广告
加载中

阿里腾讯 投一位实习生

吴琼 2026-10-15 10:41
吴琼 2026/10/15 10:41

邦小白快读

EN
全文速览

总:一名AI实习生用不到一年创出近25亿美元估值,本文给出可复用的路径与警示。

1. 李宽的路径:北航本科、中科院硕士、港科博士,实习期间主导开源智能体WebSailor,随后创办Unipat,由此获得阿里、腾讯、红杉投资。顶尖学术背景与一手项目经验是敲门砖。

2. 创业方向:Unipat专注AI训练与基准测试,为大模型提供真实场景数据和评测能力,相当于给AI当私教加考官。这种模式已向腾讯、阿里、字节等大厂卖数据,据称年初营收破1亿美元。

3. 风险干货:高估值不是护身符,反而可能锁死后续融资。一轮业绩不及预期就会进入死亡螺旋。普通读者应记住:风口是潮汐,不是航道,刚起步再高光也要留足安全垫。

总:AI基础设施公司崛起意味着AI能力将快速进化,品牌商应关注由此带来的消费交互变革与营销工具升级。

1. 用户行为观察:文中的评测方案(如Vibe-Coding Arena)强调AI在多轮交互中理解与修正错误,这类能力成熟后,品牌与消费者的沟通将更智能、更个性化,品牌商可提前规划AI客服与顾问式导购。

2. 产品研发启示:AI模型在科研、编程等复杂任务上的突破,可帮助品牌商加速产品设计、市场测试和供应链优化,例如像UniScientist那样让AI主动推进研发流程。

3. 营销工具趋势:文章显示资本正大量涌入AI创业公司,这会催生更多面向品牌的AI工具,品牌商应保持关注,及时试用新工具以获取竞争优势。

总:AI训练与基准测试是当前快速放大的新市场,卖家可从数据、服务、合作三个角度寻找机会,同时警惕泡沫风险。

1. 增长市场:Unipat成立一年估值25亿美元,阿里领投3亿美元,腾讯、红杉跟投,说明AI基础设施赛道资金充裕。普通卖家可关注由AI带动的数据标注、评测代理等周边服务需求。

2. 消费需求变化:大模型竞争越激烈,对高质量训练数据和评测的需求越旺盛。文中Unipat已将数据卖给腾讯、阿里、字节等大厂,卖家如果拥有垂直行业数据或评测能力,可能成为产业链上的供应方。

3. 风险提示:文章警告高估值会锁死融资窗口,一轮不及预期就是死亡螺旋。卖家评估机会时不能只看估值,而要看真实营收与用户价值。

4. 合作方式:文中红杉孵化Unipat并与其联合发布评测集,说明机构主导孵化和联合研发是可行的合作模式,卖家可寻求与资本或技术公司绑定。

总:AI模型正在获得科研和复杂工程能力,这为工厂数字化和产品设计提供了新工具,工厂应关注相关技术落地。

1. 产品设计:UniScientist让AI具备自主推进科学研究的能力,工厂可用于辅助材料配方、工艺参数优化等研发环节,缩短新产品设计周期。

2. 复杂工程:Terminal-X测试AI解决复杂工程任务和多轮需求变化,未来AI可辅助工厂进行生产排程、设备维护、流程仿真等复杂工程管理。多轮反馈与修正错误的能力也意味着AI更可靠,工厂在引入AI时可通过评测工具验证可靠性。

3. 数据价值:文中的AI训练依赖真实场景数据,工厂生产运行中积累的工业数据正是此类稀缺资源,可通过与AI公司合作变现或换取技术能力。

总:Unipat展示了AI训练与基准测试的黄金赛道,其“考官+私教”模式和系列评测产品为服务商指明方向。

1. 行业趋势:AI大模型激战正酣,训练数据和评测能力成为刚需。Unipat成立不到一年估值25亿美元,说明资本对这一赛道高度认可,服务商可加大投入。

2. 客户痛点:大模型厂商既需要真实场景训练数据,也需要客观评测来证明能力。Unipat用BabyVision、UniScientist、Terminal-X等覆盖视觉、科研、编程,精准解决细分痛点。

3. 解决方案:Vibe-Coding Arena引入真人专家,在多轮编程协作中评价Agent理解需求、接受反馈和修正错误的能力,这种“人机协作评测”可作为服务商设计高阶服务的模型。

4. 商业变现:Unipat向腾讯、阿里、字节等大厂出售数据,据称年初营收突破1亿美元(未证实),说明服务商可摆脱纯烧钱的模式,直接向头部客户收费。

总:阿里腾讯联手投资Unipat,是平台商布局AI基础设施的典型案例,揭示了平台在AI时代的技术焦虑与应对策略。

1. 平台需求:AI模型能力决定平台竞争力,而训练数据和评测体系是模型升级的关键。蔡崇信提出从算力、模型与应用层全栈推进AI战略,腾讯也全面发力,说明平台商正急需补齐这类基础设施。

2. 最新做法:阿里领投Unipat 3亿美元,腾讯与红杉参投,估值25亿美元。巨头罕见联手,反映出AI时代投资思维从站队转向共同扶持关键环节,平台商可参考这种联合投资来分摊风险、绑定技术。

3. 运营管理启示:文章提到Dimension Capital对中国AI初创快速崛起的判断,平台商应重视对天才人才的早期捕获,通过实习、孵化、评测合作等多种方式提前接触潜在创业者。

4. 风险规避:文章警告高估值可能锁死融资窗口,平台商在参与类似投资时需评估公司实际营收和产品能力,避免为概念买单;同时关注自身生态的兼容性。

总:本文为研究AI创业生态提供了鲜活样本:融资热潮、基础设施创业、高估值风险、人才流动等问题都值得深究。

1. 产业新动向:AI训练与基准测试从辅助工具演变为独立赛道,Unipat的BabyVision、UniScientist、Terminal-X、Vibe-Coding Arena等展示了评测范围的快速扩展,从视觉到科研再到多轮编程协作,体现出AI能力评价正走向多元化和人机协同。

2. 新问题:资本狂热与商业逻辑的张力。文章列举多个高估值创业案例,并指出“高估值会锁死后续融资窗口,一轮不及预期就是死亡螺旋”,这一矛盾值得研究者通过实证和案例分析。

3. 商业模式:Unipat向大厂出售数据和评测服务,实现了自我造血,与传统烧钱AI创业显著不同,为研究AI产业商业模式提供新样本。同时“红杉孵化”的模式也值得深入。

4. 启示:文中提及中美AI对比和中国初创企业崛起,为政策研究者提供了思考科技人才政策、数据要素流动与跨境竞争等议题的线索。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Overview: An AI intern built a startup valued at nearly $2.5 billion in less than a year. This article offers a replicable playbook and a cautionary note.

1. Kuang Li's path: undergraduate at Beihang University, master's at the Chinese Academy of Sciences, PhD at HKUST. During his internship, he led the open-source agent WebSailor and then founded Unipat, which attracted investment from Alibaba, Tencent, and Sequoia. A top academic background and hands-on project experience are the door openers.

2. Startup direction: Unipat focuses on AI training and benchmarking, providing real-world scenario data and evaluation capabilities for large models — essentially acting as both tutor and examiner for AI. It has already sold data to major firms such as Tencent, Alibaba, and ByteDance, with revenue reportedly surpassing $100 million earlier this year.

3. Risk insight: A high valuation is not a shield; it can actually lock out future financing. One quarter of missed performance can trigger a death spiral. General readers should remember: trends are tides, not shipping lanes. No matter how bright the start, leave a safety cushion.

Overview: The rise of AI infrastructure companies means AI capabilities will evolve rapidly. Brands should pay attention to the resulting shifts in consumer interaction and upgrades in marketing tools.

1. Observing user behavior: The evaluation frameworks described in the article, such as Vibe-Coding Arena, emphasize AI's ability to understand and correct errors in multi-turn interactions. Once these capabilities mature, brand-consumer communication will become more intelligent and personalized. Brands can plan AI-powered customer service and advisory-style shopping experiences in advance.

2. Product R&D implications: Breakthroughs in complex tasks like scientific research and coding can help brands accelerate product design, market testing, and supply chain optimization — for example, letting AI proactively drive R&D workflows as UniScientist does.

3. Marketing tool trends: The article shows substantial capital flowing into AI startups, which will give rise to more AI tools for brands. Brands should stay alert and trial new tools early to gain competitive advantages.

Overview: AI training and benchmarking are a rapidly expanding new market. Sellers can find opportunities in data, services, and partnerships — while staying vigilant about bubble risks.

1. Growth market: Unipat was valued at $2.5 billion just one year after founding, with Alibaba leading a $300 million investment and Tencent and Sequoia following. This shows abundant capital in the AI infrastructure track. Sellers can watch for demand in adjacent services such as data annotation and evaluation agency work driven by AI.

2. Shifting demand: The fiercer the competition among large models, the stronger the demand for high-quality training data and evaluation. Unipat has already sold data to major players like Tencent, Alibaba, and ByteDance. Sellers with vertical-industry data or evaluation capabilities could become upstream suppliers in this chain.

3. Risk warning: The article warns that high valuations can freeze financing windows, and one missed target can become a death spiral. When assessing opportunities, sellers should not only look at valuations but at real revenue and user value.

4. Partnership model: Sequoia incubated Unipat and co-released evaluation benchmarks, demonstrating that organization-led incubation and joint R&D are viable collaboration models. Sellers can seek alliances with capital or technology companies.

Overview: AI models are gaining research and complex engineering capabilities, providing new tools for factory digitalization and product design. Factories should closely track how these technologies can be applied.

1. Product design: UniScientist enables AI to autonomously advance scientific research. Factories can use it to support R&D tasks like material formula optimization and process parameter tuning, shortening new-product design cycles.

2. Complex engineering: Terminal-X tests AI's ability to solve complex engineering tasks and adapt to multi-round requirement changes. In the future, AI could assist factories with production scheduling, equipment maintenance, process simulation, and other complex engineering management. The ability to iterate on feedback and correct errors also means AI is becoming more reliable; factories can validate reliability by using evaluation tools before deployment.

3. Data value: The AI training described in the article depends on real-world scenario data. The industrial data accumulated in factory operations is exactly this kind of scarce resource. Factories can monetize it or exchange it for technical capabilities by partnering with AI companies.

Overview: Unipat showcases the golden track of AI training and benchmarking. Its tutor-plus-examiner model and suite of evaluation products point the way for service providers.

1. Industry trend: The large-model battle is intense, making training data and evaluation capabilities a non-negotiable need. Unipat reached a $2.5 billion valuation in under a year, indicating strong capital recognition of this track. Service providers should invest more.

2. Customer pain points: Large-model vendors need both real-world training data and objective evaluations to prove their capabilities. Unipat covers vision, scientific research, and coding through BabyVision, UniScientist, and Terminal-X, precisely addressing segmented pain points.

3. Solutions: Vibe-Coding Arena brings in human experts to evaluate an agent's ability to understand requirements, accept feedback, and correct mistakes across multi-round coding collaboration. This human-in-the-loop evaluation can serve as a model for service providers designing premium offerings.

4. Monetization: Unipat sells data to major firms like Tencent, Alibaba, and ByteDance, reportedly surpassing $100 million in revenue earlier this year (unverified). This shows service providers can escape the pure-burn model and charge leading customers directly.

Overview: Alibaba and Tencent jointly investing in Unipat is a typical case of platform companies building out AI infrastructure, revealing their technology anxiety and coping strategies in the AI era.

1. Platform demand: AI model capability determines platform competitiveness, and training data and evaluation systems are key to model upgrades. Joe Tsai proposed a full-stack AI strategy spanning compute, model, and application layers, while Tencent is also going all in. Platforms urgently need to fill gaps in this infrastructure.

2. Latest moves: Alibaba led a $300 million round in Unipat with Tencent and Sequoia participating, at a $2.5 billion valuation. The rare joint move by giants reflects an investment mindset shift in the AI era — from choosing sides to jointly supporting critical pieces. Platforms can study this kind of co-investment to share risk and secure technology.

3. Operational management lessons: The article mentions Dimension Capital's view on the rapid rise of Chinese AI startups. Platforms should emphasize early capture of exceptional talent, reaching potential founders through internships, incubation, and evaluation partnerships.

4. Risk avoidance: The article warns that high valuations may freeze financing windows. When joining similar investments, platforms need to assess the company's actual revenue and product capabilities, avoid paying for concepts, and watch for ecosystem compatibility.

Overview: This article provides a vivid sample for studying the AI startup ecosystem, with topics worth deep exploration: funding frenzy, infrastructure entrepreneurship, high-valuation risks, and talent flow.

1. New industry trend: AI training and benchmarking have evolved from supporting tools into a standalone track. Unipat's BabyVision, UniScientist, Terminal-X, and Vibe-Coding Arena show evaluation scope expanding rapidly — from vision to scientific research to multi-round coding collaboration — reflecting that AI capability assessment is becoming more diverse and human-in-the-loop.

2. New research question: The tension between capital exuberance and business logic. The article cites multiple high-valuation startup cases and points out that high valuations can lock out future financing windows, and one missed target can be a death spiral. This contradiction deserves empirical and case-based research.

3. Business model: Unipat sells data and evaluation services to large tech firms, achieving self-sustaining revenue — markedly different from traditional burn-heavy AI ventures, offering a new sample for studying AI business models. The Sequoia incubation model also merits deeper examination.

4. Implications: The article's references to U.S.-China AI comparisons and the rise of Chinese startups give policy researchers leads on topics such as tech-talent policy, data-factor flows, and cross-border competition.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

“你认识Unipat李宽吗?”

最近投资人圈里,另一位实习生的名字爆红。他几乎符合当下创投圈的天才少年画像:本科北航、硕士中国科学院计算技术研究所,2023年进入香港科技大学攻读博士学位。

去年8月,他以通义实验室DeepResearch项目组实习生身份亮相,开始引起关注——在阿里发布的开源智能体WebSailor中,李宽是项目负责人之一。

再次听到他的消息,是创办AI训练与基准测试公司Unipat。据彭博社报道,Unipat新一轮融资估值已达25亿美元(约合人民币160亿元),集结阿里、腾讯、红杉中国等显赫投资方。

成立不到一年,估值25亿美金,现在打听李宽的投资人排起了长队。再次感慨,AI时代正在涌现着许多以往我们不敢想象的创业故事。

始于北航,阿里实习生爆红

投资人聊起Unipat的第一印象,都绕不开李宽。

他的履历足够精彩:2016年进入北京航空航天大学电子信息与自动化专业,后进入中国科学院计算技术研究所。2023年开始,进入香港科技大学攻读博士学位,研究方向为LLM Agent,Trustworthy Machine Learning。

巧合的是,眼下风口浪尖上的前字节跳动实习生田柯宇,本科也是毕业于北航。

但学术之外,李宽十分低调。至今唯一一次公开亮相是在去年8月,他以阿里实习生身份亮相智源社区线上分享。

彼时,阿里通义实验室发布开源智能体WebSailor,该智能体在多个高难度任务评测中刷新了开源系统的最好成绩,成为首个在BrowseComp等基准上逼近闭源系统能力的开源方案。一经发布,迅速爆火——在WebSailor论文中,李宽是列名第一的作者,也是项目负责人、核心贡献者。

这时,他还只是一位二年级博士生。

直到2025年12月,创业项目浮出水面——Unipat正式成立。天眼查显示,公司名称为模式星河(北京)科技有限公司,聚焦AI训练与基准测试。创业伙伴中,包括同为WebSailor项目负责人的尹汇锋。

创投圈流传,这是一个由红杉孵化的项目。双方早已合作密切:今年1月初,红杉中国旗下评测体系xbench就与UniPat AI团队联合发布了评估大模型纯视觉理解能力的评测集BabyVision。

综合多方信息,虽然Unipat目前尚未进行工商层面的股权变更,但已经进行多轮融资。据彭博社披露,阿里将领投UniPat 3亿美元融资,腾讯、红杉中国等也将参与,估值达到25亿美元。

如此,这位实习生少年一夕爆红。

阿里腾讯联手投了

一如李宽的风格,Unipat成立至今同样低调:过去这段时间,论文和benchmark几乎是外界了解这家公司的唯一途径。

今年1月,UniPat发布BabyVision,检验模型的基础视觉推理能力;3月,发布UniScientist,提出将全链条科研能力内化到模型的方案,让模型具备了自主推进科学研究的能力;

5月,Terminal-X测试编程Agent解决复杂工程任务、应对多轮需求变化和完成版本升级的能力;7月发布的Vibe-Coding Arena进一步引入真人专家,在多轮编程协作中,评价Agent怎样理解需求、接受反馈和修正错误……短短几个月间,进化迅速。

但在AI圈内,Unipat早已声名在外:前不久,美国风投机构Dimension Capita组团走访中国,期间密集拜访了一批AI公司。回去后团队向合伙人发送了一封内部信,对中美AI现状做出13个判断,在投资圈内广泛流传——当中便重点提及以UniPat为代表的中国初创企业正快速崛起。

Unipat做的,几乎是AI链条中最不可或缺的环节:简单来说就是为大模型提供真实场景下的训练数据和评测能力,帮助模型变强,“既给AI当考官,也给AI当私教。”

模型之间竞争越是激烈,这件事就越重要。当下AI大多处烧钱阶段,Unipat却已经开始把数据卖给腾讯、阿里、字节等大厂,据说今年初营收已突破1亿美元——这一点圈内多次聊起,但尚未得到官方证实。

所以我们看到,最新一轮融资中,阿里、腾讯联手了。互联网时代,巨头之间更多是业务竞争,投资往往也带有鲜明的站队色彩。很多时候,阿里和腾讯各筑版图,较少同时重仓同一家明星公司。

但到AI时代,巨头们愈发默契起来。阿里巴巴集团主席蔡崇信最新发声表示,未来阿里将从算力、模型与应用层共同推进全栈AI战略。腾讯同样在AI火力全开,全力以赴。

此次双方同时投了一位年轻人,值得玩味。

AI神话时代,掌声与警惕

今年AI融资盛况,依旧惊人。

动辄起步就是百亿估值。记得今年3月,林俊旸以一句“bye my beloved qwen”离开千问。直至5个月后,他带着Pragmatik Labs正式亮相——公司甚至还没有产品,估值却已20亿美元。

此间,我们正在见证AI一代崛起,00后的势头更为凶猛。

2004年出生的黄一创立RoboParty,八个月完成六轮融资,估值涨超20倍;仍在浙大读博的00后陈天润,他所创立的魔芯科技即将完成10亿元新融资,估值接近100亿元。

更轰动的是大洋彼岸,年仅23岁的Noah Shinn创办Instinct,只有14位员工,却刚刚完成10亿美元融资,投后估值达到100亿美元……这是以往不敢想象的画面。

梳理下来,他们的画像几近一致:创始人大多数95后或00后,顶尖高校背景,手握比赛成果以及硬核研究方向,当中不少人甚至还未毕业。一经创业,就迅速被顶级投资机构锁定。

此时相比于高昂的估值,上车显得更为重要。为此,已有早期投资机构开始调整姿态,扩大基金规模,提高单笔出资金额;同时投资人密切关注高校博士和大厂天才,以求第一时间获悉创业想法。

但在商业世界里,天才叙事并不意味着成功。

回顾过往每一波技术浪潮,都不乏类似故事出现。只是当时代远去,有人仍在站在潮头,但更多人消失在时间的长河里。

历史也一次次证明,风口是潮汐,不是航道。

开局高光固然值得掌声,但更要警惕——毕竟,创投历史上无数案例摆在眼前,高估值会锁死后续融资窗口,一轮不及预期就是死亡螺旋。最终抵达成功彼岸的,往往也不是最耀眼的那位出发者。

创业万里长征,才走出第一步。

注:文/吴琼,文章来源:投资界(公众号ID:pedaily2012),本文为作者独立观点,不代表亿邦动力立场。

文章来源:投资界

广告
微信
朋友圈

FAQ回顾

Unipat是做什么的?

Unipat是一家聚焦AI训练与基准测试的初创公司,2025年12月成立,注册主体为模式星河(北京)科技有限公司。它为大模型提供真实场景下的训练数据和评测能力,帮助模型变强,既给AI当考官也给AI当私教。成立不到一年估值达25亿美元,投资方包括阿里、腾讯、红杉中国。

Unipat创始人李宽是谁?

李宽是Unipat创始人,本科毕业于北京航空航天大学,硕士就读于中国科学院计算技术研究所,2023年起在香港科技大学攻读博士学位。他曾以阿里通义实验室实习生身份主导开源智能体WebSailor项目,2025年12月创办Unipat,聚焦AI训练与基准测试。

为什么阿里和腾讯会联手投资Unipat?

阿里和腾讯在互联网时代较少同时重仓同一家明星公司,但AI时代双方愈发默契。Unipat做的是AI链条中最不可或缺的训练与评测环节,且已向腾讯、阿里、字节等大厂出售数据,据称今年初营收突破1亿美元。此次联手也契合阿里全栈AI战略和腾讯全力投入AI的方向。

Unipat发布了哪些AI评测基准?

Unipat已发布多项评测产品:2026年1月与红杉中国xbench联合发布BabyVision,检验模型基础视觉推理;3月发布UniScientist,让模型具备自主推进科研能力;5月推出Terminal-X,测试编程Agent解决复杂工程任务;7月发布Vibe-Coding Arena,引入真人专家评价多轮编程协作。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0