新智元报道
就在刚刚，Claude Code核心作者正式离职Anthropic！
Thariq Shihipar是Claude Code团队的核心灵魂人物、被开发者社区尊称为「解释之王」。
他在X上正式发布了一封长文告别信。
在短短45分钟，Claude团队最知名的两位成员相继离职。
到底发生了什么？
「在加入Claude Code 一年多，并亲眼见证它的规模扩展到远超我曾经的想象之后，我离开了 Anthropic。」
在告别信开篇，Thariq这样写道。
作为陪伴 Claude Code 从有争议的新工具到「无数开发者默认编码方式」的核心推手，他在最巅峰时刻，选择了离开。
紧接着，他抛出了整封信最核心的价值观：「向上竞赛」。
我真的非常热爱在 Anthropic的工作，深度认同公司的使命。最重要的是，我坚信『向上竞赛』的理念。如果我们努力且负责任地工作，我们就能抬高行业的标准，并激发其他人走向更好的结果。如果我们都能互相推着对方变得更好，事情就会变得极其美好。我将永远带着这种精神前行。
在大模型疯狂内卷的今天，Thariq 犹如一股清流。他眼里的AI竞赛，不是零和博弈，而是一场提高全人类标准、倒逼彼此变得更卓越的事业。
对于同事们，他给予了赞美——
「我在 Anthropic 的同事们，是我一生中遇到过最有才华、最周到、最能干且最真诚的人。我很高兴现在是他们在掌舵，我深深地信任他们。」
在信的结尾，他将聚光灯交还给了团队，一口气@了多位 Anthropic 的核心开发者，呼吁大家继续关注他们。
虽然告别信十分官方，但Thariq在评论区，还是发出了一个亮点评论：「等等，是不是只有我一个人觉得 Opus 5.5 变蠢了？」
Thariq的履历，仿佛开了挂。
他不只是传统研究员，而是一个横跨学术界、顶级大厂与硅谷创投圈的「六边形战士」。
早在多伦多大学攻读工程科学本科期间，他就展露出了极强的产品直觉，联合创办了聚合消息通知的插件 Chime，并在2013年直接被 HubSpot 收购。
随后，他进入科技极客的梦想圣地——MIT Media Lab深造，并担任研究科学家。
在那里，他联合主导了颠覆传统PDF的交互式学术出版平台 PubPub，试图用动态程序重塑人类知识的展现形式。
随后，带着对复杂交互的深刻理解，Thariq 一头扎进创业赛道。
他创办了浏览器端游戏创作平台 One More Multiverse，不仅成功杀入全球顶尖孵化器YC，更是一手将平台做到了百万级用户规模。
这款灵感来自跑团和Roblox的产品，斩获了超1700万美元的融资。
直到2025年，在经历了一段专注AI可解释性研究的Fellowship沉淀后，这位CEO，选择以「技术人员」的身份加入 Anthropic。
可以看出，Thariq 的每一步都在解决同一个核心问题：如何让极度复杂的系统，变成普通人一眼就能看懂、上手就能用的绝佳体验？
正是这种「游戏CEO」与「MIT学者」的双重视角，让他在加入 Claude Code 团队后，敏锐地察觉到了当前AI开发的盲点——不要再执迷于给模型无脑堆算力了，应该重塑的，是人类与AI的交互。
作为前 YC 明星创业者、MIT Media Lab 的研究员，Thariq 带着极强的产品直觉改造了 Claude Code。
此前，他曾输出以下观点。
真相一：大模型的智力已经「过剩」，算力不再是第一瓶颈
Thariq 曾尖锐地指出：现在的强模型（如 Opus 5）其实已经非常聪明了，真正的瓶颈根本不是原始算力，而是交互接口与人类自身的理解力！
能力冗余，是指一个模型在配备合适的智能体运行框架（Harness）、获得环境访问权限和代码执行能力后，能够完成的任务，远远超出了它仅靠普通对话、缺乏外部工具支持时所展现出的能力。
真相二：强模型时代，请勇敢砍掉80%的提示词！
Thariq 团队在内部实验中，对强模型直接砍掉了超80%的系统提示词，结果发现代码能力根本没有下降！
链接：https://claude.dev/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models/
他主张：过度约束反而会限制更强模型的发挥。把规则做成按需调用的模块化「Skills」，给AI留出自由发挥的空间。
真相三：告别 Markdown，HTML 才是高密度交互的未来
程序员最爱的 Markdown，正在成为人机协作的阻碍。当 AI 输出的代码和计划越来越长，人类往往会因为阅读疲劳而「盲目点击同意」。
Thariq 断言，带图表、带颜色差异比对、可交互的 HTML 界面，才是未来能让人类真正保持判断力的交互方式。
真相四：未来人类的终极职业——「算力分配者」
别怕AI写代码比你快。
在未来，你的角色将变成「计算资源分配者」。你的核心价值在于：决定把昂贵的 Agent 算力投入到哪个方向，并在关键节点利用主动提问工具（如 AskUserQuestion）消除盲点，时刻把控方向盘。
或许，Thariq Shihipar已经找到「向上竞赛」的方向，准备好下一个创业方向了？
参考资料：
编辑：Aeneas
不喜欢
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。
New Ji-won.
Just now, Claude Code's core author officially separated from Anthropic!
Thariq Shihipar is the core soul figure of Claude Code's team and is known by the developer community as "the king of interpretation".
On X, he officially wrote a long farewell letter to publish.
In just 45 minutes, the two most prominent members of the Claude team separated.
What the hell happened?
"I left Anthropic after joining Claude Code for more than a year and witnessing its expansion far beyond what I had ever imagined."
At the beginning of the farewell letter, Thariq wrote.
As the core pusher to accompany Claude Code from the controversial new tool to the "Numerous Developers' Default Coding Method", he chose to leave at the peak.
Then he threw out the core value of the letter, "to race up."
I really loved working in Anthropic, deeply endorsing the company's mission. And above all, I believe in the idea of going up. If we work hard and responsibly, we can raise industry standards and inspire others to move towards better outcomes. If we can push each other to get better, things will get better. I will always walk with that spirit.
Today, in LLM's craziness, Thariq is like a clear stream. His vision of the AI competition was not a zero-sum game, but a cause that raised the standards of humanity and forced each other to become more superior.
To his colleagues, he praises...
"My colleagues in Anthropic are the most talented, thoughtful, capable and sincere people I've ever met in my life. I'm glad they're at the helm now, and I trust them deeply.”
At the end of the letter, he returned the spotlight to his team, with a breath of @ many core developers of Anthropic, calling for continued attention to them.
Although the farewell letter was official, Thariq sent a bright comment in the comment area: "Wait, am I the only one who thinks that Opus 5.5 is stupid?"
Thariq's résumé seems to be on the loose.
He is not just a traditional researcher, he is a " hexagonal warrior" who crosses academia, top factory and Silicon Valley.
As early as he was studying undergraduate engineering science at the University of Toronto, he showed a strong product instinct, co-founding Chime, a polymer message plugin, which was acquired directly by HubSpot in 2013.
He then went to the dream site of the technocrats, MIT Media Lab, and became a research scientist.
There, he co-led an interactive academic publishing platform that subverted the traditional PDF, PubPub, trying to reshape the presentation of human knowledge with dynamic processes.
Then, with a deep understanding of complex interactions, Thariq stuck his head into the entrepreneurial track.
He created One More Multiverse, the browser-side game creation platform, which not only succeeded in killing the world ' s top incubator YC, but also made the platform a million-scale user.
This was inspired by the products of the Run and Roblox, which secured over $17 million in financing.
Until 2025, after a fall of Phillowship, which focused on an AI interpretive study, the CEO chose to join Anthropic as a "technical" person.
As can be seen, Thariq is addressing the same core issue at every step: how can an extremely complex system become a perfect experience that ordinary people can understand and use at once?
It's this double view of the game CEO and the MIT scholar who, after joining the Claude Code team, is keenly aware of the blind spots of the current AI development. - Don't be obsessed with the brainless calculation of models. It's about human interaction with AI.
As a former YC star entrepreneur and researcher of MIT Media Lab, Thariq transformed Claude Code with a strong product instinct.
He had previously given the following view.
Truth One: LLM's intelligence has become "over".
Thariq once pointed out that the powerful models of today (e.g. Opus 5) are actually very smart, and the real bottlenecks are not primitive calculus, but rather interactive interfaces with human own understanding!
Capability redundancy is the ability of a model to accomplish tasks that, with the appropriate AI agent operating framework (Harness), environmental access to permission and code implementation capacity, go well beyond what it has demonstrated when it is supported by ordinary dialogue and lack of external tools.
Truth II: The Age of Strong Models, please bravely cut down 80% of prompt!
In an internal experiment, the Tharik team cut off over 80% of the system prompt directly against the strong model, and found that the code capability had not decreased at all!
Links: https://claude.dev/blog/the-new-rules-of-content-engineering-for-claude-5-gender-models/
He argued that excessive restraint would instead limit the use of stronger models. The rules are modularized to "Skills" as needed, leaving free space for AI.
Truth III: Goodbye, Markdown, HTML is the future of high-density interaction.
The programmer's favorite Markdown is becoming an obstacle to human collaboration. As the codes and plans for AI output grow longer and longer, humans tend to "black click yes" for reading fatigue.
Tharik asserted that the interactive HTML interface with graphics, colour contrasts, is the way in which humans can truly maintain judgement in the future.
Truth IV: The ultimate profession of future human beings - the "calculator"
Don't be afraid AI.'s writing codes faster than you.
In the future, your role will become "calculating resource allocation." Your core value is to decide which direction to invest in expensive Agent algorithms and to use proactive question tools (e.g. AskUserQuestion) at key nodes to remove blind spots and keep the prosecution on track.
Maybe Thariq Shihipar has found his way up to the competition and is ready for his next start-up?
References:
Editor: Aeneas
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.