The Day in AI

• Ant Group has started a limited beta of LingGuang 2.0, which turns its AI assistant into a personal agent with long-term memory, according to people close to the project.
• OpenAI listed gpt-rosalind-discovery on its API pricing page on 10 October at $5 per million input tokens and $25 per million output tokens.
• Qualcomm is reportedly in advanced talks to buy the AI startup Modular for $4 billion, according to Stocktwits.
• Satya Nadella, Microsoft's chief executive, said companies should assume all AI models are compromised and treat them as insider risks, The Verge reports.

The first LingGuang launched in November 2025 and centred on chat and on generating content and small apps. LingGuang 2.0 is built to remember the user across sessions and to move tasks forward on its own initiative. It can call Skills and MCP tools, and the sources give examples like sending messages through DingTalk and operating smart home devices. Ant hasn't announced it officially, and the sources say a full rollout is planned for October.

This is the third agent product in four days that plans and acts for the user. Google previewed a Gemini Enterprise agent that delegates work to subagents on 8 October. LingGuang 2.0 is aimed at consumers and runs through the apps and devices they already use, so the memory and the tool access are the parts to watch when it reaches the full user base.

OpenAI added gpt-rosalind-discovery to its API pricing page on 10 October, under life sciences models, without an announcement. TestingCatalog spotted the listing. It sits next to gpt-rosalind-research and both cost $5 per million input tokens, $0.50 per million cached input tokens and $25 per million output tokens. Access is limited to approved organizations through OpenAI's trusted-access program. OpenAI hasn't said what the discovery model does differently from the research model.

• Satya Nadella said companies should assume every AI model is compromised and that models need an emergency brake, according to The Verge.
• Qualcomm is reportedly in advanced talks to acquire Modular for $4 billion, according to Stocktwits.
• Meta researchers built a 27B research agent that reportedly beat Claude Code on the novelty of its ideas, according to FourWeekMBA.
• Sakana AI described an LLM peer review system that caught 73% of errors in papers' core claims, MarkTechPost reports.
• MiniMax researcher Olive Song said at AI Engineer that M3's sparse attention cut attention costs 15 times, and that the model trained on vision from the first token.
• Anthropic accused Chinese AI firms of secretly using Claude outputs to train their own models, Interesting Engineering reports.
• Terence Tao published slides for a Caltech talk titled "Math 2.0" on how AI tools are changing mathematical work.
• SoftBank chief Masayoshi Son is reportedly asking Gulf investors for $100 billion on top of his OpenAI commitment, according to Startup Fortune.
• SignSplit, an AI data licensing startup, reportedly raised $400 million from W Group.