11-19 更新 实时
首页/ 科技公司/ 正文

决策模型使用激增,定制化成下一前沿?

科技公司 来源:TechCrunch AI 2026-11-18 02:10 0 阅读 机器翻译
决策模型使用激增,定制化成下一前沿?

10 月,Vercel 的 AI 模型路由器上超过四分之一的请求流向了决策模型,而非 LLM——对于一种仅仅一个月前才被发明的计算机智能而言,这是一份极为庞大的工作量份额。

决策模型也是 LLM,但它们不生成输出 Token;相反,它们经过调校,利用自身的通用智能在预设选项之间快速、低成本地做出选择,同时记录自身对该选择的置信程度。它们在生产环境中很有用——在生产环境中,快速决策可能比更慢、更昂贵的前沿模型所提供的推理更重要。

TypeSafe 在发布 Jev 系列模型后成为一家估值 75 亿美元的公司,该系列至今仍在用量上占据主导(至少在我们能看到指标的地方是如此,比如 OpenRouter 这类服务——上周它处理了超过 15 亿次对 TypeSafe 模型的查询)。

然而,做出一个决策模型并不难,Jev 之后涌现了数十个。一位 Amazon 工程师用一个周末就做出了一个,很快每家超大规模云厂商都有了自己的选择。周二,Microsoft CEO Satya Nadella 表示,他的公司正在推出其决策模型 Microsoft-Decision-1.1,立即在 Azure 上面向广泛使用开放。

生态系统的两端都在行动:总部位于英国的 AI 初创公司 Sqwish 今天宣布完成 300 万美元的 pre-seed 轮融资,其旗下有一款名为 kettei fat 的决策模型。联合创始人 Ushnish Sengupta 和 Federica Freddi 在剑桥读研究生时相识,Jev 问世时他们正在开发小型 LLM,他们意识到自己的架构可以被改造成一个强大的决策模型。

在这个类别中,数据似乎是模型成功的关键——TypeSafe CEO Diogo Almedia 对 TechCrunch 表示,构建 Jev 最难的部分,是利用经过统计调整的合成数据来精细校准其判断力。Sqiwsh 也一直在使用合成数据,但其独特之处在于内置的后训练工具,可用客户自己的数据提升模型的判断力。

“我认为,让我们不沦为同质化商品的原因在于,我们为客户提供同样的能力……他们可以持续上传自己的生产数据,就像我们训练自己的模型一样,他们的模型也能在各自关心的任务上不断变得更好,”CEO Ushnish Sengupta 对 TechCrunch 表示。

想要对决策模型进行排名并不容易;Sqwish 的“kettei fat”模型在一些基准测试中表现不错,但这项技术的大多数构建者都表示,真正的检验还得看实际效果:对开发者来说,通常只要把一个模型换成另一个,就能看出它在特定场景中的表现如何。

“基准测试更像是一份邀请,让客户来试用模型,而持续调优才是他们应该继续选择我们的原因,”Sengupta 说。

Microsoft 的 Decision 1.1 模型同样是该公司在成本、延迟、智能水平和校准之间仔细权衡的结果。主导该项目的工程副总裁 Achint Srivastava 也认同,过度关注基准测试是一种错误。

Srivastava 在电子邮件中写道:“我们看到的是,‘刷榜型’模型在实际企业应用中很快就会败下阵来。”“我们的测试凸显出,要在单一模型中同时实现准确性、可靠性、速度和成本效率有多么困难。Microsoft-Decision-1.1 和 Jev 是少数在全部四个维度上都表现强劲的模型,即便是在代表真实用例的‘留出’或未见过的基准测试上接受评估时也是如此。”

对 Sqwish 这样的初创公司而言,与超大规模云厂商和前沿实验室的竞争是一个机会。Freddi 预测,未来的系统将让所谓的“系统一”智能(如决策模型)与“系统二”智能(如前沿 LLM)协同工作。

她告诉 TechCrunch:“即便到了那样的未来,‘系统一’的机会规模可能也就 100 亿、500 亿美元,而 OpenAI 做这个 IPO 时,心里想的是你知道的 1 万亿、2 万亿美元市值。”“这是我们全部的关注所在;而对他们来说,这只是一个副业项目。”

当你通过我们文章中的链接购买时,我们可能会赚取少量佣金。这不会影响我们的编辑独立性。

Tim Fernholz 是一名记者,撰写有关科技、金融和公共政策的报道。他长期密切跟踪报道私营航天产业的崛起,著有《Rocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race》。此前,他在全球商业新闻网站 Quartz 担任高级记者十余年,职业生涯始于华盛顿特区的政治记者。你可以通过发送电子邮件至 [email protected],或通过 Signal 向 tim_fernholz.21 发送加密消息来联系 Tim,或核实来自 Tim 的联系是否真实。

正在加载下一篇文章

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(As decision model use skyrockets, is customization is the next frontier?)

In October, more than one in four requests on Vercel’s AI model router went to decision models, not LLMs — an enormous share of work for a kind of computerized intelligence that was invented only a month earlier.

Decision models are LLMs that don’t produce output tokens; instead, they’re tuned to use their general intelligence to quickly and cheaply decide between preset options, while registering its level of confidence in the choice. They’re useful in production environments, where fast decisions can be more important than the reasoning provided by a slower and more expensive frontier model.

TypeSafe became a $7.5 billion company after releasing its Jev family of models, which continues to dominate usage (at least where we can see metrics from services like OpenRouter, which handled more than 1.5 billion queries to TypeSafe’s models last week).

However, it’s easy to make a decision model, and dozens were released in the wake of Jev. One Amazon engineer made one in a weekend, and soon every hyperscaler had its own option. On Tuesday, Microsoft CEO Satya Nadella said that his company is rolling out its decision model, Microsoft-Decision-1.1, for wide use on Azure immediately.

The action is happening on both ends of the ecosystem: Sqwish, a UK-based AI startup behind a decision model called kettei fat, announced a $3 million pre-seed round today. Cofounders Ushnish Sengupta and Federica Freddi, who met as Cambridge grad students, had been developing small LLMs when Jev arrived on the scene, and they realized their architectures could be repurposed as a powerful decision model.

Data appears to be the key to a successful model in this category — TypeSafe CEO Diogo Almedia told TechCrunch that the hardest part of building Jev was finely calibrating its judgement using statistically-adjusted synthetic data. Sqiwsh, too, has been working with synthetic data, but its unique offering is built-in post-training tools that use a customer’s own data to improve the model’s judgement.

“What makes us not a commodity, I think, is the fact that we are offering our customers the same power…they can keep uploading their production data, and just the same way that we trained our models, their models can also keep getting better at the tasks that they care about,” CEO Ushnish Sengupta told TechCrunch.

Attempting to rank decision models is difficult; Sqwish’s “kettei fat” model has performed well on some benchmarks, but most builders of the technology say that the real proof is in the pudding: it’s generally simple for developers to sub in one model for another and see how it does in a specific context.

“The benchmark is more of an invitation to try the model out, and then the continual tuning is the reason why why they should stick to us,” Sengupta says.

Microsoft’s Decision 1.1 model is also the result of the company carefully balancing across cost, latency, intelligence and calibration. Achint Srivastava, the engineering VP who led the project, agreed that too much focus on benchmarks is a mistake.

“What we see is that ‘benchmaxed’ models fall flat under real enterprise usage very quickly,” Srivastava wrote in an email. “Our testing underscores just how challenging it is to deliver accuracy, reliability, speed and cost efficiency in a single model. Microsoft-Decision-1.1 and Jev are among the few that perform strongly across all four dimensions, even when evaluated on ‘held out’ or unseen benchmarks reflective of real use cases.”

For startups like Sqwish, competition with hyperscalers and frontier labs is an opportunity. Freddi predicts that future systems will see so-called System One intelligence, like decision models, working in concert with System Two intelligence, like frontier LLMs.

“Even in that future, System One is maybe a $10 billion, $50 billion [opportunity], and OpenAI is doing this IPO with you know $1 trillion, $2 trillion market value in mind,” she told TechCrunch. “This is all we care about; whereas for them, this is just a side project.”

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Tim Fernholz

Tim Fernholz is a journalist who writes about technology, finance and public policy. He has closely covered the rise of the private space industry and is the author of Rocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race. Formerly, he was a senior reporter at Quartz, the global business news site, for more than a decade, and began his career as a political reporter in Washington, D.C. You can contact or verify outreach from Tim by emailing [email protected] or via an encrypted message to tim_fernholz.21 on Signal.

Loading the next article

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 TechCrunch AI(As decision model use skyrockets, is customization is the next frontier?),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗