Microsoft CEO criticizes Anthropic for being too strict in content control of Fable model
CEO Satya Nadella said on Wednesday that Anthropic's practice of placing various restrictions on queries submitted by users to its high-end artificial intelligence model Fable makes no sense. According to a transcript of Nadella's speech, he told engineers developing Microsoft's Copilot artificial intelligence tool: "If you have used Fable, you will find that it inexplicably refuses various instructions. When have you ever seen a creative tool subject to such strict content control? This makes no sense at all." According to a technical support page, when end users ask Fable about large-scale model building related issues or other topics, Anthropic sometimes calls older versions of the model to generate responses. Many users have complained on social media that the model frequently refuses normal questions. Anthropic said when it released Fable 5 in early June that the company was working to reduce the number of false interceptions of normal requests. Just three days after the model went online, Anthropic suspended the use of Fable in order to comply with the U.S. government's export control directives; on July 1, the company resumed the model service and stated: "Compared with the old version, the new version of the security protection mechanism will mark slightly more normal requests that are risk-free." Nadella's comments come as corporate executives increasingly favor more cost-effective artificial intelligence models - models that are not produced in the best-funded labs but are capable of in-house software development and other business tasks. Nadella's remarks were tantamount to criticism of his important partners and customers. Anthropic's Claude Code software development tool is growing in popularity among programmers and non-technical talent. In November last year, Microsoft announced that it would invest $5 billion in Anthropic, and the startup also committed to invest $3 billion in purchasing Microsoft Azure cloud services. This year, Microsoft released the enterprise office assistant Copilot Cowork, which integrates Anthropic's large model capabilities. Investors have been worried that large models that can quickly generate software code will impact Microsoft's business; and Microsoft allocates tens of billions of dollars every quarter to expand data centers. Since the beginning of this year, Microsoft’s stock price has fallen by 17%. The composite index rose 11%. Recently, Nadella proposed that companies should be able to cost-effectively develop and customize large models themselves and access internal data without having to outsource internal data to other institutions, such as companies that specialize in developing large models. In a blog post on Sunday, he cited Palantir CEO Alex Karp, who said technology companies "want to own their own production tools." Microsoft has launched the Azure AI Foundry service. Developers can choose from more than 11,000 large models on this platform, including multiple models from Anthropic and OpenAI. Nadella told R&D engineers: "It is impossible that there are only two companies in the world that hold token computing resources, and all other companies can only rent them from them. This business model is completely inconsistent with economic logic." Token is a unit of measurement for measuring the computing power consumption of artificial intelligence models.