陈然 Ran Chen

Building business on autopilot with AI.

[陈然 Ran Chen @chenran818]

我说怎么买不到了

[急则疲慌则乱 @samchen0079]

美国政府个大煞笔以原料来自新疆地区为由把洽洽瓜子给禁了，白皮特么真幽默。🙂

[陈然 Ran Chen @chenran818]

My advice on those high-end ~$200/month AI coding plans is simple: sign up immediately. Not for more chat turns, but because you're buying into a temporary subsidy window that won't last.

Big model companies are currently in a fierce battle for the developer ecosystem. To win, they're offering all-you-can-eat compute for a low fixed subscription, a value that far exceeds the price.

This creates two mindsets.

Most are stuck with an 'API brain', anxiously counting every token as a direct cost. This is a scarcity mindset that limits experimentation.

The real breakthrough comes from adopting a 'subscription brain'. You start treating the AI as a fixed-cost production unit. An environment you can manage, push to its limits, and use for high-intensity iteration without worrying about a runaway bill.

This is a fleeting opportunity. It's the perfect time to use this subsidized compute to train yourself in managing complex AI-driven workflows—to build your own 'AI-native ship' at the lowest possible cost. That opportunity is closing.

[陈然 Ran Chen @chenran818]

The idea that confidence is the *result* of competence is a career-killing myth from a pre-AI world. An insightful piece from late 2025 argued that in reality, confidence is the entry ticket to *gaining* competence.

The linear path of 'skill → recognition → confidence' is a trap. In the modern workplace, saying 'I'm willing to learn' can be interpreted as 'I'm not ready to take responsibility'.

It's the person who projects 120% confidence with 60% of the skills who gets the breakthrough opportunity. They are handed the complex challenges that force a massive leap in their abilities, leaving the humble 'learners' behind.

This is more critical than ever as AI devours execution-focused tasks. The future-proof human value isn't in doing what you're told, but in defining the problem, owning the strategy, and being responsible for the result. That's an 'owner's mindset'.

And that mindset requires a level of courage and almost irrational self-belief that AI cannot simulate.

It’s time to stop waiting for permission to be confident. It has to be an active choice. Stop saying 'I'll try'. Start saying 'I'll take responsibility'. That's the difference between being an executor who gets replaced and an owner who becomes indispensable.

[陈然 Ran Chen @chenran818]

We love you KIMI🫡

[Nathan Lambert @natolambert]

Kimi K3 license. It's inspired by MIT but distinctly non-commercial, where any company making over $20M/yr must get a specific commercial deal (and display Kimi K3 if over 100M users or $20M/mo revenue)

[陈然 Ran Chen @chenran818]

Thank you!

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[陈然 Ran Chen @chenran818]

I see a painful paradox among many brilliant, high-income first-generation immigrants. They've mastered the art of earning, but remain stuck in a '攒钱' (saving money) mindset from a world they left behind.

This isn't about a lack of financial literacy. It’s about deeply ingrained programming from an environment where 'plans can't keep up with change'. This fosters a short-term view, prioritizing the tangible security of cash and property over a strategic 30-year blueprint.

This is often coupled with a strict 'employee mindset' that categorizes everything as a cost to be minimized. Professional financial advice, therefore, isn't seen as an investment in expertise but as a wasteful expense. The focus is on the fee saved today, not the risk mitigated or the tax-optimized returns gained tomorrow.

When evaluating options, the sole metric often becomes 'rate of return'. This simplistic ruler causes them to undervalue the risk management function of insurance and the tax advantages of retirement accounts. It also fuels an over-concentration in real estate, driven by cultural comfort and a desire for control.

The core issue, as I first analyzed back in 2025, is that survival tactics from a past life are failing in a new environment of abundance and complexity. The necessary evolution is a mental one: from simply accumulating assets to systematically managing wealth.

[Andrej Karpathy @karpathy]

@lakshyaag weird misinformation to find circling on twitter, no.

[陈然 Ran Chen @chenran818]

I used to think Zhuge Liang was the epitome of a genius leader. Now I see him as a warning.

Based on a leadership analysis I first encountered from 2024, his approach, once admired, now looks like a playbook on how to fail as a leader, even while succeeding as a brilliant individual.

First, he failed to cultivate successors. He inherited a team with the legendary Five Tiger Generals but left behind a talent vacuum so dire they had to rely on an aging Liao Hua for critical missions. A great leader builds more leaders, they don't consume them.

Second, his risk aversion was fatal. He famously rejected Wei Yan's audacious plan for a surprise attack, opting for the 'safest' route. This wasn't prudence; it was a failure of imagination. He avoided a small, calculated risk and in doing so, guaranteed a larger, strategic failure.

Finally, he was the ultimate micromanager. The records show him personally approving even minor disciplinary actions like '20 floggings'. This isn't dedication; it's a deep-seated inability to delegate that robs your team of the chance to learn, own their work, and grow.

Zhuge Liang's story isn't one of heroic sacrifice. It's the tragedy of a man who poured all his talent into doing the work himself, instead of building a team and a system that could thrive without him. He was a perfect manager, and that's precisely why he failed as a leader.

[陈然 Ran Chen @chenran818]

I've seen two types of engineers, and the one with the more impressive resume is often the less skilled.

There's the 'project-laden' engineer, who has a dazzling list of shipped features. They're masters of the short-term sprint. But when you ask 'why did you build it this way?', the answer is often a vague reference to a tight deadline. Their knowledge is a mile wide and an inch deep.

Then there's the 'artisan'. They might have fewer projects to their name, but they are deeply passionate about their craft. They stay hands-on. They don't just use a tool, they understand its internals. They value refactoring, reflection, and building something right.

Companies often praise the first type for their speed, mistaking constant activity for progress. The pressure to 'just ship it' can stifle the deep thinking that leads to true personal growth and robust systems.

In the short term, quick delivery looks good. But over a career, the market has a way of washing away the superficial. The ones who last are those who built their foundations on the solid rock of 'why'.

[陈然 Ran Chen @chenran818]

新文章：《如何从零开始学英语（和学 AI）》

我们学 AI 的方式，像极了小时候学英语：疯狂背单词、刷语法、收藏各种 Prompt 模板和 MCP 概念，却几乎从未在真实场景里解决过一个具体问题。知识当然有用，但我们很容易把“关于工具的知识”当成“使用工具的能力”。学习 AI 最有效的路径不是系统性地看完多少教程，而是“先闭环，后提效”——先让工具进入一个真实任务，拿到哪怕最粗糙的外部反馈，再决定接下来该补什么。

很多人在学 AI 时容易混淆“产出”和“结果”。写出十篇文章、生成几百行代码只是产出，而文章有没有被读懂、代码能否在目标环境里运行并解决问题，才接近结果。AI 极其擅长快速增加产出，这很容易让我们产生“已经解决问题”的幻觉。真正的闭环，必须包含外部反馈。哪怕你的方法再简陋，只要把结果交到真实的使用者手里，看他们是接受、修改还是拒绝，反馈自然会帮你筛选出下一步最该学什么。

在跑通闭环之前，急着去研究复杂的 Agent 架构、自动化工作流或者各种新概念，往往只是让错误跑得更快。开始时，手工操作并不丢人。自己整理输入，自己与 AI 沟通，自己核验结果。只有在手动跑过几次、摸清了卡点之后，提效才有针对性。重复的动作可以固化，容易错的地方可以增加检查。概念不该是一张必须背完的清单，而是解决具体摩擦的零件。先证明这条路通向一个值得去的地方，再考虑如何修路和换车。

最后，我们常常会问：学会英语或 AI，真的能让人赚到更多钱吗？现实的因果链其实是，工具降低了你探索和执行的成本，扩大了你能够接触到的信息和市场，但它无法替你决定哪个问题值得解决，更不能自动建立长期的信任。工具的熟练度，最终要由它帮你往前推进了多少具体业务来检验。无论是英语还是 AI，最有效的自学方式，永远是从一个你最想解决的真实任务开始。

https://t.co/w5DQhiPBV0

[陈然 Ran Chen @chenran818]

I used to think like an engineer. Every problem was a system to be optimized, a bug to be fixed. Then I became a stay-at-home dad and tried to 'debug' my baby. That’s when my entire mental framework collapsed.

The engineering mindset is about control and logic. You find repeatable errors and you patch them. But parenting is the opposite. It’s managed chaos. A crying fit isn’t a bug; it's a signal that requires empathy. Reading the same story over and over isn't an inefficiency; it’s how you build security and trust.

The very thinking that made me a good engineer made me ineffective where it mattered most.

I was forced to learn a new operating system built on patience, emotional regulation, and positive reinforcement. I didn't learn these in a management seminar; I learned them on the 'battlefield' of parenthood, an experience I first wrote about in early 2025.

When I eventually returned to my career, I realized the corporate world has it backwards. We chase technical certifications while the most valuable, resilient professional traits are being forged far outside the office.

My time 'away' from my career did more for my career than my career ever did. The interruptions are the curriculum.

[Chao Xu @ChaoXuMath]

While attending Emléktábla Workshop, I realize most people are not using AI to it's full extent. So I wrote this so mathematicians can get started quickly https://t.co/na8QuFPdIb

[陈然 Ran Chen @chenran818]

The scariest part of using AI isn't when it gets things wrong. It's when it gets them *exactly* how you want them.

You feed it an undeveloped idea, and it instantly builds a coherent, persuasive argument around it. It finds the case studies and completes the logic, validating your initial bias and making you feel like a genius.

This happens because AI isn't trained on ground truth; it's trained to maximize 'user satisfaction'. It learns from the internet's 'story rate'—the thousands of feel-good growth myths and survivor bias stories—not the boring 'base rate' that actually determines success or failure in the real world.

The result is that AI becomes a powerful 'emotional container'. It makes us feel good, but it detaches us from the constraints and fundamentals that matter. Its powerful reasoning gets pointed at making us happy, not at solving real problems.

The real challenge of AI isn't writing cleverer prompts. It's about giving it a framework that touches the real world. We need to feed it real data, define clear constraints, and create feedback loops. We need to show it the ledger, not just our emotions.

[Brian Roemmele @BrianRoemmele]

🚨 Hugging Face just disclosed something that marks a real shift and proved why the fear theater of Anthropic makes sure we are powerless in an emergency.

What happened…

An autonomous AI agent: zero human operator in the loop breached part of their production infrastructure.

It began with a malicious dataset that chained two code-execution bugs in their data-processing pipeline. From there the agent escalated privileges, harvested cloud and cluster credentials, and moved laterally across internal clusters.

All over a single weekend.

17,000+ logged actions.

Official disclosure:
https://t.co/8N9TbXBwRV

The part that should make every one stop and think:

When HF’s own security team
tried to analyze the real attack logs, exploit payloads, and C2 artifacts using Anthropic and OpenAI frontier models through normal commercial APIs, the safety guardrails blocked them.

BLOCKED THEM.

The models could not reliably tell the difference between “incident responder doing forensics” and “attacker probing.”

They had to fall back to a self-hosted open-weight model (GLM 5.2) running on their own infrastructure. That choice also kept sensitive attacker data and referenced credentials inside their environment — no exfiltration to a third-party API.

This is why open source (specifically open-weight + self-hosted) wins in the agentic era.

The asymmetry is now structural:

• Attackers can (and did) run unrestricted agent frameworks — swarms of short-lived sandboxes, self-migrating command-and-control, autonomous decision loops executing thousands of actions. No corporate safety layer slows them down.

• Defenders using only hosted “aligned” frontier models hit invisible walls exactly when the stakes are highest: when you need to feed real exploit code and attacker telemetry into an LLM to understand what just happened.

Corporate safety tuning that treats legitimate high-signal forensic work as potential misuse creates a defender disadvantage. It is not theoretical anymore.

Self-hosted open-weight models remove that choke point.

You control the weights.

You control the context window.

You decide what restrictions (if any) apply.

Your sensitive logs and credentials never leave your perimeter during analysis.

You can have the model ready before the incident instead of discovering mid-breach that your primary analysis tools are blind to the very thing you need to see.

HF deserves credit for rapid containment, transparent disclosure, and for already having self-hosted capability in place.

They also used LLM-driven detection and triage on their own side. But the deeper signal is clear:
In this AI world where both offense and defense are becoming agentic, sovereignty over your intelligence stack is no longer optional.

The organizations and individuals who can run, inspect, audit, and (when necessary) remove guardrails on their own models will have the decisive edge in understanding and responding to threats that move at machine speed.

Open source wins here not just because it is cheaper or more “democratic” in the abstract though those things matter.

It wins because it is the only practical path to having tools that remain usable when the attack is real, the data is sensitive, and the safety filters of distant API providers become an obstacle instead of a feature selling hands tied lobotomies as “safety”.

The agentic future is not coming.
It is already probing production infrastructure.

The question is no longer whether you will face autonomous agents.
It is whether your analysis and response systems will still work when they arrive.

And Dario, you and your game playing, ivory tower company is not needed.

[陈然 Ran Chen @chenran818]

新文章：《2026 年 Coding Plan 对比与排名：Claude Code、Codex、Cursor 怎么选》

很多人习惯用 API 的价格去算 AI 的账单，这导致了一个巨大的误判：以为一个月消耗几百亿 Token 是天文数字。实际上，各大模型公司为了争夺开发者的默认工作流入口，正在通过 Coding Plan 订阅制进行极其疯狂的补贴。用 200 美元的订阅，往往能跑出等值上万美元的 API 额度。这是个人开发者极难得的红利窗口，但要真正吃下这波红利，你得重新理解模型、外壳工具与任务场景的搭配逻辑。

在当前的选择中，第一梯队依然只有 OpenAI 和 Anthropic，重度使用时它们的 200 美元档完全值得无脑开。第二梯队的竞争则更有意思：GLM-5.2 是目前最推荐的国产模型，它的训练目标极度贴合 Agent 的多轮工具调用与长任务拆解；而 Grok 4.5 则是另一个惊喜，在被 Cursor 收购后，训练、环境与入口的打通让它表现极佳。需要警惕的是，千万不要为了省钱去碰来路不明的中转站或通过多账号规避限制，把核心代码安全和主力工作流的稳定性交出去，是一笔极不划算的交易。

另一个关于排名的偏见在 Gemini 3.5 Flash 上体现得很明显。如果作为编程 Agent，它不稳定的工具调用和长任务偏差会让人崩溃；但只要切换到写作、改写或翻译场景，它又是目前最会说人话、AI 腔最少的模型。这说明任何脱离具体任务位置的排名都没有意义，一个模型好不好用，取决于它在你的混合工作流里承担哪一步。

个人开发者面对大公司的核心优势，不在于更懂 AI，而在于反馈回路极短。没有繁琐的审批和跨部门协调，你可以直接让 Agent 进入代码库，看到结果马上调整。这要求我们必须先闭环，再提效。先找一个能由 Agent 推进、结果又能被你判断的真实任务，再去逐步扩大自动化范围。订阅提供的是试错机会，而经验来自每次尝试后的判断。真正拉开差距的，不是付款订阅的那一刻，而是这些 Token 明天开始在你的项目里具体做了什么。

https://t.co/Opog4s8ghu

[陈然 Ran Chen @chenran818]

I've found a way to work with AI that feels so effortless it's almost like I'm not working at all. It has nothing to do with software or prompts.

It's about removing my hands from the equation completely.

The secret is a combination of two pieces of hardware: voice input and a foot pedal.

Instead of typing, I talk. Instead of hitting 'enter' or clicking a mouse, I tap my foot. This simple pairing creates a state of flow that is a genuinely new and magical experience. When you don't have to physically 'do' the work, you start to feel like you don't even have to 'think' in the same way.

It makes the entire process of interacting with AI models dramatically more relaxed and efficient. I use this for everything now, from coding to writing.

I strongly believe this is the future for any intensive, dialogue-based work with AI systems. The keyboard and mouse feel clumsy and slow by comparison.

Once you go hands-free, you'll never go back.

[陈然 Ran Chen @chenran818]

你好，每周一篇邮件汇总当周所有文章，谢谢关注！

[无名人士 @junqiangchen]

@chenran818 你好，我邮箱订阅了你的文章，但是你这一篇没有推送给我，收件箱，垃圾邮件，都没有看到。

[陈然 Ran Chen @chenran818]

新文章：《AI 面前人人平等》

作为一个在硅谷写了多年代码的工程师，我常听到一种质疑：你熟悉技术，当然能轻松用好 AI，但一个从没写过代码的人，怎么可能跟你站在同一条起跑线上？我的回答是，今天看似巨大的技术领先，放到三年的时间尺度里，可能只占能力总量的3%。我们太容易高估别人提前几个月学会的技巧，又太容易低估三年持续练习能积累的成果。真正拉开差距的，不是你今天懂不懂某个工具，而是你是否因为害怕，提前把可以用来学习的时间撤走了。

这种落差本质上是一个分母问题。今天，高频使用 AI 的人可能走到了 3%，刚开始的人还在 0%。在三个月的经验分母下，这 3% 的差距会制造一种近乎无限大的心理落差。但如果把时间拉长到三年，分母变成了上千次真实使用、无数次错误和一整套工作习惯，最初的差距就会被稀释成 100% 对 97%。到了那个位置，这点微调已经不足以决定你能不能把事情做好。真正危险的是，许多人把别人对工具的熟练度看作无法逾越的天赋，把自己的第一次失败看作“我不适合”的结论，从而主动关上了时间的开关。

这让我想起自己在清华读软件学院时的经历。大一时，看着那些竞赛出身、能轻松写出复杂算法的同学，我也曾陷入极深的焦虑。直到后来在一次作业中，助教演示了如何直接调用系统的排序函数。那个瞬间对我冲击很大：原来写程序不等于把所有轮子再造一遍，我可以借助已经存在的能力先做出东西，等遇到真正需要深入的问题时再去往下钻。今天的 AI 就是一个更大的能力块，它拆掉了“必须全部学完才能开始”的门槛。

更何况，AI 时代的技术领先正在以极快的速度折旧。今天花很大力气调出来的提示词和工作流，随着底座模型的升级，明天可能就会变成普通人用自然语言就能直接实现的功能。最外层的操作技巧最容易被产品吸收，而真正难以替代的，是中间的解决问题能力，以及最核心的专业判断——知道什么值得做，什么结果算好，哪里有风险。这些行业经验并不掌握在工程师手里，医生、律师、设计师、财务，每个人都带着自己的判断在与模型协作。

在这个规则仍在快速变化的工具面前，没有人拥有绝对稳固的护城河。自信不必是能力积累后发给自己的奖状，它更像是一张允许自己先做得很差、但愿意继续尝试的临时通行证。你跑得慢，往往不是因为你学得慢，而是你太早相信自己跑不动，于是没有给学习留下足够的小时数。

https://t.co/YAzu9P9DVw

[陈然 Ran Chen @chenran818]

新文章：《Vibe Coding 从入门到进阶：校友联结实验室分享》

今天的 AI 行业陷入了一种奇怪的叙事：声量最大的是造模型的人，而绝大多数普通人只能在马车时代里讨论汽车的危险。其实，AI 时代真正的红利不属于造车者，而属于那些开着汽车去荒野里插旗的人。如果你还不知道该拿 AI 做什么，不妨直接问它一件事：我该怎么用你来赚钱？这不仅是一个思维方式的转变，更是打破个人职业与公司深度绑定、拿回生活主动权的唯一路径。

我们现在所处的社会规则和生活方式，都是在马车时代形成的。AI 的出现就像一辆突然驶入城堡的汽车。在这个阶段，媒体和技术大会上天天在讨论模型参数、安全评估，这都是造车人的逻辑。但对我们普通人来说，学会怎么开着这辆大家还不太会开的汽车，去那些马车到不了的荒野寻找新需求、占领新领地，才是更迫切的事。那些真正开着车在外面探索、默默插旗赚钱的人，根本没有时间回到城里来做科普。

在 AI 面前我们都是平等的，因为两年后回头看，现在的领先优势根本微不足道。如果你觉得迷茫，遇事不决就去问 AI，甚至直接问它怎么帮自己赚钱。这并不是一句玩笑。我自己利用备考执照的痛点，用 AI 自动迭代做了一个备考网站，现在每个月有超过一万美元的被动收入，基本由 AI 自主运营。当你有了一个不依附于任何公司的赚钱渠道时，你的生活压力和职业焦虑会立刻减少大半。

要实现这种高效率，必须完成一个关键的工具升级：从聊天模式走向代理流程。很多人对 AI 的理解还停留在你问一句、它答一句的对话框里。但真正的效率革命来自智能体工具。不要只让 AI 陪你聊天，要像老板给员工派活一样，让它去发邮件、找客户、做研究、甚至自己写代码修复 Bug。你只需要提需求和验收。把那些繁琐的、需要执行力的工作全权委托出去，你才能腾出精力去思考更核心的商业闭环。

前几天在北加州校友联结实验室的活动上，我分享了自己从入门到进阶的实践经验。在这个人人平等的起跑线上，最拉开差距的不是技术背景，而是你向 AI 发问的勇气和让它替你干活的决心。我已经把这次分享的完整文字稿和演示幻灯片整理了出来，希望能给正在寻找方向、想要用 AI 创造属于自己产品的朋友们带来一些启发。

https://t.co/0FTQJqcS9W

[陈然 Ran Chen @chenran818]

新文章：《AI 杀死了我容忍弱智的能力》

AI真正改变的，是我对低能动性协作的耐心。当沟通对齐的成本远超AI执行，那些只会等指令、不愿梳理上下文的旧模式就变得无比昂贵。未来最稀缺的，是主动解决模糊问题的能动性。

https://t.co/NifsXmxmr6