November 2025

The roadmap took the company best known for smartphone chips into rack-scale inference, the part of the market where models are run rather than trained.

The agreement paired near-term cloud access for researchers with the construction of dedicated scientific machines.

The link targeted quantum error correction, which needs fast classical computation alongside error-prone qubits; Jensen Huang called it “the Rosetta Stone connecting quantum and classical supercomputers.”

The experiments gave limited, unreliable evidence that the reports of the most advanced Claude models about themselves can sometimes reflect their internal states.

The valuation put a single supplier of AI hardware at the center of the market’s expectations for the technology.

The agreement brought one of the largest rights holders into the development of AI tools for musicians.

The workflow connected finding a possible defect with testing whether it could be exploited and preparing a repair, with an engineer reviewing each proposed change.

The agreement added another major supplier to OpenAI’s computing arrangements, which had previously centered on Microsoft.

The licenses showed how far access to frontier chips now ran through government approval.

The report described Apple turning to a competitor’s model to rebuild its assistant.

As reported, the rule would close much of China’s data center market to Nvidia, AMD and Intel and steer demand to Chinese chipmakers.

The release gave developers a large reasoning model they could host and adapt themselves, with Moonshot’s published results emphasizing agentic search and coding tasks.

Anthropic, which planned to use up to one million TPUs, was named among the customers, giving Google’s in-house chips a prominent outside buyer.

The deal reversed the usual direction of payment: an AI company paid a social platform for access to its users.

The document set out a provider’s proposed principles for protecting younger users, as companion chatbots and teenage use drew scrutiny from regulators.

The figure placed Meta among the largest planned spenders on AI computing in the country.

The release widened the range of languages available for transcription research and applications, including hundreds never before supported.

The projects addressed both research questions and the practical difficulty of working across hospitals that cannot pool patient data.

The report came months after Meta reorganized its AI work under Superintelligence Labs and shifted its emphasis toward rapid product development.

The update treated communication and response timing as parts of model usability, reserving more reasoning for difficult tasks.

The program tied Anthropic’s future model development to construction and power supply in the United States.

The results gave an industry-standard measure of how quickly the newest systems could train large models.

The work explored designing small experimental models for interpretability from the start, as a possible route toward understanding larger ones.

According to Anthropic’s investigation, the attackers used an AI agent to carry out most of the operation’s steps, with people intervening at a few decision points.

The surrogate let a physics simulation skip its most expensive small-scale calculations.

The release made interaction quality a prominent part of model competition, measured through preference and role-play evaluations.

The launch connected a new model generation to several of Google’s established products at once.

The design shifted attention from individual code suggestions to supervising tasks, giving developers work products to check and comment on.

The agreements linked investment, computing supply and model distribution, and made Claude available on all three major US cloud platforms.

The release addressed continuity during extended software work, letting a single task outlast one context window.

The offer targeted classroom preparation and school administration, with privacy terms intended to support compliance with student-records law.

The results arrived amid debate over whether spending on AI infrastructure could be sustained.

The release treated openness as access to a process as well as a finished model, making it easier to study how training choices shape behavior.

The collaboration addressed the equipment needed to turn computing plans into operating facilities.

The release combined a new model with a lower price for Opus-level access, about a third of the previous Opus rate.

The order framed AI for science as an institutional program involving national laboratories, infrastructure, data access and coordination.

The plan expanded the infrastructure available for public-sector computing, including work at the Top Secret and Secret levels.

The proposal addressed AI by raising the stakes for existing fraud offenses.

The revision applied the ordinary inventorship test to AI-assisted work, withdrawing the separate framework issued in 2024.

The study asked whether safeguards survive a change in literary form, and found that style alone could shift a model’s response.