DeepSeek AI News xAI updates its development documentation, announcing that starting November 7th, invisible text watermarks will be added by default to API outputs for Grok 4.7, 4.6, 4.5, 4.3, and three versions of the Grok 4.20 models. This covers not only the EU but also global and regional APIs, with even structured outputs and tool call parameters being marked. It remains unclear whether the Grok web version and X's built-in chat will be synchronized.

The watermark is implemented by slightly adjusting word selection probabilities during model generation, leaving identifiable statistical patterns in the text. Copy-pasting directly won't remove the watermark, but shortening or heavily rewriting the text may reduce detection reliability.

This move responds to transparency requirements under the EU's AI Act, which became applicable on August 2nd. Although xAI has not signed the EU’s voluntary transparency guidelines, it must still comply with legal obligations. Developers can disable the watermark per model, but must ensure compliance with relevant legal exemptions and assume responsibility for marking and disclosure.

Google introduced SynthID text watermarking for Gemini as early as 2024, while Anthropic began enabling watermarks globally for new Claude models in August this year. OpenAI announced in October a phased rollout of watermarks for ChatGPT and Codex in the EU region, while allowing global API users to opt-in voluntarily—default remains off.

Beating AI News: AI inference platform Baseten is planning a new funding round of approximately $2 billion. Bloomberg disclosed that this round will be priced in two tiers: the first tier at a valuation of $26 billion, and the second at $30 billion. Existing investor Sands Capital is expected to lead the round, though the deal has not yet been finalized.

Baseten's valuation was only $5 billion in January this year, and it recently completed a $1.5 billion financing round at a $13 billion valuation in June. If the new round closes at the higher tier, its valuation will have increased sixfold within just nine months. Notably, the June round was also reported to have used dual-tier pricing, at $11 billion and $13 billion respectively.

Baseten primarily helps enterprises deploy and run AI models, handling GPU compute scheduling and inference optimization. Its clients include Cursor, Notion, and OpenEvidence. The company disclosed in June that its revenue grew 20 times over the past year, and inference usage increased 40 times.

As enterprises increasingly deploy open-source and custom models, inference platforms are experiencing rapid growth.