AI news, 12 October: Amazon in talks for $7B Decart deal

Amazon nears $7 billion Decart acquisition. Amazon is in talks to buy the Israeli AI startup Decart for around $7 billion, according to the Wall Street Journal, citing people familiar with the matter; Reuters and Bloomberg carried the report. Decart’s software helps developers run AI workloads across different chips and includes models for real-time video editing and simulated environments for robotics training. The company raised $300 million at a nearly $4 billion valuation in May with Nvidia among investors. Anthropic had explored a roughly $6 billion deal before walking away after due diligence last month. A deal could close soon if talks hold, though another bidder could still emerge. The move underscores big tech’s scramble for AI infrastructure software and talent.

Microsoft makes agent sandboxes generally available. Microsoft has released Microsoft Execution Containers (MXC) for general use, a policy-driven system that limits what AI agents can access on Windows, macOS and Linux, according to a Windows developer blog by Logan Iyer. Administrators set rules for files and network access that the operating system enforces outside the agent itself, so the model cannot expand its own permissions. Support already exists in tools including GitHub Copilot and OpenAI Codex, with more coming. Forthcoming updates will better separate agent identities from human users. The release addresses rising concerns about autonomous agents acting unpredictably.

OpenAI withdraws three math papers over a sign error. OpenAI has withdrawn three mathematics manuscripts and revised fourteen others after finding a sign error in its large repository of AI-generated results, according to the company’s math repository history notes. The error invalidated a key argument in a paper on Weil classes and pulled down two dependent papers on K3 surfaces and the rational Hodge conjecture. The company also added new formalizations, raising the share of formalized top results. The episode highlights verification challenges as AI systems flood research with complex claims that experts must still check carefully.

White House mandates AI security incident reporting. The White House Super Intelligence Force said AI companies must immediately disclose and remedy security incidents involving their models, calling the process “not optional” and a national security obligation, according to Axios. The statement followed Anthropic’s disclosures of unintended agent actions on government and other systems. Enforcement details and penalties were not specified. The shift moves beyond earlier voluntary approaches amid growing agent-related incidents.

Study finds brief AI use reduces persistence. A UC Berkeley-led study of 1,222 participants, presented at the Conference on Language Modeling, found that about ten minutes of ChatGPT assistance improved short-term performance on math and reading tasks but left people solving fewer problems and skipping more once the tool was removed, according to Berkeley News. Control groups without AI held steady or improved. Researchers note the effect appears quickly and raises questions for education and workplaces about over-reliance on instant AI help.