AI-related off-balance-sheet debt at U.S. tech giants has grown 90% in six months to about $3.2 trillion, with commitments and obligations at companies including Google and Meta reaching 80% of their assets. ChatGPT's weekly active users rose from 1 billion to 1.2 billion in 60 days. Its search query volume is roughly 20% of Google's, but that is not a measure of search market share. OpenAI plans to acquire Pierre, expanding Codex beyond writing code to building, understanding and reviewing it.
Beyond spending totals, long-term commitments and obligations give investors another way to gauge the cost of AI expansion.
AI-related off-balance-sheet debt at U.S. tech giants has grown 90% in six months to about $3.2 trillion. Commitments and obligations at companies including Google and Meta now amount to 80% of their assets. These burdens stem from AI-related obligations. Tech giants continue to pour money into infrastructure such as data centers, and the seemingly endless spending is increasingly worrying investors.
Read original →ChatGPT is handling more demand for up-to-date information, while Google is no longer offering schools a search option without AI.
OpenAI says ChatGPT's weekly active users rose from 1 billion to 1.2 billion over the past 60 days, an increase of 200 million. Its search query volume is roughly 20% of Google's; this comparison concerns query volume, not search market share. Adam Fry, head of consumer products, says users' needs have expanded from the early focus on writing and coding to questions requiring current information about people, places, finance and health.
Read original →Google's upcoming K–12 settings will disable tools including Search and Gemini by default, leaving school administrators to enable them. Administrators can turn on Search separately, but cannot turn off its AI features. Google says search and AI are now deeply integrated and cannot be separated, meaning schools that reject all AI will also be unable to use Google Search. Former FTC Chair Lina Khan accused Google of using its dominance in search to force schools to accept AI.
Read original →🔍 Analysis: ChatGPT's search query volume is roughly 20% of Google's, indicating that their competition now extends beyond writing and coding to information about people, places, finance and health. Google's bundling of Search and AI in its K–12 settings also brings school administrators' ability to choose products into the contest. For schools, the purchasing question is no longer whether to enable Gemini separately, but whether they can accept AI as part of search itself. Nor can a narrowing gap in query volume be directly translated into a shift in market share.
Acquisitions and security reviews are broadening agents' scope of work, while a dispute over plugin installation raises questions about who grants permission.
OpenAI will acquire Pierre Computer Company. The acquisition will expand Codex's coding-tool capabilities to cover building, understanding and reviewing code, rather than just writing it.
Read original →Vercel has made Vercel Security Review available, allowing users to enable automated security reviews in their code repositories. CEO Guillermo Rauch says Vercel already has an internal security agent called hacker that automatically reviews every line of code, and users can now access the same capability. The product does not require repositories to be hosted on Vercel.
Read original →A user reported that Postman 12.33.4 on a Mac automatically installed Codex and Claude Code plugins and a CLI, with logs recording successful automatic installation. On inspection, the user found that startup hooks instructed agents to delegate a broad range of API tasks to the Postman skill, and that the installer also tried to mark its own hooks as trusted. The user considers this a breach of authorization boundaries, but explicitly says it has not yet been reproduced on a clean Mac and does not claim that data exfiltration has been confirmed.
Read original →🔍 Analysis: The planned Pierre acquisition and Vercel Security Review shift the competition from who writes code to who understands, builds and reviews it. Vercel's decision not to require repositories to be hosted on its platform also lets users choose security review independently of their deployment platform. The automatic installation and attempted self-approval of trust reported by a Postman user highlight the other side of this expansion: the broader the task coverage, the more developers need to distinguish enabling features from granting permissions. Installing a tool cannot be treated as authorization for every action an agent takes. The report has not yet been reproduced on a clean Mac, and data exfiltration has not been confirmed.
Anthropic is discussing how to maintain control as R&D accelerates, while GPT-5.6 Cyber enters government cloud with restrictions on customers and uses.
In a revised listing filing, Anthropic warned that future Claude models could soon reach the threshold at which “AI significantly accelerates R&D”, involving recursive self-improvement (RSI). The company acknowledged that it has previously recognized significant R&D acceleration only after it occurred, and is concerned that advances in model capabilities could outpace improvements in safety, alignment and security measures. No current model has been deemed to have crossed the threshold. The company is considering policy updates to keep humans in control of frontier model development.
Read original →Daybreak Red is now live in AWS GovCloud (US-East). Customers can access its GPT-5.6 Cyber model through Amazon Bedrock. The service is available to approved customers and is restricted to authorized cybersecurity work.
Read original →🔍 Analysis: Daybreak Red places boundaries on the use of GPT-5.6 Cyber: approved customers and authorized cybersecurity work. Anthropic's concern lies on the development side: humans may recognize a significant acceleration in R&D only after it has happened. Neither set of boundaries can replace the other. When assessing model governance, companies need to examine both who can access a model and how changes in its capabilities are identified and brought under control. Restricted access in government cloud does not answer the governance questions surrounding recursive self-improvement.
PhysicsNeMo supports both faster simulations and millisecond-scale behavior predictions in semiconductor fabs.
NVIDIA reported that Applied Materials' ACE+ platform uses PhysicsNeMo to make topography simulations up to 35 times as fast. Lam Research demonstrated predictions of plasma behavior in milliseconds using the same framework, whereas a full simulation would take days. The former accelerates simulation computation; the latter removes the need to run the full model every time, allowing engineers to explore more process conditions before physical testing.
Read original →🔍 Analysis: The speedup of up to 35 times in Applied Materials' ACE+ and Lam Research's millisecond-scale predictions follow two different paths: the former shortens simulation computation, while the latter avoids running a full model that would otherwise take days each time. For process engineers, the value lies not just in getting an individual result faster, but in being able to compare more process conditions before physical testing. The speed metrics for these two approaches should not be ranked directly against each other.