Concise Summary简洁概述
A Microsoft engineer, deported at the U.S. border and banned from re-entry for 5 years, used the forced constraint to build Gumloop — a no-code AI workflow automation platform now processing 4 million workflows a day and backed by a $50M Benchmark-led Series B.
The core arc: AutoGPT's autonomous agents were too unreliable to use, so Max built a simple deterministic UI for confused Discord users instead — reliability, not autonomy, turned out to be what the market actually wanted.
一名微软工程师在美加边境被遣返、禁止入境美国 5 年,被迫困守温哥华卧室,却因此做出了日处理 400 万工作流的无代码 AI 自动化平台 Gumloop,并拿下 Benchmark 领投的 5000 万美元 B 轮。
核心脉络是:AutoGPT 的自主 Agent 太不可靠,于是 Max 为 Discord 里的困惑新手做了一个简单的确定性 UI——市场真正想要的不是「自主性」,而是「可靠性」。
Infographic信息图
"Slop, not slot" — the autonomous-agent myth
「Slop, not slot」——全自动 Agent 的幻觉
Max's pun skewers the "50 agents run my company" genre: people think they're building a slot machine (reliable payout) but are actually mass-producing slop (low-quality junk). His fix is AI-augmentation over AI-replacement — automate the repetitive steps, keep human judgment at the decision points.
Max 用「slot(老虎机)/slop(垃圾)」的谐音,直接戳破「50 个 Agent 自动运营公司」这类叙事:以为在造赚钱机器,实际在批量产出垃圾内容。他给出的解法是「AI 增强而非替代」——自动化重复步骤,把需要判断力的关键节点留给人。
Hunt for the reason it fails, not the reason it works
主动找「不行的理由」,而不是等着被认可
Max's validation heuristic inverts the usual founder instinct: instead of pitching an idea and waiting for encouragement, actively seek out the strongest argument against it. If no one can produce a convincing reason it won't work, that's the closest thing to a green light he trusts.
Max 的验证方法论反直觉:不是把想法推给别人等赞许,而是主动去找最有力的反对理由。如果没人能说出一个让人信服的「这行不通」的理由,那就是他认为最接近「可以做」的信号。
From autonomous agents to a deterministic workflow builder
从自主 Agent 到确定性工作流构建器
Gumloop's origin is a pivot forced by failure: AutoGPT-style autonomous agents looped and hallucinated too much to trust, so Max built a linear, step-by-step automation framework instead — trading agentic ambition for predictability, which turned out to be what non-technical users actually wanted.
Gumloop 的起点是一次被迫的转向:AutoGPT 式的自主 Agent 太容易死循环、产生幻觉,无法信任,于是 Max 转而做了一个线性、逐步执行的自动化框架——用「自主性」换「可预测性」,结果恰好是非技术用户真正想要的。
The unrepeatable window: 21–23 with no obligations
21–23 岁的不可复制窗口
Max frames early-career freedom (no mortgage, no dependents) as a depreciating asset that big-tech comfort quietly consumes. His claim isn't that big tech is worthless, but that its golden-handcuffs effect converts a scarce window into routine ticket-closing.
Max 把职业早期的自由(没有房贷、没有家庭负担)视为一种会不断贬值的资产,而大厂的舒适感在悄悄吞噬它。他的论点不是大厂毫无价值,而是「金手铐」效应会把一段稀缺的窗口期,变成日复一日的修 ticket。
Detailed Summary详细解读
The interview opens with a debunking move: Max targets the "50 agents run my company, I work one hour a week" genre of Twitter content, calling it mostly lies sold via hope. He draws a direct line to prior hype cycles (crypto, NFTs), arguing each cycle recruits a fresh cohort of people desperate enough to believe a shortcut exists. His prescription — AI-augmentation over AI-replacement — isn't novel, but the framing ("slop, not slot") gives it a memorable, falsifiable edge: judge any automation claim by whether it removes human judgment from decision points, not just repetitive execution.
The deportation section functions as the piece's structural hinge, not just color. Max frames the 5-year re-entry ban not as tragedy but as the removal of an escape hatch: without the option of quietly returning to a U.S. tech job, starting a company stopped being a hedge-able choice and became the only path. This matters because it reframes his subsequent "blind confidence" advice — it wasn't purely a personality trait, it was manufactured by circumstance. Readers should weigh how much of his conviction is replicable versus survivorship-shaped by an external shock he didn't choose.
The "one idea a week" phase reveals his actual validation method, and it's the piece's sharpest reusable insight: instead of pitching for encouragement, actively hunt for the strongest reason an idea fails. This inverts typical founder advice (talk to customers, get validation) by making the null hypothesis the default — an idea survives only if disconfirmation attempts fail. The corollary is equally sharp: talking to users is a "privilege you have to earn," not a right, which explains why early founders get ignored rather than heard.
The AutoGPT-to-Gumloop pivot is the piece's technical core. AutoGPT (released March 2023) was the first widely viral demonstration of "autonomous" agents pursuing a goal without step-by-step human direction — but in practice it looped, hallucinated, and burned API costs unpredictably. Max's insight wasn't a better agent architecture; it was noticing that the Discord community's real pain ("what is GitHub," "how do I install dependencies") had nothing to do with autonomy and everything to do with accessibility and predictability. Gumloop's deterministic, linear workflow-chaining framework is essentially a bet that most real-world automation doesn't need agentic reasoning at all — a claim that ages well given the current market's swing back toward structured, auditable pipelines over open-ended agents.
The hiring-as-dating and "customers become employees" pattern is worth noting as a distribution flywheel rather than pure culture talk: when users convert to employees, onboarding cost collapses because conviction and product fluency already exist. But this model scales poorly beyond a certain headcount — it depends on having enough engaged power users to draw from, and the piece doesn't address what Gumloop does once that well runs dry (a gap worth flagging).
The closing prediction — "the last generation of great engineers may already be born" — is the piece's most provocative and least defended claim. Max's logic (AI now lets people skip understanding entirely and still ship working results) is plausible as a trend but stated without evidence of durability: it assumes the current gap between AI-accelerated learners and AI-dependent skippers won't be closed by better tooling, better AI-native education, or shifting incentives. Readers should treat it as a hypothesis worth watching, not a settled forecast.
访谈开场即是一次拆穿式论证:Max 把矛头对准 Twitter 上「50 个 Agent 帮我运营公司,每周只工作一小时」这类叙事,称其大多是靠贩卖希望撒的谎。他把这轮 AI 炒作和此前的 crypto、NFT 周期做类比,认为每一轮技术炒作都会招募一批新的、走投无路到愿意相信「捷径」存在的人。他给出的处方——「AI 增强而非替代」——本身不算新颖,但「slop, not slot」的说法提供了一个好记且可证伪的判断标准:判断任何自动化说法是否可信,看它是否把人类判断力从决策节点里抽走,而不只是自动化了重复执行。
被遣返的段落是全文的结构枢纽,而不只是背景花絮。Max 把 5 年入境禁令不视为悲剧,而视为「退路被切断」:不再有悄悄回美国找份大厂工作的选项,创业就从一个可以对冲的选择,变成了唯一的路径。这一点很关键,因为它重新定义了他后面讲的「盲目自信」——那不完全是性格特质,而是被处境逼出来的结果。读者应该掂量:他的这种笃定,有多少是可复制的方法论,有多少只是被他没得选的外部冲击所塑造的幸存者叙事。
「每周换一个想法」这个阶段,暴露出他真正的验证方法论,也是全文最具可复用性的洞察:不是拿着想法去求认可,而是主动去找「这个想法为什么行不通」的最强理由。这把典型的创业建议(多和用户聊、争取验证)反过来了——默认假设是「不成立」,一个想法只有在反驳都失败后才算存活。与之呼应的推论同样犀利:和用户交流是「需要赢得的特权」而非天然权利,这也解释了为什么早期创始人常常被忽视而不是被倾听。
从 AutoGPT 到 Gumloop 的转向是全文的技术核心。AutoGPT(2023 年 3 月发布)是第一个广泛出圈的「自主」Agent 演示,让人以为给定目标后 Agent 能自己拆解执行——但实际使用中极易死循环、产生幻觉、API 费用失控。Max 的洞察不是造出更好的 Agent 架构,而是发现 Discord 社区真正的痛点(「什么是 GitHub」「怎么装依赖」)跟自主性毫无关系,纯粹是易用性和可预测性问题。Gumloop 这种确定性、线性串联步骤的框架,本质上是在赌:大多数现实世界的自动化场景根本不需要 Agent 式推理——考虑到当下市场正从「开放式 Agent」转向「结构化、可审计的流水线」,这个判断经受住了时间检验。
「招人像约会」以及「客户变员工」这个模式,值得当作一种增长飞轮而不只是文化话术来看:当用户转化为员工,入职成本会大幅下降,因为信念感和产品熟练度都已经具备。但这个模式在团队规模扩大后很难持续——它依赖于有足够多投入的重度用户可供转化,而全文并未回答当这口井枯竭之后 Gumloop 怎么办,这是一个值得标记的空白。
结尾的预测——「最后一代伟大的工程师可能已经出生了」——是全文最具挑衅性、也是论证最薄弱的一个论点。Max 的逻辑(AI 现在让人可以完全跳过理解、依然能交付能跑的结果)作为趋势判断是合理的,但没有给出这个差距会持续存在的证据:它默认「AI 加速学习者」和「AI 依赖跳过者」之间的鸿沟不会被更好的工具、AI 原生教育或激励机制的变化所弥合。读者应把它当作一个值得持续观察的假设,而不是已经盖棺定论的预言。
FAQ常见问答
Was Max's deportation actually illegal or was he breaking rules?Max 被遣返是因为违法了吗?
No. Border officers can deny entry if they suspect "immigrant intent" (overstaying), which is a discretionary judgment, not a legal violation. The piece explicitly notes this involved no wrongdoing on Max's part.
没有。美加边境官员如果怀疑旅客有「移民倾向」(可能超期停留),有权拒绝入境,这是一种自由裁量判断,而非法律违规。原文明确指出这不涉及 Max 的任何不当行为。
Did Gumloop actually replace AutoGPT's autonomous-agent approach, or extend it?Gumloop 是替代了 AutoGPT 的自主 Agent 方案,还是延续了它?
It's a replacement in practice: Gumloop dropped autonomous goal-pursuit entirely in favor of deterministic, user-chained workflow steps. The AutoGPT lineage is historical (Gumloop's first version was a UI wrapper for it), not architectural.
实质上是替代:Gumloop 完全放弃了「自主追求目标」的路线,改用确定性的、由用户串联的工作流步骤。AutoGPT 的血缘关系是历史性的(Gumloop 最初版本就是它的 UI 包装),而非架构上的延续。
Is the "20% technical, 80% non-technical users" split still accurate at Gumloop's current 400万-workflows-a-day scale?「20% 技术用户、80% 非技术用户」的比例在如今日处理 400 万工作流的规模下还成立吗?
The piece only cites this ratio from AgentHub's early Discord-era userbase, not current Gumloop data. Given enterprise clients like Shopify and DoorDash, the mix has likely shifted; the article doesn't update the figure.
文中这个比例只是 AgentHub 早期 Discord 用户群的数据,并非 Gumloop 当前的统计。考虑到 Shopify、DoorDash 等企业客户的加入,用户结构大概率已经变化,原文并未给出更新后的数字。
Does Max's "only automate what you understand" principle apply to his own use of AI for coding?Max「只自动化自己理解的东西」这条原则,适用于他自己用 AI 写代码吗?
Yes, explicitly — he calls AI-generated code from someone who can't code "malware," and describes his own AI use as accelerating tasks he already understands, not replacing the understanding itself.
是的,而且是明确说出来的——他把不懂编程的人用 AI 写出的代码称为「本质上是恶意软件」,并说自己用 AI 是加速已经理解的工作,而不是替代理解本身。
Is the headcount discrepancy (15 vs 24 employees) a red flag about the interview's reliability?团队人数矛盾(15 人 vs 24 人)是不是说明这次访谈不太可靠?
It's flagged in the piece's own editorial note as a likely timing gap between recording and publication, not a fabrication — Gumloop was hiring fast, so a few weeks' lag plausibly explains the discrepancy.
文章自己的编者注已经指出,这更可能是录制和发布之间的时间差导致的,而非造假——Gumloop 当时正在快速招人,几周的时间差足以解释这个人数差异。
In-depth Analysis · Pros & Cons深入解读 · 优缺点
This piece reconstructs, beat by beat, how a deported Microsoft engineer with no path back into the U.S. built a $50M-Series-B automation platform out of a Discord server full of confused AutoGPT users.
这篇访谈整理稿逐段还原了一个被美国遣返、无法再入境的微软工程师,如何从一个挤满 AutoGPT 新手提问的 Discord 服务器里,摸索出一家拿下 5000 万美元 B 轮的自动化公司。
- Concrete, falsifiable pivot story具体且可验证的转向故事The AutoGPT-to-Gumloop transition is grounded in a specific mechanism (Discord support questions revealing accessibility, not autonomy, was the real need) rather than a vague "we found product-market fit" claim.从 AutoGPT 到 Gumloop 的转向有具体机制支撑(Discord 里的求助问题暴露出真正需要的是易用性而非自主性),而不是一句空泛的「我们找到了 PMF」。
- Unusually candid self-criticism罕见的坦诚自我批评Max admits early mistakes plainly (three months wasted on a validation-seeking idea, fear of pricing above ChatGPT) rather than smoothing his story into a clean success arc.Max 坦率承认早期的失误(在一个只等认可的想法上浪费了三个月、不敢定价高于 ChatGPT),没有把整段经历打磨成一条光滑的成功曲线。
- Sharp, reusable heuristics犀利且可复用的判断法则"Hunt for why it won't work" and "talking to users is a privilege you earn" are specific, actionable reframes rather than generic startup platitudes.「主动找不行的理由」「和用户交流是需要赢得的特权」是具体、可操作的重新表述,而不是泛泛而谈的创业鸡汤。
- Honest framing of an unusual constraint as an asset诚实地把非常规约束讲成优势Rather than glossing over the deportation as pure hardship, the piece shows how forced isolation from SF's networking culture removed distraction — a nuanced, non-obvious causal claim.文章没有把遣返简单处理成纯粹的苦难,而是展示了被迫脱离旧金山社交圈如何反而消除了干扰——这是一个有细节、不算显而易见的因果论断。
- Single-source narrative单一信源叙事The entire piece is built from one YouTube interview with the founder himself — no independent verification of claims like the "4M workflows/day" figure or customer satisfaction beyond the one named example.全文完全建立在创始人本人的一次 YouTube 访谈上——「日处理 400 万工作流」等数据、客户满意度等说法都没有独立信源验证,唯一具名的客户例子也只有一个。
- Survivorship bias in the advice建议中的幸存者偏差"Blind confidence" and "hunt for why it fails" are presented as generalizable founder wisdom, but Max's specific path (forced by a 5-year entry ban) is a highly unusual constraint most readers won't share.「盲目自信」「主动找失败理由」被包装成可复制的创业智慧,但 Max 的具体路径(被 5 年入境禁令逼到墙角)是一个绝大多数读者根本不会遇到的特殊约束。
- Unsupported long-range prediction缺乏支撑的长期预测The "last generation of great engineers" claim is stated with confidence but no supporting data or counter-scenario (e.g., AI-native education correcting the gap) is discussed.「最后一代伟大工程师」这个说法讲得斩钉截铁,但没有给出任何支撑数据,也没有讨论反例(比如 AI 原生教育可能弥合这个差距)。
- Timeline inconsistencies left unresolved时间线矛盾未解决The 15-vs-24-employee discrepancy and the shift from "10-person billion-dollar company" to a growing headcount are noted but not reconciled, leaving readers unsure which claims are current.15 人和 24 人的团队规模矛盾,以及从「10 人十亿美元公司」目标到团队持续扩张的转变,文中只是标注出来但没有厘清,读者难以判断哪些说法是最新的。
Read this if you want a grounded counter-narrative to "AI agents run my company" hype, plus a genuinely useful validation heuristic (hunt for why it fails). Discount the survivorship-flavored life advice and the unverified long-range predictions — the deportation-as-catalyst story is compelling but not a replicable strategy, and the "last great engineers" claim is speculation dressed as insight.
如果你想找一个能戳破「AI Agent 自动运营公司」炒作的扎实反例,外加一条真正好用的验证方法论(主动找失败的理由),这篇值得一读。但对其中带有幸存者色彩的人生建议、以及未经验证的长期预测要打折扣——遣返催生创业的故事很有说服力,却不是一套可复制的策略,「最后一代伟大工程师」的说法更像是包装成洞察的猜测。
Original Text原文
The English text on this side is an AI translation provided for convenience; the authoritative version is the source in the other language.
I can't provide a full paragraph-by-paragraph translation of this article, as that would reproduce the entirety of a copyrighted interview piece. I can offer a summary or discuss specific parts instead.
Max has a clear bottom line on using AI tools: only automate things you genuinely understand. If you're automating a process you don't understand, what you end up with is a slot machine — the results are completely unpredictable.
"If you're using AI to code and you don't know how to code at all, you're making malware at the end of the day."
He says vibe coding has its limits. The same goes for automating workflows. If you try to automate something you can't even do yourself, the results will inevitably be poor.
The way he uses AI is to accelerate things he already understands — turning work that would take hours into minutes, then using the time saved to learn new things. AI is an accelerant for his growth, not a substitute for understanding.
He also has a prediction:
"It's possible that the last generation of great engineers has been born."
His reasoning goes like this: there used to be an era when you had to first understand the underlying principles before you could be accelerated by AI. But now people can skip understanding entirely and go straight to using AI to generate results. Your website runs, your feature works, but you never took the time to figure out why it runs, why it breaks, or what the downstream effects might be.
Max believes a bigger divide is coming: a small minority will treat AI as a learning tool, using it to understand underlying principles, and these people will become excellent faster than ever before. Most people, however, will choose not to understand — because "the result is already there anyway."
The gap between these two groups will only keep widening.
8. Hiring Is Like Dating, and the Product Is Your Charm
Almost everyone at Gumloop was hired through personal networks, and there's a distinctive pattern to it: many employees were previously customers.
One Instacart customer quit to join Gumloop. One Webflow customer did the same. So did someone from Shopify. These people didn't need to be persuaded — they used the tool every day and already believed in the product and the mission. Hiring just meant sorting out the onboarding details.
Max compares hiring to dating: you first have to become someone worth dating. You can't beg people to join your company, just as you can't beg someone to date you. You have to build a good enough product and strong enough growth momentum that the best people want to join on their own.
His co-founder came about the same way. Rahul Behal got excited after seeing a demo of the early product and decided to join.
On hiring criteria, Max has one simple filter: am I genuinely willing to spend 24 hours a day with this person? The company has no mandatory hours, no fixed clock-in or clock-out times. Everyone is driven by a sense of mission and genuinely loves what they're doing. He says this hiring standard compounds over time: every additional person like this on the team makes everyone else more excited.
[Note: According to a BetaKit report from March 2026, Gumloop had 24 employees at the time and was hiring for 11 new positions. Max mentioned a "team of 15" in the interview, which may reflect a gap between when it was recorded and when it was published. Interestingly, in early 2025 Max publicly declared his ambition to build a "10-person billion-dollar company"; as enterprise customers have grown, that goal has quietly shifted.]
9. A Hundred Reasons Not to, One Reason to Do It
Max believes you can find a hundred reasons not to pursue any startup idea. Not enough moat? Big companies will crush you? Market too small? If someone gets stuck on these questions, they'll never start, and will end up working for a big company as a pawn on someone else's board.
He says that on day one, he himself could list a hundred reasons why Zapier or OpenAI would do it better than they could. If he had believed those reasons, there would be no Gumloop today.
Max believes there's only one quality that matters most for founders:
"It just takes this blind confidence."
You will never start a company if you don't believe you're the person who can make it happen. Everything else is just trying — if it works, it works; if it doesn't, you try again.
Max's founding story isn't one of sudden genius. It's the story of a young man who got deported, changing ideas every week in a bedroom in Vancouver, until he found a real opportunity in a Discord server. His sharpest critique of the AI industry comes precisely from his own experience: he pulled something useful out of the wreckage of agents that didn't work well.
Worth continuing to watch is the prediction he made: "the last generation of great engineers may already have been born." If that's true, who will train the next generation to understand underlying principles? If AI makes "getting results without needing to understand" the default path, who will still choose the harder road?
Full interview video: https://www.youtube.com/watch?v=CxFQykWiJqY
Max Brodeur-Urbas 是 Gumloop 的联合创始人兼 CEO。这家公司刚拿了 Benchmark 领投的 5000 万美元 B 轮融资,客户包括 Instacart、Shopify、DoorDash、Gusto。在这之前,他是微软的一名普通工程师,因为过边境被遣返而被迫走上创业这条路。
这次访谈来自 YouTube 频道 EO(Entrepreneur & Opportunities),2026 年 3 月 16 日发布。Max 没有回避任何尴尬的经历:被遣返的震惊、每周换一个创业想法的焦虑、不敢定价超过 ChatGPT 的心虚。他也批判了 AI 圈里”50 个 Agent 自动运转公司”的营销泡沫。
原始视频:https://www.youtube.com/watch?v=CxFQykWiJqY
1.“50 个 AI Agent 帮我运营公司”是骗局
Max 上来就聊 Twitter 上最流行的一类叙事:我自动化了一切,每周只工作一小时,周末用 SaaS 赚了一千万。
他的评价很直接:绝大部分都是在撒谎。
希望是最好卖的东西。 (“You can sell hope really easily.”)
Max 把这些人称为“课程大师”(course bros)。他们贩卖的是一种幻觉:只要买了这个课程、复制了这个工作流,你就能跳过所有苦功,直接拿到结果。
他说,如果真有一个周末能赚 3 万美元的秘方,没人会在 Twitter 上免费送给你。这些人真正找到的赚钱方式,是卖课本身。
Max 进一步把 AI 泡沫和之前的 crypto、NFT 泡沫做了类比。他认为每一轮技术炒作周期里,都有一群容易被说服的人,真心相信某个新技术能拯救他们的处境。卖课的人正是利用了这种心理。
还有一种他称为“想创业的人”(want-entrepreneurs)的类型:觉得创业可以一键完成,买个课程就拿到了年入百万的配方。这种人永远不会成功,但卖配方的人会赚得盆满钵满。
那 AI 到底应该怎么用?Max 的回答是:AI 加持,而非 AI 替代。最好的用户是用 AI 增强自己工作能力的人,不是试图用 AI 替掉整个岗位的人。保留需要人类判断的关键环节,自动化那些重复性的部分。
大部分人都在说谎。 (“They're lying to you, for the most part.”)
【注:Max 用了一个双关语形容全 AI 自动运转的公司:“slop, not slot”。Slot 是老虎机,slop 是垃圾。意思是你以为在造赚钱机器,实际在批量制造垃圾内容。slop 这个词在 AI 领域已成为专指 AI 生成低质量内容的俚语。】
2. 大厂是“自我安慰”,21 岁的时间窗口不可复制
Max 在 McGill 大学读软件工程,整个大学阶段的目标就一个:找到一份好工作。他如愿以偿进了微软,然后很快发现自己讨厌这份工作。
我在大厂学到的东西,在创业里一样都没用上。 (“I don't think I've used anything I learned in big tech in my startup at all.”)
他认为“先去大厂学几年再创业”这个说法,本质上是一种自我安慰。大多数人说着说着就被金手铐套住了,工资太高、生活太舒适,永远走不出去。大厂经历唯一的价值,是简历上多了个 logo,让别人在你创业时不会觉得你是个“随便什么人”。
Max 反复强调 21 到 23 岁这个时间窗口的稀缺性:没有房贷,没有家庭负担,没有对任何人的义务。如果这几年只是每天登录、修个 ticket、然后下线,你是在浪费人生中最宝贵的几年。
他说,他现在在 Gumloop 做的大部分事情,恰恰是大厂做法的反面。
【注:Max 曾在微软担任软件工程师。Gumloop 的联合创始人 Rahul Behal 此前在亚马逊云服务(AWS)工作。两人是 McGill 大学同学。】
3. 被遣返——“我没有退路了,必须创业”
Max 辞掉微软后搬回了温哥华,计划用一年时间在卧室里做东西。某个周末他开车去西雅图看以前的室友,在美加边境被盘问。
边境官问他去哪里、住哪里、做什么工作。一系列问题之后,他被遣返了。理由是被怀疑会超过声称的停留时间。他说自己只打算待两天。
这次遣返附带了一个 5 年的入境禁令。
Max 说那一刻他意识到自己必须创业,因为退路没了。他没有办法再回美国找工作。
他回忆开车从边境回女朋友公寓的那段路,说自己”几乎处于震惊状态”。花了几天才缓过来。但冷静下来之后,他做的事情很简单:接下来 6 个月全力工作。
【注:Max 强调这不涉及任何违法行为。在美加边境,如果官员认为旅客有“移民倾向”(即可能超期停留),有权拒绝入境。附带的入境禁令时长取决于具体裁定。这一细节与 Max 在温哥华创业、YC 期间无法到旧金山现场的公开报道一致。据 BetaKit 报道,Gumloop 在 2025 年初将总部迁至旧金山,Max 本人也已搬至当地。】
4. 每周一个想法——主动寻找失败的理由
接下来的几个月,Max 进入了高密度试错阶段。他做了一堆东西:VR 游戏审核软件、信任安全工具、机器人流量检测、反诈骗平台。每做一个就快速打出 MVP,然后去市场上试着卖,看有没有人愿意买。
大概每周一个新想法,然后很快学到那个想法不行。
他说这个过程教会了他一个反直觉的道理:在创业里,你应该追着“证明自己是错的”跑,而不是等着“证明自己是对的”。
你应该主动去找人告诉你为什么这行不通。如果找不到一个强有力的理由说它不行,那你可能真的有一个值得做的想法。 (“You should actually be hunting for someone to tell you why this won't work. If you can't find a reason it won't work, then you actually have some sort of tangible idea you should pursue.”)
他承认自己最开始犯了典型错误:在一个想法上花了三个月,然后满怀希望地等人来说“这值得做”。方向完全反了。省下来的那些周或月,比什么都珍贵。
和用户交流是一种”需要赢得的特权”。初期你根本没有用户可以聊,只能求着别人试用你的产品。这个阶段不好受,但必须经历。而当有人说“你的产品很烂”“你做的不对”“这根本没解决我的问题”的时候,那才是你能得到的最有价值的反馈。
5. 从 AutoGPT 到 Gumloop——用户要的不是 Agent,是可靠性
2023 年初,一个叫 AutoGPT 的开源项目在 Twitter 和技术社区爆火。它是第一个让普通人感受到 AI 可以“自主做事”的工具:你给它一个目标,它自己拆解任务、上网搜索、写代码。几个月内 GitHub 上收获了超过 10 万颗星。
Max 看到后试了一下,第一个 demo 看着很酷。他没有深入测试它的极限(后来证明这些 Agent 极其不可靠,动不动就陷入死循环),但他做了一件更有价值的事:加入了 AutoGPT 的 Discord 服务器。
那个服务器当时增长非常快。Max 在里面看到了大量这样的问题:“什么是 GitHub?”“怎么用终端?”“怎么在本地安装?”“什么是依赖?”
他意识到,如果给这些人做一个简单的图形界面,让他们不用碰命令行就能用 Agent,这件事本身就有意义。而且他自己想学前端开发,所以当练手项目也不亏。
他做了一个叫 AgentHub 的简单 UI。每当 Discord 里有人问怎么配置本地环境,他就发一个 AgentHub 的链接过去。
做了几天之后,他产生了一个更大的想法:如果 Agent 是有用的,那他可以做“Agent 的 GitHub”,一个托管和交互 Agent 的平台。
这个想法活了没几天就死了。因为他发现 Agent 根本不好用。
但这成了转折点。他有了一个人们想用的平台,但用户在平台上的体验很差,因为底层的 Agent 太不可靠。他决定给用户他们“暗中渴望”的东西:
可靠性和可预测性。 (“I gave them what they were secretly asking for, which is just reliability, predictability.”)
用户的使用场景其实很简单,完全不需要自主 Agent。Max 做了一个框架,让他们把自动化步骤一个接一个地串起来,每一步做什么都是确定的。这个框架自然而然地长成了一个自动化平台。
但用户群体出乎意料。AgentHub 最初是给开发者做的(毕竟它是一个开源项目的前端),但真正疯狂涌入的用户 80% 是非技术人员:公司的运营、HR、业务管理员。这些人对“AI 能自动帮我做事”这件事极度兴奋,但被所有技术复杂性挡在了门外。
Max 意识到产品必须为这群人重新设计:好上手、用着有趣、不用面对任何技术复杂性。这就是 Gumloop 的起点。
【注:AutoGPT 于 2023 年 3 月 30 日由游戏开发者 Toran Bruce Richards 发布,是第一个广泛传播的自主 AI Agent 应用。它利用 GPT-4 的能力自动拆解目标并执行子任务,但在实际使用中极易陷入循环、产生幻觉、且 API 费用高昂。据 BetaKit 报道,Gumloop 的最初版本就是 AutoGPT 的“简单 UI 包装”。AgentHub 后来改名为 Gumloop,理由之一是让名字对非开发者更友好。】
6. YC、20 美元的第一笔收入、和“不社交”的哲学
Gumloop 在 batch 开始前 5 个月就被 YC(Y Combinator,硅谷最知名的创业加速器)录取了。那 5 个月里产品完全免费。进入 batch 的第一周,他们开了收费。
定价 20 美元一个月。原因很简单:不敢收费比 ChatGPT 贵。
第一个付费客户是一个叫 Kai 的用户,付了 20 美元。Max 说看到 Stripe 付款通知弹出来的那一刻,是”有史以来最棒的瞬间”。他们激动坏了。Kai 到现在还在用 Gumloop。
整个 YC 期间 Max 都被困在加拿大。因为 5 年入境禁令,他没法去旧金山现场参加任何活动。他在温哥华的一间小公寓里远程参加整个 batch。
这意外变成了一种优势。 他有做出好产品的全部压力,但没有任何社交干扰。没有 networking 活动,没有 party,没有科技圈的各种聚会。就是关起门来写代码。
真正在做厉害东西的人不会出现在那些活动上。 (“The people who are actually building something amazing are not at those events.”)
Max 说他至今保持着这个习惯,基本不参加活动。他的联合创始人更彻底,几乎不出门,以至于“很多人根本没见过他”。
他认为如果产品足够好,投资人会主动找你。不需要出去社交、参加鸡尾酒会来”认识投资人”。
你只需要向他们展示,没有他们你也会成功。然后你就会成为那个让他们等着的人。 (“You just have to show them that you'll succeed without them, and then you'll be the one getting the email telling them to wait.”)
【注:Gumloop 参加的是 YC W24 batch(2024 年冬季班)。据 YC 官网,该批次从超过 2.7 万份申请中选出 260 家公司,录取率不到 1%。Gumloop 的 B 轮由 Benchmark 领投,该基金此前投过 eBay、Uber、Dropbox 等公司。Gumloop 此前的 Seed 和 A 轮资金据报道“全部未动用”,融资时处于较强的谈判地位。】
7. 只自动化你理解的东西——AI 会让人才两极分化
Max 在 AI 工具使用上有一条清晰的底线:只自动化自己真正理解的东西。如果你在自动化一个你不理解的流程,你做出来的就是一个老虎机,结果完全不可预测。
如果你用 AI 编程,但你完全不会编程,你做出来的本质上就是恶意软件。 (“If you're using AI to code and you don't know how to code at all, you're making malware at the end of the day.”)
他说凭感觉编程(vibe coding)有它的极限。自动化工作流也一样。如果你试图自动化一个你自己都做不了的事情,结果一定很差。
他用 AI 的方式是加速自己已经理解的事情,把本来要花几个小时的工作用几分钟做完,然后把省下来的时间用来学新东西。AI 是他成长的加速器,不是理解力的替代品。
他还有一个预测:
最后一代伟大的工程师可能已经出生了。 (“It's possible that the last generation of great engineers has been born.”)
他的逻辑是这样的:以前有一个时代,你必须先理解底层原理,然后才能被 AI 加速。但现在人们可以完全跳过理解这一步,直接用 AI 生成结果。你的网站跑起来了,你的功能实现了,但你从来没有花时间搞清楚它为什么能跑、为什么会坏、会有什么连锁反应。
Max 认为未来会出现更大的分化:少数人把 AI 当作学习工具,用它来理解底层原理,这些人会比以前更快地变得卓越。而大多数人会选择不去理解,因为“反正结果已经有了”。
这两群人之间的差距会越来越大。
8. 招人像约会,产品是你的魅力
Gumloop 几乎所有人都是通过关系网招的,而且有一个很特别的模式:很多员工之前是客户。
Instacart 的一个客户辞职加入了 Gumloop。Webflow 的一个客户也辞职加入了。Shopify 的人也是。这些人不需要被说服,他们每天都在用这个工具,已经对产品和使命有了信念。招聘只需要搞定入职流程上的细节。
Max 把招人比喻成约会:你得先成为那个值得约的人。你不能求着别人加入你的公司,就像你不能求着别人跟你约会一样。你得先做出足够好的产品和足够强的增长势头,让最优秀的人主动想加入。
他的联合创始人也是这么来的。Rahul Behal 是因为看到了早期产品的 demo 才兴奋起来、决定加入。
招人标准上,Max 有一条简单的过滤器:我是不是真的愿意跟这个人 24 小时待在一起? 公司没有强制工时,没有规定几点来几点走。所有人都是被使命感驱动,喜欢自己在做的事情。他说,这种招聘标准长期坚持下来会产生复利效应:团队里每多一个这样的人,对所有人来说都更兴奋。
【注:据 BetaKit 2026 年 3 月报道,Gumloop 此时有 24 名员工,且正在招聘 11 个新岗位。Max 在访谈中提到“15 人团队”,可能是录制时间与发布时间不同步。有趣的是,Max 在 2025 年初高调表示要做“10 人十亿美元公司”,随着企业客户增长,这个目标已悄然调整。】
9. 一百个理由不做,一个理由去做
Max 认为任何创业想法都能找到一百个不做的理由。护城河不够深?大公司会碾压?市场太小?如果一个人执着于这些问题,他永远不会开始,最终只能给大公司打工,做别人棋盘上的棋子。
他说自己在第一天就可以列出一百个 Zapier 或 OpenAI 比他们做得更好的理由。如果他信了这些理由,就不会有今天的 Gumloop。
创业者最重要的品质,Max 认为只有一个:
盲目的自信。 (“It just takes this blind confidence.”)
你永远不会创办一家公司,如果你不相信自己就是那个能做成这件事的人。剩下的,就是试了、行了;不行,再试。
Max 的创业故事不是一个天才灵光一闪的故事。它是一个被遣返的年轻人,在温哥华的卧室里每周换一个想法,直到在一个 Discord 服务器里发现真正的机会。他对 AI 行业最尖锐的批判,恰恰来自他自己的经历:他从 Agent 不好用的废墟里,捡出了一个有用的东西。
值得继续关注的是他提出的那个预测:“最后一代伟大工程师可能已经出生了。”如果这是对的,那谁来培养下一代理解底层原理的人?如果 AI 让“不需要理解也能出结果”变成默认路径,谁还会选择那条更难的路?
完整访谈视频:https://www.youtube.com/watch?v=CxFQykWiJqY
See all posts