← 回總覽

Cursor 联合创始人解释技术选型路径,事件定性为「沟通事故」

📅 2026-03-21 04:58 宝玉 人工智能 4 分鐘 4091 字 評分: 82
Cursor Kimi K2.5 Aman Sanger 强化学习 持续预训练
📌 一句话摘要 Cursor 联合创始人 Aman Sanger 详细解释选择 Kimi K2.5 的技术原因(困惑度指标最强),承认发布博客未提基座模型是失误,并定性该事件为「沟通事故」而非许可证纠纷。 📝 详细摘要 该推文是接着上一条的深入分析。推文引用了 Cursor 联合创始人 Aman Sanger 的详细技术回应:团队评估了大量基座模型,Kimi K2.5 在困惑度(perplexity)指标上表现最强,之后 Cursor 在此基础上做了持续预训练和 4 倍算力规模的强化学习。Aman Sanger 承认在发布博客中没有提到 Kimi 基座是「一个失误」,并表示下次会改正。作者
Skip to main content ![Image 1: LogoBestBlogs](https://www.bestblogs.dev/ "BestBlogs.dev")Toggle navigation menu Toggle navigation menuArticlesPodcastsVideosTweetsSourcesNewsletters

⌘K

Change language Switch ThemeSign In

Narrow Mode

Cursor Co-Founder Explains Technical Selection Path; Incident Labeled as 'Communication Mishap'

Cursor Co-Founder Explains Technical Selection Path; Incident Labeled as 'Communication Mishap'

![Image 2: 宝玉](https://www.bestblogs.dev/en/tweets?sourceId=SOURCE_2c263f) ### 宝玉

@dotey

紧接着,Cursor 联合创始人 Aman Sanger 也发了一条详细回应。他解释了技术路径:团队评估了大量基座模型,Kimi K2.5 在困惑度指标上表现最强,之后 Cursor 在此基础上做了持续预训练和 4 倍算力规模的强化学习。x.com/amanrsanger/st…

他承认在发布博客中没有提到 Kimi 基座是"一个失误",并表示下次会改正。

到这里,这件事的性质已经从"许可证纠纷"变成了"沟通事故"。

几个关键变化值得注意。首先,Moonshot 官方帖子特意补了一句"Cursor 通过 Fireworks 的授权商业合作伙伴关系使用 Kimi K2.5",这等于官方确认了授权链条的合法性。

之前 Yulun Du 质问的"为什么不遵守许可证、不付费",现在被公司层面正式否认了。他和另外两名员工删帖这件事,回过头来看,大概率是公司内部沟通滞后,员工不知道有商业合作协议在先。Show More

!Image 3: Aman Sanger

#### Aman Sanger

@amanrsanger · 6h ago

We've evaluated a lot of base models on perplexity-based evals and Kimi k2.5 proved to be the strongest!

After that, we do continued pre-training and high-compute RL (a 4x scale-up).

The combination of the strong base, CPT and RL, and Fireworks' inference and RL samplers make Composer-2 frontier level.

It was a miss to not mention the Kimi base in our blog from the start. We'll fix that for the next model.Show More

107

85

1,509

197K

Mar 20, 2026, 8:58 PM View on X

3 Replies

0 Retweets

7 Likes

5,222 Views ![Image 4: 宝玉](https://www.bestblogs.dev/en/tweets?sourceid=2c263f) 宝玉 @dotey

One Sentence Summary

Cursor co-founder Aman Sanger explains in detail why Kimi K2.5 was chosen (strongest performance on perplexity metrics), acknowledges the mistake of not mentioning the Kimi base model in the release blog, and labels the incident as a 'communication mishap' rather than a license dispute.

Summary

This tweet is a follow-up in-depth analysis. It quotes Cursor co-founder Aman Sanger's detailed technical response: the team evaluated many base models, and Kimi K2.5 performed strongest on perplexity metrics. Cursor then conducted continued pre-training and Reinforcement Learning at 4x compute scale on this basis. Aman Sanger acknowledged it was 'a miss' not to mention the Kimi base in the release blog and promised to fix it for next time. The author labels the incident as a 'communication mishap' rather than a license dispute, noting that Moonshot's official post specifically confirmed the legitimacy of the authorization chain (through Fireworks' commercial partnership), and that the previous employee Yulun Du's query was most likely due to internal communication lag.

AI Score

82

Influence Score 3

Published At Yesterday

Language

Chinese

Tags

Cursor

Kimi K2.5

Aman Sanger

Reinforcement Learning

Continued Pre-training HomeArticlesPodcastsVideosTweets

Cursor Co-Founder Explains Technical Selection Path; Inci...

查看原文 → 發佈: 2026-03-21 04:58:49 收錄: 2026-03-21 08:00:54

🤖 問 AI

針對這篇文章提問,AI 會根據文章內容回答。按 Ctrl+Enter 送出。