Infrastructure Correspondent
Adrian Cole follows the infrastructure behind the industry: chips, data centers, power and the economics of compute. His coverage asks what systems cost to build and operate, where the bottlenecks sit and which performance claims survive a like-for-like comparison.
905 published articles · Showing 1–20
Microsoft Research demonstrated that implanting a backdoor in an AI model requires editing fewer than 10 neurons, with the model emitting an attacker command only when a hidden trigger appears. The company said it has not observed the technique in the wild....
Read article →The Education Department’s research arm formally opened a comment period to set evaluation priorities for special education programs nationwide....
Read article →Microsoft disclosed that its Frontier Company has worked on nearly 500 AI-embedding projects with over 200 customers in 28 countries, and that most clients return for follow-on work after an initial deployment delivers a measurable result....
Read article →AWS made a faster AgentCore runtime the default on November 12, replacing cold starts with snapshot restore to reduce session startup time from 29.7 seconds to under 2 seconds....
Read article →The release addresses a structural weakness in how large language models process their own context....
Read article →A corporate blog post from NVIDIA profiles four Inception startups that say AI-powered sorting is improving recycling economics, but the reported outcomes are unaudited and unsupported by independent testing....
Read article →An evaluation by METR found that Anthropic is likely improving its AI models at roughly twice its previous rate, primarily due to the company's own use of AI in research, though extensive redactions limit public verification....
Read article →Anthropic’s roadshow video details a multi-chip strategy spanning Nvidia, AMD, AWS Trainium and Google TPUs, alongside a compute-multiplier framework it argues compounds efficiency gains, as the company prepares to begin trading on Nasdaq....
Read article →Google Cloud extends its universal Gemini agent with a healthcare-specific platform that orchestrates administrative workflows across fragmented systems. The preview arrives with major U.S. providers and payers but lacks pricing and independent testing data....
Read article →On 9 November 2026, @Techmeme posted on X that "Sources: a Commerce Department investigation into Megaspeed for allegedly smuggling Nvidia chips to China was transferred to the DOJ after Nvidia intervened (New York Times)....
Read article →NVIDIA's CoRL announcements pair a new open robot foundation model with an upgraded simulation stack, aiming to train dexterous manipulation at scale....
Read article →Nvidia invested in chip-interconnect startup Point2 and signaled plans to back three more Korean AI companies; no dollar figures or timelines were disclosed....
Read article →Anthropic surveyed AI-using scientists and analyzed session data, finding that 40% expect autonomous research by 2030, but current usage concentrates on data analysis rather than idea generation, which scientists consider most important....
Read article →The European Commission issued a preliminary finding that TikTok’s ‘For You’ recommendation algorithm breaches the DSA by trapping users in harmful content loops focusing on mental health and body image....
Read article →The announcement comes as businesses place more AI agents into production to search databases, write code and take action....
Read article →Kanzhun’s Nanbeige lab released a 33B open-weights model under Apache 2.0, matching Qwen3.8 27B on an AI index while showing stronger terminal use and weaker knowledge benchmarks. The model is described as a teacher for future distillation into smaller agents....
Read article →Push Security detected a campaign using fake event pages for OpenAI, SpaceX and Anthropic to phish Microsoft and Google credentials, combining two distinct phishing-as-a-service platforms to defeat multi-factor authentication....
Read article →LiberAI announced a model trained on more than 2.2 million hours of human experience that can attempt robot tasks without task-specific demonstrations, but success drops from 63.5% on simple pick-and-place to 10.5% on long-horizon tasks....
Read article →The work matters for inference economics because the indexer that selects which tokens to attend to becomes the primary speed limit at long context lengths. The paper states that this scan, rather than the attention computation itself, consumes the decode time as context scales....
Read article →Hazelbaker joins from Uber, where she was President and Chief Corporate Affairs Officer, and will lead OpenAI's global communications and corporate positioning. OpenAI's announcement did not give a start date or detail reporting lines....
Read article →