已成立的软件公司必须围绕人工智能重建其产品，而不是把人工智能仅仅作为另一项功能添加进去，SAP 首席人工智能战略官 Sean Kask 说。他在周四于都灵的 Wave by Vento 会议上发表了讲话。
Established software companies have to rebuild their products around AI, not add it as one more feature, SAP’s chief AI strategy officer Sean Kask said. He spoke at Wave by Vento in Turin on Thursday.
“正如那句著名的调侃——每家公司都在变成软件公司，每家软件公司也必须变成人工智能公司，否则它们将会消亡，”Kask 说。
“Like the famous quip that every company is becoming a software company, every software company has to become an AI company, or they’ll perish,” Kask said.
CNBC 的 Carolin Roth 主持了题为《模型欧洲打造》的会议。会议围绕 SAP 收购德国初创公司 Prior Labs 展开，Prior Labs 专注于构建表格基础模型：为表格数据设计的人工智能模型。SAP 于七月完成了此次收购，并承诺在四年内投入超过 10 亿欧元，将 Prior Labs 打造成前沿人工智能实验室。
CNBC’s Carolin Roth moderated the session, titled “The Model Europe Built”. It centred on SAP’s purchase of Prior Labs, a German startup that builds tabular foundation models: AI models made for data held in tables. SAP completed the deal in July. It has committed more than €1bn over four years to grow Prior Labs into a frontier AI lab.
为何关注表格？大型语言模型通过预测文本中的下一个词进行学习，Kask 说。从数学上讲，它们无法处理表格中的数据，例如电子表格的行和列。表格模型能够进行数值预测和分类，这是语言模型无法实现的。
Why tables Large language models learn by predicting the next word in text, Kask said. Mathematically, they cannot handle data held in tables, such as the rows and columns of a spreadsheet. Tabular models can make numerical predictions and classifications that language models cannot.
他说，大约 80% 的业务数据是非结构化的，如文本、电子邮件和图像。但他指出，表格中那 20% 的数据却产生了公司 80% 的价值。大型人工智能实验室则更专注于扩展语言模型。
About 80% of business data is unstructured, such as text, email and images, he said. But he argued that the 20% held in tables generates 80% of a company’s value. The big AI labs focus on scaling language models instead, he said.
为何选择 Prior Labs？Kask 说，SAP 在几年前就开始内部构建表格模型，并在生产环境中使用自研的 SAP‑RPT‑1。Prior Labs 之所以合适，是因为其模型在 TabArena 基准测试中名列前茅。他还提到，Prior Labs 已经将该模型应用于癌症诊断和银行交易，SAP 也在进行相关研究。
Why Prior Labs SAP started building tabular models in-house a couple of years ago, Kask said, and uses its own, SAP-RPT-1, in production. Prior Labs fitted because its model leads the TabArena benchmark, he said. Prior Labs was also already applying it to cancer diagnosis and bank transactions, and SAP was doing research with it.
Kask 表示，SAP 承诺将 Prior Labs 保持为独立实验室，拥有自己的研究空间。其模型将保持开放权重，供研究人员和初创公司基于此进行二次开发。
SAP promised to keep Prior Labs as a separate lab with room for its own research, Kask said. Its model will stay open weight, so researchers and startups can build on it.
过去，他说，机器学习团队会为每个问题构建一个模型，可能需要数周甚至数月的时间。一家公司可能会训练和调优 100 个模型。而使用表格基础模型，只需几天即可完成同样的工作，并且在准确性上超越 XGBoost 等方法。
In the past, he said, a machine learning team built one model for each problem, which could take weeks or months. A company might train and tune 100 models. With a tabular foundation model, one model does that work in days, he said, and beats methods such as XGBoost on accuracy.
Kask表示，SAP不自己构建大型语言模型，因为成本高昂，而且这些模型正在趋于同质化。客户可以使用来自谷歌、OpenAI、Anthropic和Mistral的模型，而SAP内部则使用了超过100个模型。公司会对每个用例在所有模型中进行测试，并选择表现最佳的一个。
Picking the best model SAP does not build its own large language model, Kask said, because it is costly and the models are converging. Customers can use models from Google, OpenAI, Anthropic and Mistral, and SAP uses more than 100 models internally. It tests each use case against all of them and picks the best one.
他表示，对SAP而言，“以AI为先”意味着要重新构建其界面。用户可以用自然语言提问，系统就会在屏幕上生成他们所需的表格。
For SAP, AI-first means rebuilding its interface, he said. Users can ask in plain language, and the system generates the tables they need on screen.
Kask表示，欧洲不应贸然进入大型语言模型这一“红海”。他说道，在一个模型推出六个月后就变得过时的情况下，投入100亿美元并不会让任何一家公司脱颖而出。相反，欧洲应该专注于构建能让这些模型发挥实际价值的资产。
Europe’s place Europe should not jump into the “red ocean” of large language models, Kask said. Spending $10bn on a model that is outdated in six months does not set anyone apart, he said. Europe should build the assets that make those models useful instead.
他表示，SAP的知识图谱连接了50万个表格和700万个字段，这使得其AI代理具备足够高的准确性，能够用于实际工作。但并非所有欧洲企业都选择了这条路。周二，Mistral推出了Large 4，这是一款拥有1万亿参数的开源模型。
SAP’s knowledge graph links 500,000 tables and 7 million fields, he said, which makes its AI agents accurate enough to use at work. Not every European company is taking that route. On Tuesday, Mistral launched Large 4, a 1-trillion-parameter open-weight model.
Kask表示，SAP每售出一美元的软件，就能为其合作伙伴生态系统带来6到10美元的收入。SAP会将初创公司的技术授权到自己的产品中，通过其商店转售合作伙伴的产品，有时还会直接收购这些公司。周二，该公司同意收购比利时AI劳动力初创企业TechWolf。
Every dollar of SAP software sold generates $6 to $10 for its ecosystem of partners, Kask said. SAP licenses startups’ technology into its products, resells partners through its store, and sometimes buys them. On Tuesday, it agreed to buy TechWolf, a Belgian AI workforce startup.