身就偏乐观。尝试 n 次，至少成功一次的概率是：
P（至少成功一次）= 1 − (1 − p)^n，p ≈ 2.4%
这张表里最重要的数字，是第一行的 0%。不管你学了多少课、收藏了多少教程、说了多少句"AI 味太浓"，在这个公式里，n 都等于 0。做的人面对一个很低但不为零、而且随经验上升的概率；不做的人面对一个确定的数字：零。
那 n 要怎么加上去？看单次成本。我自己那几个没赚到钱的项目，全部基础设施跑在免费套餐上，固定支出就是每月一两百块的工具订阅。放在五年前，同样一个产品至少需要一个前端、一个后端、一个设计师，几个月的工资。
这也是我说"原始资本积累"是当下最优解的原因。原始资本不只是钱，它是第一批客户、第一份能拿出手的案例、对一个细分市场的真实理解，以及"我做成过一次"的手感。更重要的是，钱能买来时间：有了第一笔原始资本，你下一次选择做什么，可以不被焦虑和房租推着走。在变化这么快的时期，能从容做选择，本身就是最稀缺的资产。
六、开发者下场的最小动作
把上面所有数据压缩成一句可执行的话：选一个有人愿意付钱的具体问题，用最低的成本做出来，第一天就开始收费。
为什么第一天就收费？因为收费是唯一诚实的验证方式。点赞、收藏、"这个好棒"都不算数，只有付款算数。公式里 p 的提升，几乎全部来自对"什么东西有人买"的判断变准，而这种判断只能从真实的付款和拒绝里学到。
为什么挑具体的问题？因为宽泛的"AI 工具"没有人买，具体的"帮我省掉一次影棚拍摄"有人买。HeadshotPro 只做职场头像，不做写真、不做修图、不做证件照。
为什么用最低的成本？因为你大概率会失败，而且不止一次。成本压得越低，你能负担的 n 就越大。
动手前有三个检查，帮你少掷几次注定输的骰子：
- 价格锚检查 客户心里的价格，还停在旧成本吗？找那些"过去必须请人做、现在 AI 能做八成"的事：影楼写真、外包动画、产品详情页、行业报告、企业内部小工具。如果客户已经默认"这东西 AI 几块钱就能弄"，窗口就关了，别进。
- 复购检查 客户会不会付第二次钱？妙鸭栽在一次性付费上，9.9 元做完写真，用户就没有理由再回来，增长全靠拉新。能按月收费的，优先于按次收费的。
- 拥挤度检查 你有没有别人一时进不来的东西？AI 漫剧成本降了九成，结果半年上线 22 万部。工具降低的门槛，对所有人都降低了。真正能让你在拥挤里活下来的，往往不是 AI，而是 AI 之外的东西：你在某个行业干过五年的经验，你手上的二十个客户，你比别人更懂的那个细分人群。AI 是杠杆，杠杆要撬在你自己的支点上。
至于从 a 到 n 的学习清单，删掉。留下当前这个产品真正需要的那一项，学完就用。剩下的，等客户逼你去学的时候再说。
七、三个反驳，我先回
"你这就是幸存者偏差。"对，所以第三节放了那张表：26.1% 的产品零收入，做到年入 10 万美元的只有 2.4%。我从没说下场就能成，我说的是不下场一定不成。这两句话的区别，就是这篇文章的全部。
"说到底你也在贩卖焦虑。"这个质疑我认真想过。区别在于：贩卖焦虑的人最后会给你一个付款链接，让你买他的解药；这篇文章结尾没有付款链接，我给的唯一建议是去找一个愿意付你钱的人。如果你读完更焦虑了，关掉它，去做一件能收到 9.9 元的小事，焦虑比读十篇文章消得更快。
"不是每个人都适合做生意。"这条我承认，我回答不了。有人有房贷、有孩子，没有任何试错的余量；有人天生不擅长和人打交道，让他去收第一笔钱比让他加班还难受。对他们来说，"下场"的成本比我算的高得多。这个问题我放在最后，留给你。
最后
每个时代都有红利。PC 时代有，互联网时代有，移动互联网时代有，现在轮到 AI。每一次，红利都不是全民暴富，它只留给那些已经准备好、并且真的动手的人。窗口开着的时候，大多数人在评论，在焦虑，在学习；窗口关上以后，大多数人开始说"早知道就好了"。
你可以嘲笑那些动手的人，嘲笑他们的作品有 AI 味，嘲笑他们赚的是小钱。但别忘了第五节那个公式：他们每试一次，n 就加一；你每嘲笑一次，n 还是零。 你错过的不是几次暴富，是一条本来可以一次次试、慢慢走通的路。
文中所有数据的一手来源（a16z 原文、Stripe 报告、TrustMRR 分析、斯坦福论文、快手财报报道）我都整理好了，后台回我「hongli」就行。
你现在是在场上还是场边？最近一次靠自己的产品收到钱是什么时候，评论区留言告诉我。 下一期我挑一个被点名最多的场景，从头拆怎么做。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You've probably met some of these programmers.
Someone sent an application written in AI, the first in the comment area was: "A taste of AI." "Most praise.
Someone just got laid off and turned on "AI is another scam of capital," with the text: "I've read it."
And one of the most common and most distressing things: it was painted on the off-duty subway, "How ordinary people use AI as a side business", and it was a collection; it was bought on weekends in Agent, and it was bought in prompt; and it was in a collection of dozens of courses that were actually opened. Every day I feel like I'm falling behind again. Every day I'm ready to start.
Let's start with the position, so you don't think I'm here to sell classes: I've got a lot of "living, working, not making money" projects -- running on a free set, nobody ever paid for it. So this isn't a success story, it's a judgment after all the numbers are checked: At this current node, the best solution for ordinary developers is to use AI to get business results and complete their original capital accumulation. It's not a score for someone else's work, it's not an anxiety drive to learn from a n and then water in a collection basket.
I. A dividend, not a windfall, is a time Bad
A lot of people understand the dividend as "the wind" as if they could fly. This understanding harms people: it allows people to either go blindly in there or to conclude that the wind is fake because someone falls.
In commercial terms, dividends are a time lag: production costs are falling faster than market prices. During this time, the client also priced the results at the old cost, which you are already producing at the new cost. The middle difference is profit.
Māori window = customer price at old cost anchor - your new production cost
How fast is the cost of this round going down? a16z Counted: LLM with equal capacity, the cost of reasoning is about 10 times lower per year. In November 2021, GPT-3 capacity (MMLU 42 points) required $60 million token; by the end of 2024, the equivalent level was only 0.06 dollars, down 1,000 times in three years. Epoch AI ' s subsequent measurements are more radical: the median year drops from 50 times to 200 times after January 2024.
And the price? The price anchor in the client's heart is still in the age of "to do." The price is determined by the perception, which is far from keeping pace with the token price.
There's a more intuitive version of the time difference for a code writer: five years ago, you made the same product, you needed a front end, a back end, a designer, a few months' pay; now you add an AI tool to one person, at a cost of thousands of dollars a year. The cost of a single attempt is down a few orders of magnitude -- this is the most friendly part of the dividend for programmers: you don't need to bet on the whole, you just have to try a few more times.
In the original LLMflation of a16z, the conclusion is that equal capacity is 10 times cheaper per year.
ii. The window is real: don't look at emotions, look who's paying.
The simplest way to judge whether a thing is a scam is to see if someone pays for it. Emotions can be faked and payment records difficult.
Stripe analyses the 100 AI companies with the highest revenue on the platform: the median time to reach $1 million in annualized income was 11.5 months, four months faster than the fastest in the SaaS gold age; 24 months in annualized income of $5 million, and 37 months in the head of SaaS in 2018. The case is even more exaggerating: Lovable 17 million USD, Bolt 20 million USD, Cursor 100 million USD three years.
The number of one company is no less. Danny Postma does an AI occupational license, HeadshotPro, with an annual recurrent income of about $3.6 million; Maor Shlomo ' s Base44 was purchased six months online by Wix with $80 million in cash; Pieter Levers has no employees, no financing, and earns more than $3 million a year.
There's sample in the country. Manus came online in March 2025, with annualized revenues breaking $100 million at the end of the year; in December 2025, Meta announced purchases at $2 billion, and on April 27, 2026, was stopped by the NDC. The rhythm of Kling is even more dramatic: on March 25, 2026, Open AI announced the closure of Sora, the fourth quarter of publish, the same day's fast-track financial paper Kling 2025, revenues of $340 million, single-month revenue breakthroughs of $20 million in December, annualized income performance rate of $240 million, and annual income of approximately $1.04 billion. The same track, the same day, the family was leaving, the family was collecting money.
These are companies, away from individuals. But one thing in common: they make the difference between "old price anchor" and "new production cost". The difference between HeadshotPro and the 2023-exploded duck camera is that it sells to people at work to make LinkedIn's face, the scene is narrower and more demanding, and the client is willing to pay more for what looks like a professional.
The official blog of Stripe, AI, is rewriting the growth rules.
The title of the report in the Securities Times is the full story - Sora's exit was the same day that Kling doubled his income.
III. Cold water first: more than half of the products have not earned money
See here if you start to bleed, it's time to pour cold water. It's full of survivors, and survivors' deviations are the biggest scam in the industry, bigger than the "AI is a scam" itself.
More honest figures from Trust MRR. This platform is directly linked to payment channels such as Stripe, showing verified real revenues. In September 2026, researchers captured all 10,150 products on the platform:
Two details worth staring at. First, AI products and non-AI products have almost the same income probability (49.3 per cent versus 49.0 per cent), with a median monthly income of one \(161, one \175) - AI changed the cost of making the product, not the basic question of "no one buys." Second, of the 8,817 products captured in August, the developers ' tool category is zero-income disaster zone: 69 per cent was confiscated. The most people make tools and the least people get money.
It's a magnified version of the same table. AI Comics 2025 market size was $16.8 billion, which sounds like a windfall; however, 22.19 million new Douyin hits in the first half of 2026, 98.7 per cent below the profit/loss line. The cost is down 90%, the threshold is down, everyone comes in.
My duck camera is the specimen of the window closed. In July 2023, it was set on "9.9 dollars to achieve real freedom of writing" one-night fire, peaking at 4000 people, waiting for a dozen hours, breaking 600,000 a day; then the same product piled up and the user's price anchor went from "hundreds of the studio" to "a few dollars." The team was disbanded at the end of September, 2025, less than two years after the explosion. The window is real and closed faster than you think. That summer in the circle of friends, the one who laughed at "Ai writing a thousand people" didn't get a penny.
Tech Startups' analysis of the full amount of Trust MRR data, the title is the answer.
Ongwan: The market of $16.8 billion, and the cost of $500 per minute, is in the same story.
Chongqing Daily reproduces the Red Star Capital Bureau: the team was dissolved at the end of September 2025, with minimal operation.
iv. Why do we prefer to hoard lessons and not end up?
The window is real, and the failure rate is real. Reasonable action should be to increase the number of attempts and reduce the single cost. But most of the candidates are on the other side: commenting, and learning with anxiety.
The cost of commenting is zero. Saying "AI is too strong" doesn't require any skills, and it gives you a sense of aesthetic superiority and a sense of prompt arrival. The cost of the next event is real: it takes time to expose the work, to face the odds of failure and ridicule. Between "zero cost, instant return" and "cost, delayed return, probability failure", inertia will make your choice.
The word "AI tastes" is perfectly valid as an aesthetic judgment, and I'm tired of that instant tone myself. But it makes little sense as a commercial judgement: the customer pays for the result — the problem is solved, time is saved, money is earned. The critics of "AI taste" are evaluating the works, the payers are evaluating the results, and the doubles are not in a single market. And "AI tastes" is an engineering problem, and models are getting better every generation, and the moats that they keep are getting lighter every year.
It's more interesting to talk about it than to comment, to learn with anxiety. This type of economics is classic sample: DelaVigna and Malmendier, who have tracked for three years the three United States gymnasiums with 7,752 members, with a monthly fee of over $70, on average only 4.3 times a month, at a rate of over $17 a month - and who could have bought 10 cards at $10 a time. The average cost per person is about $600. They overestimate themselves for the future.
Buying lessons is the same psychology: the moment you pay, it's not the ability, it's the "I'm trying." MOOC's data are even more straightforward: the study published in 2019 in Science by Reich and Ruipérez-Valiente analysed 565 lessons at MIT and Harvard at edX, 12,670,000 enrolments, with only 3.13 per cent of certificates obtained; the completion rate for paid trainees was between 44 and 56 per cent. The gap is not in the quality of the course, it is in terms of goals and costs.
A time check. According to China’s Web Audio-Visual Development Study (2025), short video is used on a single-day basis for 156 minutes, approximately 949 hours a year, almost half a full-time working year. You said you didn't have time for a side job, all the time, just on a video of how ordinary people get into 10,000 dollars on AI.
There is also a more solid article for programmers: the tools of this era are changed every few months. You took three months to finish the frame, maybe the day you finished. Learning without customer demand is equivalent to running backwards on an accelerated forward conveyor belt. The real good thing is on a needs-based basis: clients want a function, I learn how to do it; products are being distributed, I learn how to put them. Each study is tied to a specific, paid problem.
Two of the most common scenery statements have been removed.
"Ai stole my job. "This is true, it should not be downplayed. The Stanford Brynjolfsson team updated its study in August 2026: In AI ' s high-exposure occupations, 22 to 25-year-olds are 19% less likely to be employed than their peers in low-exposure occupations, a gap of 13% a year ago. But the same study added that there was no massive unemployment in the economy as a whole and no similar gap for experienced people - - The shock is concentrated in entry positions that rely on "codifiable knowledge" and occurs mainly through low recruitment, not downsizing. Translating into business logic: AI underpresses the price of "selling standardized labour by job", the ability to judge and be responsible for results, and the price is not down. The rational response is to move from "pricing labour" to "pricing results."
"AI is a scam." As of December 2025, China generated 602 million AI users, a penetration rate of 42.8 per cent, an annual increase of 141.7 per cent. Valuation can have bubbles, many companies will die, many narratives will be broken, and I agree. But bubbles and scams are different: the Internet bubble broke down in 2000, a large number of companies died, and Amazon didn't disappear. Declaring the bubble as a fraud is logically unworkable, and the operational effect is only one: to find a decent reason for not to end up.
Pew's survey in June ended this year: 52 per cent of adults in the United States said they were more worried than excited about AI, which was 37 per cent in 2021, and 47 per cent of AI experts said they were more excited than worried, and only 11 per cent of the general public. The closer to this technology, the more optimistic the people who are working with it, the more anxious the people are. Anxieties and distance are related, not much to the truth.
Thesis of the digital economy laboratory in Stanford, updated in August, was preceded by a subheading question mark.
Pew ' s age group data from the June survey - the most optimistic group of 18-29 years old five years ago - now have the highest rate of concern.
success rate is the exact zero.
Put the previous data in a simple model. It is assumed that the probability of annualizing the income of $100,000 per product trial is p. According to Trust MRR, p is about 2.4% - this is already optimistic, and those who want to open their incomes are optimistic themselves. The probability of success at least once is:
P (at least one success) = 1 − (1 - p)^n, p ≈ 2.4%
The most important number in this table is 0% in the first line. No matter how many lessons you've learned, how many lessons you've collected, how many words you've said "AI is too strong", in this formula, n is equal to zero. Those who do do do have a low but zero probability of rising with experience; those who do not face a certain number: zero.
So how do you add? Look at the single cost. My own few projects that did not make money, the entire infrastructure ran on a free set, and fixed spending was the subscription of $1,200 a month. Five years ago, the same product required at least a front-end, a back-end, a designer for several months.
That's why I said "primitive capital accumulation" is the best thing to do now. The original capital is not just money; it's the first customer, the first case that comes out, the true understanding of a sub-market, and the sense of "I did it once." More importantly, money buys time: with the first seed capital, you can choose what to do next without anxiety and rent. In times of such rapid change, the choice of choice is itself the most scarce asset.
VI. Minimum action by developers
Compress all the above data into an enforceable sentence: select a specific issue where someone is willing to pay, at the lowest cost, and start charging on the first day.
Why is it charged on the first day? Because fees are the only honest way to verify. It doesn't count. It counts. The increase in the p in the formula is almost entirely due to a change in the judgement about what is bought, which can only be learned from real payments and refusals.
Why pick specific questions? Because no one buys the broad "AI tools", specifically "help me save a studio." HeadshotPro only makes a job appearance, doesn't write real, doesn't make drawings, doesn't take photos.
Why the lowest cost? Because you'll probably fail more than once. The lower the cost, the bigger you can afford.
There are three checks before you do it.
- The price anchor check, the customer's price, still at the old cost? Looking for things that "must have to be done in the past, and now AI can do 80 percent": cinema truth, outsourcing animation, product details pages, industry reports, small tools within the enterprise. If the client has defaulted "this thing AI can get a few bucks," the window closes.
- Repurchase checks. Will clients pay a second time? The mackerel was planted on a one-time paid, and after 9.9 dollars, the user had no reason to come back, and the growth depended on freshening. A person who is able to charge a monthly fee has priority over a person who is charged a secondary fee.
Do you have anything else you can't get in? AI lost 90% of the cost of the comics, leading to a six-month roll-out of 220,000. The lowering of the threshold of the tool has lowered for all. What really keeps you in a crowded place is not AI, but something outside AI: your five years of experience in a trade, your 20 clients, the segmented people you know better than anyone else. AI is the leverage, the leverage to get on your own feeder.
Delete the list of studies from a to n. Leave the one that's really needed for the product. The rest, when clients force you to learn.
Seven or three rebuttals. I'll go first.
"You're the survivor's deviation. " Yes, so section III includes the table: 26.1 per cent of product zero income, with only 2.4 per cent of $100,000 per year. I never said it would happen. I said it wouldn't happen. The difference between the two sentences is the whole of the article.
"In the end, you're selling anxiety." I thought about that. The difference is that the person who sells anxiety eventually gives you a payment link to buy his antidote; there's no payment link at the end of the article, and the only advice I give is to find someone who's willing to pay you. If you're more anxious to read it, turn it off, do a little thing that can get $9.9, and it's faster than reading ten articles.
"Not everyone is fit to do business. "This one I admit, I can't answer. There are people who have mortgages and children, without any margin of error, and there are people who are not born to be good at dealing with people, and it is harder for him to take the first payment than to work overtime. For them, the cost of "end" is much higher than I'm counting. I put it at the end, leave it to you.
Finally.
Every era has a dividend. There are PCs, there are Internet, there are mobile Internet, and now it's AI. In each case, the dividend is not for people to be rich; it is for those who are ready and who do. When the window was open, most people were commenting, worried, studying; when the window closed, most people started saying, "Just know."
You can laugh at those who do it, at the al taste of their work, at the small money they make. But don't forget the fifth formula: they add one every time they try, and you laugh every time you laugh, or nothing. What you missed was not a few fortunes, but a road that could have been tried over and over again.
I've got all the first-hand sources of data in the text (a16z original, Stripe report, Trust MRR analysis, Stanford thesis, fast-hand financial report), and I'll just go back to my "hongli" backstage.
Are you in or out? When did you last get the money from your own product? Next issue I'll pick the most famous scene, how to start from the beginning.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.