
就在Anthropic以有史以来最大规模IPO上市、投资者对这家AI实验室的热情高涨的同一天,远在赫尔辛基Slush初创企业大会上,当天最热门的话题却是“开源一切”。
On the same day that Anthropic went public in the largest IPO on record, with investor enthusiasm for the AI lab through the roof, far away at the Slush startup conference in Helsinki, the topic du jour was open-source everything.
这就是Fireworks AI等公司正在追逐的巨大机遇,因为企业正在寻找方法,避免为AI需求支付前沿模型的高昂价格。
That’s the enormous opportunity being chased by companies like Fireworks AI, as enterprises seek ways to avoid paying frontier prices for their AI needs.
Fireworks为其他企业运行AI模型。它专注于开放权重模型——任何人都可以下载,并帮助客户利用自身数据对这些模型进行微调,使其运行速度更快、成本更低,且比Anthropic、OpenAI等巨头的模型更难被竞争对手复制。客户包括Uber、Shopify和Notion。联合创始人赵辰宇(Chenyu Zhao)和我在当天接近尾声时同台对话。
Fireworks runs AI models for other businesses. It specializes in open-weight models, which anyone can download, and helps customers tune them on their own data so they run faster and cheaper than the models of Anthropic, OpenAI and other heavyweights and are harder for rivals to copy. Customers have included Uber, Shopify and Notion. Co-founder Chenyu Zhao and I spoke on stage toward the end of the day.
赵辰宇和许多其他人认为,企业正转向像他这样的平台,因为前沿模型价格昂贵,而且在大多数情况下,他说,除非你需要最新、最强的通用智能,否则这些模型并非必要。此外,还有一种日益增长的担忧:前沿实验室会吸收过多客户的专有知识,哪怕只是通过获取答案时使用的提示词。微软CEO萨提亚·纳德拉(Satya Nadella)在7月就提出过这一担忧。
Zhao and many others believe companies are gravitating toward platforms like his because frontier models are pricey and, in most cases, he says, unnecessary unless you need the very latest and greatest generalized intelligence. There are also growing concerns that frontier labs absorb too much of their customers’ know-how, even if only through the prompts used to get answers, a worry Microsoft CEO Satya Nadella raised in July.
Fireworks今秋也遭遇了首次重大挫折,披露了两起安全事件。在其中一起事件中,黑客获取了存储为机密的少量账户的API密钥和令牌;在另一起事件中,攻击者入侵其云服务商,进而渗透到开发系统,窃取了客户数据,包括账单记录和部分数据集。
Fireworks has also gotten its first black eye, having disclosed two security incidents this fall. In one, hackers reached API keys and tokens stored as secrets for a limited number of accounts; in the other, attackers breached its cloud provider, made their way into a development system, and nabbed customer data, including billing records and some datasets.
Notion表示,不足1%的工作区AI聊天记录被泄露,但已与Fireworks断开连接。Uber称,黑客通过一家供应商窃取了约15万名司机的姓名、地址、出生日期和驾照号码,TechCrunch确认该供应商为Fireworks。Fireworks将不得不赢回这些客户的信心。
Notion says AI chat transcripts from fewer than 1% of its workspaces were exposed, but it has disconnected from Fireworks. Uber says hackers stole the names, addresses, birth dates and driver’s license numbers of about 150,000 drivers through a vendor that TechCrunch identified as Fireworks. Fireworks will have to win back the confidence of customers like them.
投资方在夏季给 Fireworks 估值 175 亿美元,据报道该公司正处于“早期谈判”阶段,准备再融一轮,赵琳拒绝置评。
Fireworks, assigned a $17.5 billion valuation by its investors in summer, is reportedly in “early talks” for yet another round, though Zhao declined to comment.
看来这赛道谁都不缺钱。竞争对手 Together AI 和 Baseten 近几个月相继宣布完成巨额融资。周三,Palo Alto Networks 宣布将收购 Applied Compute,这家成立仅一年的初创公司由前 OpenAI 研究人员创立,与 Fireworks 一样,帮助企业利用自有数据训练和运行定制 AI 模型。交易金额 45 亿美元,全为股票支付。以下是赵琳的部分观点。对话经编辑删减,以求简明。
There’s money chasing everyone in the category, seems like. Rival companies Together AI and Baseten have announced massive rounds in recent months. On Wednesday, Palo Alto Networks said it will buy Applied Compute, which, like Fireworks, helps companies train and run custom AI models on their own data. The deal for the startup, founded just a year ago by former OpenAI researchers, is $4.5 billion, all of it in stock. Here’s some of what Zhao had to say. This conversation has been edited for length and clarity.
您在 Google 工作了十年,曾管理 Google Cloud Vertex AI 的一个 50 人工程师团队。为什么当时选择离开?
You spent a decade at Google, including managing a 50-person engineering team on Google Cloud Vertex AI. Why leave when you did?
遇到优秀的联合创始人,五分钟内就能明显感觉到他们非常聪明。跟 Lin(Qiao,乔琳)和 Dima(Dzhulgakov)聊完我就决定了。找到真正好相处、合作愉快的人也很重要。
When you meet great co-founders, it’s obvious within five minutes that they’re really smart. I was sold after talking to Lin [Qiao] and Dima [Dzhulgakov]. It’s also important to find people who are genuinely nice, people you’ll enjoy working with.
如何向从未听说过 Fireworks 的 CEO 介绍公司?
How do you describe Fireworks to a CEO who has never heard of it?
我把 Fireworks 定义为专用智能平台。OpenAI 和 Anthropic 在构建通用智能。我们使用开放模型——通常更小、更便宜、更快——让企业加入自有数据进行专业化定制。这样你就拥有自己的 AI,而不必依赖巨型 AI 集团。如果只是用 Opus 或 Sonnet,那不具备差异化优势。开放模型目前落后多少?
I describe Fireworks as a specialized intelligence platform. OpenAI and Anthropic are building general intelligence. We use open models, which are often smaller, cheaper and faster, and let companies add their own data to specialize them. That way you own your AI rather than being beholden to giant AI conglomerates. If you’re just using Opus or Sonnet, that’s not a differentiator. How far behind are open models right now?
几年前我会说落后一年、一年半。后来出现了 DeepSeek 和一大批优秀的开放模型。现在最多落后六个月,可能更接近三个月。模型差距这么小,为什么企业转换意愿不强?
A few years ago I’d have said a year, a year and a half. Then came DeepSeek and a waterfall of fantastic open models. These days it’s at most six months behind, probably closer to three. If the models are so close behind, why are businesses slow to switch?
前沿模型已成为家喻户晓的品牌,在你找到产品与市场契合点之前,使用最好的模型风险较低。但一旦用户突破 1 万或 10 万,且你每月在 Token 上花费 1 万到 10 万美元,就是考虑开源模型的时候了。
Frontier models are a household brand, and before you find product-market fit, using the best model is low risk. But once you pass 10,000 or 100,000 users and you’re spending $10,000 to $100,000 a month on tokens, it’s time to think about open source.
7 月,微软 CEO 萨提亚·纳德拉警告称,使用专有 AI 模型的公司要付出双重代价,一次是金钱,另一次是专有技术。你认为该警告产生了实质性影响吗?
In July, Microsoft CEO Satya Nadella warned that companies using proprietary AI models pay twice, once in money and again in know-how. Do you think that warning had a material effect?
我们认为未来将是一个“双模型世界”,前沿模型与开源模型并行。你需要最强的智能来进行高层规划,但基础算术不需要前沿模型。这也是为什么路由技术如此受关注,以及为什么人们现在考虑的是每任务成本,而不只是每 Token 成本。
We think we’ll live in a two-model world, with frontier and open models in parallel. You want the best intelligence for high-level planning, but you don’t need a frontier model for basic arithmetic. That’s why there’s so much interest in routing, and why people now think about cost per task, not just cost per token.
Anthropic 今日上市。你如何看待其估值?一种情形是它继续巩固自己在寡头格局中的地位;另一种情形是它开始失去吸引力。
Anthropic went public today. What do you think of its valuation? There’s one scenario in which it just continues to cement its place in this sort of oligopoly; there’s another where it starts to lose traction.
公开记录在案:我今晚大概会给经纪人打电话,买几股。这是一项长期押注。10 年后 AI 使用规模将完全不同,这需要惊人的资本,这也是这些公司上市的原因。
On the record: I’ll probably call my broker tonight and buy a few shares. This is a long-term bet. In 10 years AI usage will be at a completely different scale, and that takes an incredible amount of capital, which is why these companies are going public.
Fireworks 近期发生了几起安全漏洞。你如何向客户解释?
Fireworks has had a couple of security breaches recently. What do you tell customers?
每家公司大概都面临着前所未有的攻击频率。就我们的事件而言,很大一部分动机是“LLM 劫持”:攻击者在各供应商中搜寻 API 密钥和机密,以便在非法市场倒卖 Token。我们的承诺是对推理流量零数据留存,目前没有迹象表明该承诺已被破坏。
Every company is probably seeing an unprecedented frequency of attacks. For our incident, a lot of the motivation was LLM jacking: hunting through providers for API keys and secrets to sell tokens on illicit marketplaces. Our promise is zero data retention for inference traffic, and there’s no indication that has been compromised.
Notion 已与 Fireworks 断开连接。你们是否失去了其他客户?
Notion has disconnected from Fireworks. Did you lose other customers?
我不认为我可以公开点名受影响的具体客户。包括 Notion 和 Cursor 在内的部分客户已自行公开。只有我们的一小部分客户遭遇了数据影响。考虑到你们承诺零数据留存,这些客户数据是如何泄露的?
I don’t think I can publicly name the specific customers that were impacted. Some, such as Notion and Cursor, have gone public already. It was a small subset of our customers that had this data impact. How did that customer data get exposed, given your zero-retention promise?
Notion 使用批量预测,即上传存储在静止状态下的数据集,因此它本质上不是零保留产品。我们正将这些服务迁移到在线流程,以便仅在需要时保留数据。
Notion was using batch prediction, where you upload a dataset that’s stored at rest, so it’s not a zero-retention product by nature. We’re moving those services to an online flow so we keep data only as long as we need it.
黄仁勋上周离开了英伟达,英伟达自 2024 年起就支持 Fireworks。这会影响你们吗?
Jensen Huang left Nvidia last week, and Nvidia has backed Fireworks since 2024. Does that affect you?
这对行业是一次打击,我们对黄仁勋及其家人表示同情。但英伟达一直是开源的强力支持者,我认为这种情况不会改变。我们预计英伟达将继续投资 Nemotron,并且我们对继续与其合作感到兴奋。
It’s been a blow to the industry, and we feel for Jensen and his family. But Nvidia has always been a strong supporter of open source, and I don’t see that changing. We expect it to keep investing in Nemotron, and we’re excited to keep partnering with them.