Move Over, OpenAI: How the Startup Behind Cursor Became the Hottest, Vibiest Thing in AI
让位于 OpenAI：Cursor 背后的初创公司如何成为 AI 领域最热门、最炫的事物

Stripe CEO Patrick Collison is another Anysphere investor. “I’ve been extraordinarily impressed by the Cursor founders,” Collison told me, describing himself as “very bullish” on the company. “They remind me of the very best early-stage teams that I’ve seen.”
Stripe 首席执行官帕特里克·科利森也是 Anysphere 的投资者。“我对 Cursor 的创始人印象非常深刻，”科利森告诉我，形容自己对这家公司“非常看好”。“他们让我想起我见过的最优秀的早期团队。”

With investors clamoring for a stake in the company, Anysphere’s valuation has shot up from $400 million in May to $2.5 billion in November. Now Truell and his co-founders are talking to Andreessen Horowitz, Thrive Capital and other investors about even more money, which could further lift Anysphere’s valuation to $10 billion. (Accel is also involved in those discussions—a previously unreported detail— one of only a few new investors, another indication of the enormous interest surrounding Anysphere.)
随着投资者争相希望在公司中占有一席之地，Anysphere 的估值从 5 月的 4 亿美元飙升至 11 月的 25 亿美元。现在，Truell 和他的联合创始人正在与 Andreessen Horowitz、Thrive Capital 和其他投资者洽谈更多资金，这可能会进一步将 Anysphere 的估值提升至 100 亿美元。（Accel 也参与了这些讨论——这是一个之前未报道的细节——是少数几个新投资者之一，另一个迹象表明 Anysphere 周围的巨大兴趣。）

Meantime, Anysphere’s annual recurring revenue has surpassed $200 million as of this month—a figure that also hasn’t been previously reported—according to a person with direct knowledge of the company’s financials. It’s an astonishing amount considering that Cursor’s launch came in January 2023. It all adds up to a stunning reality: Anysphere is one of the fastest-growing startups ever, and what Truell and his co-founders have built is a bona fide AI rocket ship with a trajectory that stands out even among other AI startups hurtling into the stratosphere.
与此同时，Anysphere 的年度经常性收入截至本月已超过 2 亿美元——这一数字也是之前未报道的——根据一位对公司财务状况有直接了解的人士。考虑到 Cursor 于 2023 年 1 月推出，这个数字令人惊讶。这一切加起来形成了一个惊人的现实：Anysphere 是有史以来增长最快的初创公司之一，Truell 和他的联合创始人所建立的是真正的 AI 火箭船，其轨迹在其他迅速进入太空的 AI 初创公司中也显得格外突出。

And there’s an indication Anysphere may not suffer from the same fate that’s befalling many other AI startups: fast revenue growth, fueled by Silicon Valley’s eagerness to experiment with AI, that disappears after customers play around with a tool and decide it’s nothing special.
而且有迹象表明，Anysphere 可能不会遭遇许多其他 AI 初创公司所面临的相同命运：快速的收入增长，得益于硅谷对 AI 实验的热情，但在客户玩过一个工具后决定它并没有什么特别之处。

Anysphere’s net revenue retention rate for customers who joined in March 2024 is 250%, according to the person familiar with the company’s financials, a metric that measures a company’s ability to hold onto existing customers and increase their spending. For perspective, a one-year net retention rate of 115% is wonderful.
根据一位熟悉该公司财务状况的人士，Anysphere 在 2024 年 3 月加入的客户的净收入留存率为 250%，这一指标衡量了一家公司留住现有客户并增加其支出的能力。为了提供一个参考，一年的净留存率达到 115%是非常出色的。

In another sign of Cursor’s rising profile, other startups have lately taken to pitching themselves as the Cursor of their own industry, hoping to signal their own promising take on automation technology. Recent Y Combinator classes, for example, included Onlook, a Cursor for designers; Lightski, a Cursor for engineering documents; and Artifact, a Cursor for hardware design.
在 Cursor 日益崛起的另一个迹象中，其他初创公司最近开始自我推销，称自己是各自行业的 Cursor，希望能传达他们对自动化技术的有希望的看法。例如，最近的 Y Combinator 班级中包括了 Onlook，一个为设计师服务的 Cursor；Lightski，一个为工程文档服务的 Cursor；以及 Artifact，一个为硬件设计服务的 Cursor。

“In the next five years, we really believe it will be possible to build a tool that automates almost all of software engineering,” Truell told a crowd gathered in downtown San Francisco for the AI Engineer World’s Fair last June. In those remarks, he described a future in which “individual engineers can build systems that are much more complex than even entire engineering teams” can today.
“在接下来的五年里，我们真的相信可以构建一个几乎可以自动化所有软件工程的工具，”Truell 在去年六月于旧金山市中心举行的 AI 工程师博览会上对聚集的人群说道。在这些讲话中，他描述了一个未来，“个人工程师可以构建的系统比今天整个工程团队能够构建的系统要复杂得多。”

Why all the hype about Cursor? It can’t magically turn novice coders into genius ones overnight. But developers see it as a significant leap forward in code-editing software—perhaps on par with the impact that tools like Microsoft Word had on writing and editing in the 1980s and ’90s.
为什么大家对 Cursor 如此 hype？它不可能在一夜之间将初学者变成天才。但开发者们认为它是代码编辑软件的一次重大飞跃——或许与 1980 年代和 1990 年代微软 Word 对写作和编辑的影响相当。

Like ChatGPT, it allows coders to work using natural-language prompts. For example, a programmer can highlight a chunk of code and then ask Cursor to explain what the code does. But Cursor has become beloved by coders almost overnight chiefly for its ability to predict what code they want to write or edit, as well as its knack for spotting errors and for integrating right into their existing workflows—something other tools have struggled to do. By building a smoother, more automated coding workspace, Cursor says it has attracted over 30,000 enterprise customers.
像 ChatGPT 一样，它允许程序员使用自然语言提示进行工作。例如，程序员可以突出显示一段代码，然后要求 Cursor 解释该代码的功能。但 Cursor 凭借其预测程序员想要编写或编辑的代码的能力，以及发现错误和无缝集成到现有工作流程中的能力，几乎一夜之间就赢得了程序员的喜爱——这是其他工具所难以做到的。通过构建一个更流畅、更自动化的编码工作空间，Cursor 表示它吸引了超过 30,000 家企业客户。

More broadly, Cursor fits into a significant shift within Silicon Valley toward automating more of the lengthy process of software development. Salesforce CEO Marc Benioff has vowed not to hire any software engineers at all this year after his company saw AI increase his engineers’ productivity by 30%. Google’s Sundar Pichai has said his engineers are using AI to write a quarter of all new code, while Y Combinator’s Garry Tan said 25% of startups from a recent group in his accelerator had written nearly all their code using AI. And Meta Platforms’ Mark Zuckerberg has predicted the emergence of AI equipped with the same skills as a “good midlevel engineer” within the next nine months.
更广泛地说，Cursor 适应了硅谷在自动化软件开发漫长过程中的重大转变。Salesforce 首席执行官 Marc Benioff 发誓今年不会雇佣任何软件工程师，因为他的公司发现 AI 提高了工程师 30% 的生产力。谷歌的 Sundar Pichai 表示，他的工程师正在使用 AI 编写四分之一的新代码，而 Y Combinator 的 Garry Tan 则表示，他加速器最近一组中的 25% 的初创公司几乎全部使用 AI 编写了他们的代码。Meta Platforms 的 Mark Zuckerberg 预测，在接下来的九个月内，将出现具备“优秀中级工程师”相同技能的 AI。

Still, any number of bigger companies could interrupt Anysphere’s ascent, such as OpenAI with ChatGPT, Microsoft with GitHub Copilot or Anthropic with its Cursor-like tool Claude Code. Given how Cursor has completely captured Silicon Valley’s interest, it would be impossible for these much larger companies to ignore how Cursor has captured the market’s attention.
尽管如此，任何数量的大公司都可能打断 Anysphere 的上升，比如 OpenAI 的 ChatGPT、微软的 GitHub Copilot 或 Anthropic 的类似 Cursor 的工具 Claude Code。考虑到 Cursor 已经完全吸引了硅谷的关注，这些更大的公司不可能忽视 Cursor 如何吸引市场的注意。

Anysphere co-founder Aman Sanger acknowledged the reality of competition in a rare public appearance on the “Lex Fridman Podcast” in October—and expressed blithe confidence in the startup’s ability to stay ahead. “The Cursor a year from now will need to make the Cursor of today look obsolete,” he said. “Microsoft has done a number of fantastic things, but I don’t think they’re in a great place to really keep innovating and pushing on this in the way that a startup can.”
Anysphere 联合创始人 Aman Sanger 在十月的“Lex Fridman Podcast”上罕见公开露面，承认了竞争的现实，并对初创公司的领先能力表示轻松自信。他说：“一年的 Cursor 需要让今天的 Cursor 看起来过时。”他补充道：“微软做了很多出色的事情，但我认为他们并不处于一个真正能够以初创公司那样持续创新和推动的好位置。”

Perhaps it’s no surprise that the young men disrupting the world of software coding have all been coders themselves since childhood.
或许并不奇怪，这些年轻人正在颠覆软件编码世界，他们从小就是程序员。

Truell, son of two journalists, spent his childhood fascinated by robotics and working on coding games. In high school, he co-created a programming game called Halite, in which thousands of players controlled a bot using different programming languages. (It became a hit online among teenage programmers.) And while attending New York’s elite Horace Mann School, he and his friend Sanger—one of his eventual Cursor co-founders—both excelled in math and science, and were accepted at the Massachusetts Institute of Technology.
特鲁尔，son 两位记者中的一位，从小对机器人技术充满了兴趣，并致力于编写游戏。在高中时，他共同创作了一款名为 Halite 的编程游戏，数千名玩家使用不同的编程语言控制一个机器人。（这款游戏在青少年程序员中迅速走红。）在纽约的精英霍拉斯·曼学校就读期间，他和他的朋友桑格——他未来的 Cursor 联合创始人之一——在数学和科学方面都表现出色，并被麻省理工学院录取。

While at MIT, Sanger and Truell met up with their other two future co-founders, who also were studying computer science and math: Sualeh Asif, the former valedictorian of his Karachi, Pakistan, secondary school, who had represented his county at the International Mathematical Olympiad; and Arvid Lunnemark, a former Math Olympiad champion.
在麻省理工学院期间，Sanger 和 Truell 结识了他们的另外两位未来联合创始人，他们也在学习计算机科学和数学：Sualeh Asif，曾是巴基斯坦卡拉奇中学的优秀毕业生，曾代表他的国家参加国际数学奥林匹克；以及 Arvid Lunnemark，一位前数学奥林匹克冠军。

On summer break in 2019, Truell, who had just finished his freshman year and was interning at Google, met investor Ali Partovi for a coffee at the Computer History Museum in Mountain View, Calif. Truell was hoping to win a spot in Partovi’s prestigious Neo Scholars program, and at the meeting, Partovi asked him to take a coding test—using just pen and paper. Truell finished the assessment in record time: 15 minutes.
在 2019 年夏季假期，刚刚完成大一的 Truell 在加利福尼亚州山景城的计算机历史博物馆与投资者 Ali Partovi 见面喝咖啡。Truell 希望能获得 Partovi 的著名 Neo Scholars 项目的名额，在会议上，Partovi 要求他进行一次编码测试——仅使用纸和笔。Truell 以创纪录的时间完成了评估：15 分钟。

“Pretty much from the first meeting, I said to myself, ‘I would invest in anything this guy does,’” recalled Partovi, who has earned a reputation for spotting talent early, with previous investments in Facebook and Dropbox.
“几乎从第一次会议开始，我就对自己说，‘我会投资于这个家伙做的任何事情，’”Partovi 回忆道，他因早期发现人才而声名显赫，曾投资于 Facebook 和 Dropbox。

Still, Partovi’s program mainly accepted juniors and seniors, and he didn’t want to bend the rules for Truell. But the two stayed in touch, and a year later, both Truell and Sanger became Neo Scholars, earning internships at Octant, a biotech company, and You.com, an AI search engine startup, respectively.
尽管如此，帕尔托维的项目主要接受大三和大四学生，他不想为特鲁尔破例。但两人保持联系，一年后，特鲁尔和桑格都成为了 Neo Scholars，分别在生物技术公司 Octant 和人工智能搜索引擎初创公司 You.com 获得了实习机会。

As Truell, Sanger, Asif and Lunnemark neared graduation, the friends were weighing whether to pursue master’s degrees to do more research in AI and computer science. If they wanted to go even deeper into AI, the traditional route would have been to remain in academia.
当特鲁尔、桑格、阿西夫和伦内马克接近毕业时，朋友们在考虑是否攻读硕士学位，以便在人工智能和计算机科学领域进行更多研究。如果他们想更深入地研究人工智能，传统的途径就是留在学术界。

But in 2021, things were changing in the private sector, and the founders sensed they didn’t need to stay in school. OpenAI was quickly gaining momentum and had released a groundbreaking paper about scaling laws—measures of how quickly AI can develop—while Microsoft had recently debuted GitHub Copilot.
但在 2021 年，私营部门发生了变化，创始人们意识到他们不需要继续上学。OpenAI 迅速获得了动力，并发布了一篇关于扩展法则的开创性论文——衡量人工智能发展速度的指标——而微软最近推出了 GitHub Copilot。

After all four co-founders graduated, Partovi stepped in with seven figures’ worth of seed money for Truell and Sanger. Their initial focus was on AI-based software for mechanical engineers working on three-dimensional designs. But they struggled to make software they felt could be turned into a lucrative business, so they scrapped the concept and decided to start over by joining the debut of OpenAI’s startup accelerator program, Converge.
在四位联合创始人毕业后，Partovi 为 Truell 和 Sanger 提供了七位数的种子资金。他们最初的重点是为从事三维设计的机械工程师开发基于人工智能的软件。但他们在制作能够转化为盈利业务的软件方面遇到了困难，因此他们放弃了这个概念，决定重新开始，加入 OpenAI 的初创企业加速器项目 Converge。

In 2022, and during the OpenAI accelerator program, Truell and Sanger lived with their MIT classmates turned friends, Asif and Lunnemark. As the four friends joined forces to do the earliest work on what would become Cursor, they also got to be part of a small group of people who received early access to GPT-4, OpenAI’s latest model, before its public release.
在 2022 年，在 OpenAI 加速器项目期间，Truell 和 Sanger 与他们的 MIT 同学兼朋友 Asif 和 Lunnemark 同住。当四位朋友联手进行 Cursor 的早期工作时，他们也成为了一小群在公开发布之前获得 OpenAI 最新模型 GPT-4 早期访问权限的人。

It didn’t take long for developers to get excited. Sam Seely, the co-founder of Knock, a New York–based company that makes product software for other companies, first began experimenting with Cursor six months ago and quickly found himself intrigued by its ability to edit code with natural language.
开发者们很快就兴奋起来了。Knock 的联合创始人 Sam Seely，这是一家位于纽约的公司，为其他公司制作产品软件，他在六个月前首次开始尝试 Cursor，并迅速对其用自然语言编辑代码的能力产生了兴趣。

“It’s a breakthrough,” Seely said. “I haven’t used Copilot since.”
“这是一个突破，”Seely 说。“我从那以后就没有使用 Copilot 了。”

As Cursor has become the toast of the Valley, Truell and his co-founders, who kept their team small in the beginning, have more recently rushed to bolster the startup’s ranks. Over the past year, Anysphere has grown from 10 people to over 50. The company’s founders use a rigorous recruitment process: Prospective engineers go through three technical interviews and then a two-day stint working from the Anysphere offices with its staff.
随着 Cursor 成为硅谷的明星，Truell 和他的联合创始人们在最初保持团队规模较小的情况下，最近急于壮大初创公司的队伍。在过去的一年里，Anysphere 的员工人数从 10 人增长到 50 多人。公司的创始人采用严格的招聘流程：潜在的工程师需要经过三轮技术面试，然后在 Anysphere 的办公室与员工一起工作两天。

Those time-consuming trials happen in the company’s headquarters in North Beach, a historically Italian neighborhood in San Francisco. As the company expands, the team is considering making a mini-campus area in North Beach, with several Anysphere offices in separate buildings on the same block.
这些耗时的试验在公司位于旧金山北滩的总部进行，北滩是一个历史悠久的意大利社区。随着公司的扩展，团队正在考虑在北滩建立一个迷你校园区域，在同一街区内设立多个 Anysphere 办公室。

Sanger has publicly boasted that the startup hasn’t spent a dollar on marketing. It has done some video conference panels with users to answer questions about the software, and its co-founders are thinking about other low-cost ways to further bolster their company’s presence, like hackathons and in-person meetups for Cursor superfans. Still, they seem to know they need greater professional help in getting Cursor into the marketplace, and lately they have hired more people focused on sales and marketing the product. (Today, around 80% of the company’s staff are researchers and engineers.)
Sanger 公开表示，这家初创公司在营销上没有花费一美元。它与用户进行了一些视频会议，回答有关软件的问题，其联合创始人正在考虑其他低成本的方式来进一步提升公司形象，比如黑客马拉松和 Cursor 超级粉丝的面对面聚会。不过，他们似乎知道在将 Cursor 推向市场方面需要更多的专业帮助，最近他们雇佣了更多专注于销售和营销产品的人。（目前，公司的员工中大约 80%是研究人员和工程师。）

Anysphere has built Cursor using large language models from other AI companies, first getting its start from OpenAI models and then benefiting from Anthropic’s coding-savvy models, including Claude 3.5 and the newer Claude 3.7 Sonnet. But as Anysphere has expanded, its relationship with Anthropic has shifted from that of friends to potential rivals.
Anysphere 使用其他 AI 公司的大型语言模型构建了 Cursor，最初从 OpenAI 模型开始，然后受益于 Anthropic 的编码能力强的模型，包括 Claude 3.5 和更新的 Claude 3.7 Sonnet。但随着 Anysphere 的扩展，它与 Anthropic 的关系已经从朋友转变为潜在的竞争对手。

In February, Anthropic launched Claude Code, a possible Cursor competitor. The two tools have some technical differences, but developers say they’re similar enough that Claude Code could steal some customers away from Cursor. OpenAI has also released a product that allows users to use ChatGPT to write code, Canvas.
在二月份，Anthropic 推出了 Claude Code，一个可能的 Cursor 竞争对手。这两个工具在技术上有一些差异，但开发者表示它们足够相似，以至于 Claude Code 可能会从 Cursor 那里抢走一些客户。OpenAI 也发布了一款产品，允许用户使用 ChatGPT 编写代码，名为 Canvas。

For now, developers think Cursor remains a superior product—particularly thanks to its ability to predict code edits, spot errors and keep engineers in the loop while also automating some of their work.
目前，开发者认为 Cursor 仍然是一个更优秀的产品——特别是由于它能够预测代码编辑、发现错误并让工程师保持在信息循环中，同时还自动化他们的一部分工作。

Anthropic’s connection with Anysphere has been fruitful for both companies, Anthropic’s product lead Michael Gerstenhaber told me. Anysphere has passed on customer feedback about the models to Anthropic, which has used those notes to further hone its technology and fuel research work.
Anthropic 与 Anysphere 的合作对两家公司都很有成效，Anthropic 的产品负责人 Michael Gerstenhaber 告诉我。Anysphere 已将客户对模型的反馈传递给 Anthropic，后者利用这些反馈进一步完善其技术并推动研究工作。

“We need to keep our position [as] the best model in the world for coding,” Gerstenhaber said. “And this is how we do it: it’s by feeling pressures from Cursor’s customers to get them the best model, and finding all the ways we can to learn about that.”
“我们需要保持我们在全球编码模型中的最佳地位，”Gerstenhaber 说。“而这就是我们做到这一点的方法：通过感受到 Cursor 客户的压力，给他们提供最佳模型，并寻找所有可以了解这一点的方法。”

Still, some in Silicon Valley see a possible brewing war. “The obvious threat to any of these companies is competition from the foundational models,” said Ethan Kurzweil, a founder of Chemistry, a VC firm, who doesn’t currently have a stake in either Anthropic or Anysphere. The question for large language models is, he says, is simple: “Is what Cursor is doing a threat, or do you want to enable them?”
尽管如此，硅谷的一些人看到了一场可能正在酝酿的战争。“这些公司的明显威胁是来自基础模型的竞争，”Chemistry 的创始人 Ethan Kurzweil 说，他目前在 Anthropic 或 Anysphere 都没有股份。他表示，对于大型语言模型，问题很简单：“Cursor 所做的事情是威胁，还是你想要支持他们？”

Gerstenhaber downplayed any sense of a corporate rivalry between Anysphere and Anthropic, describing Claude Code as merely a “research project.”
Gerstenhaber 轻描淡写地表示 Anysphere 和 Anthropic 之间并不存在企业竞争关系，将 Claude Code 仅仅描述为一个“研究项目”。

Microsoft, meanwhile, introduced a free version of GitHub Copilot in December, seemingly in an attempt to keep developers from continuing to flock over to Cursor or other alternatives.
微软在 12 月推出了 GitHub Copilot 的免费版本，似乎是为了防止开发者继续涌向 Cursor 或其他替代品。

Another issue for Anysphere is making sure Cursor can withstand the pressure of virality. The company’s own online forum is littered with complaints about Cursor failing or breaking down. Some of those issues may be attributable to Cursor’s reliance on Anthropic’s technology to operate—Anthropic has struggled to keep up with the demand.
Anysphere 面临的另一个问题是确保 Cursor 能够承受病毒式传播的压力。该公司的在线论坛上充满了关于 Cursor 故障或崩溃的投诉。这些问题中的一些可能归因于 Cursor 依赖 Anthropic 的技术来运行——Anthropic 一直难以跟上需求。

Truell and his co-founders are focused on developing their own frontier models to help power new features in Cursor. They have recently prioritized building their own LLMs and are aggressively hiring people with experience in post-training, which involves fine-tuning a model as it tries to make its own AI more sophisticated. Currently, Anysphere runs over 200 million queries a day using its own models and has procured thousands of GPUs to handle this work, according to someone with knowledge of the matter.
Truell 和他的联合创始人专注于开发自己的前沿模型，以帮助推动 Cursor 的新功能。他们最近优先考虑构建自己的LLMs，并积极招聘具有后期训练经验的人，这涉及在模型尝试使其自己的 AI 变得更复杂时进行微调。根据一位知情人士的说法，目前 Anysphere 每天使用自己的模型处理超过 2 亿个查询，并已采购了数千个 GPU 来处理这项工作。

The capacity problems are unquestionably a real concern, but investor Ali Partovi for one has the fullest confidence in the Anysphere founders’ ability to figure it out, likening their abilities to those of “Bill Gates and Larry Ellison and Larry Page and Sergey Brin and Mark Zuckerberg.”
容量问题无疑是一个真正的担忧，但投资者阿里·帕尔托维对 Anysphere 创始人解决这个问题的能力充满信心，将他们的能力比作“比尔·盖茨、拉里·埃里森、拉里·佩奇、谢尔盖·布林和马克·扎克伯格”。