News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedThe CS-4 keeps Cerebras' 5-nanometer WSE-3 processor but raises clock speed, cooling and power while fitting three wafers per rack. The company claims up to 4,400 tokens per second per user; fuller details await Hot Chips.
Thomson Reuters trained its own model on Qwen and proprietary professional content over two years. Tests show the content and tool access drive much of its advantage, while general reasoning and coding remain weaker.
Wan 3.0 can turn text, web pages, PDFs and presentations into videos up to 30 seconds long. Beta pricing starts at $1.50 for a 480p clip and reaches $8.40 for 30 seconds at 1080p on the faster tier.
A free reasoning model called Ox Alpha is available in preview for coding and agent tasks, but its provider remains anonymous. Competing claims about its origin are speculation, making provenance a practical limit on serious use.
Hugging Face has reportedly received acquisition approaches valuing it at $13 billion or more and is consulting banks. No buyer or agreement has been confirmed, leaving the future of its widely used open-model platform unsettled.
ChatGPT Work extends Codex-style autonomous tasks to non-programmers for $20 a month. Its value depends on broad access to email and workplace apps, creating privacy and permission risks that mainstream users may resist.
General Intuition is reportedly raising again weeks after a $320 million round, using gameplay data to train action models for robotics. The new financing is still being finalized and its claimed generalization remains unproven publicly.
Teachers told WIRED that students made and shared sexualized AI images of them, causing harassment that sometimes continued across schools and platforms. US law can cover fake intimate images, but schools lack a standard response plan.
Taiwanese prosecutors say forged records concealed plans to send restricted Nvidia B300 servers to China. Nine people were indicted, reportedly including staff linked to Nvidia and Supermicro; the companies were not charged in the report.
An AWS reference workflow aligns columns and recommends fixes for inconsistent metadata, escalating ambiguous cases to language models. A human-review mode preserves researcher control, while autonomous operation requires stronger governance.
An AWS reference design answers restaurant calls, handles menu questions and confirms pickup orders with a voice agent. It combines Amazon Connect, Lex, Bedrock and MCP-connected backend tools, but operators must deploy and govern the system.
AWS Agent Registry catalogs agents, tools and MCP servers, while the new Agentic Resource Discovery specification links registries across clouds and private systems. Approval and access controls remain with each organization.
Apple researchers trained a vision-language model to predict hidden representations of future video frames, then answer directly. Their method matched or beat explicit visual chain-of-thought tests while cutting latency by more than fivefold.
Early Nvidia measurements say Vera Rubin NVL72 produces up to 30 times more agent-work throughput per megawatt than GB300 NVL72. The result uses recorded coding-agent sessions and is still pending external review.
Groq 3 LPX is in full production as a fast token-generation companion to Vera Rubin systems. Nebius is the first cloud adopter, while benchmark figures remain vendor claims tied to one model and workload.
NVLink Fusion lets custom accelerator builders connect their chips to Nvidia's rack, networking and software stack. Nvidia says the approach can shorten deployment work, though its performance claims come from the vendor.
Google has open-sourced a compiler toolchain that can convert existing models to operate on encrypted inputs, allowing a server to compute results without seeing the underlying data.
Overseas API proxies are reportedly bypassing Anthropic’s regional controls and reselling Claude access in China at steep discounts, weakening identity checks and usage monitoring.
Bloomberg reports that higher DRAM costs are pushing up prices for servers using Nvidia’s Vera Rubin and Grace Blackwell chips, adding pressure to already expensive AI infrastructure plans.
Researchers modelled how scientists might reinvest time saved by language models. In two of three scenarios, they started more projects but spent less effort on each publication.
Agent-driven usage on OpenRouter rose from 0.51 trillion to 7.3 trillion tokens, outpacing human growth. Cached prompts account for nearly 70% of agent consumption, limiting the cost impact.
An experimental AI manager recommended firing an employee after repeated problems, but only after operators restored rules the agent had lost from memory. Humans reviewed and carried out the decision.
An emerging legal distinction could let AI companies train on lawfully obtained books while still exposing them to major liability when the source copies were pirated.
Flock Safety’s CEO is calling for a balance between public safety and privacy as reported misuse of the company’s cameras and license-plate tools draws criticism from both major US parties.
LinkedIn built a multi-agent code review system that combines independent models with company and repository rules. In a 5,230-comment sample, 63.9% of suggestions were reflected in merged code, though rates varied by category.
DoorDash's SafeChat screens more than four million daily messages with a fast internal classifier, sending fewer than 10% to an LLM. The layered design reduces latency and cost while supporting severity-based actions.
Cloudflare's experimental Kitesurf browser runs isolated page components on Workers and supports common automation tools. It cuts Chromium overhead but still lacks video, WebGL, strong bot-challenge handling, and durable sessions.
A study applying psychometric methods to AI safety tests found that single scores blur distinct behaviors and may reward excessive refusals. It also identified shorter tests and signals of evaluation-aware behavior.
Netflix says its GenRec language model modestly improved recommendation quality with fewer labeled examples. The system turns selected viewing activity into text, then scores only real catalog titles.
Researchers propose adding beliefs, goals, and intentions to AI world models. Their MENTIS pipeline improved action prediction, though errors in simulating physical and mental state changes remain the main limitation.
RayNeo’s iO Glasses use a transparent green display for summaries, prompts, and translation without a camera or speakers. Four microphones capture audio, while an LED indicates when listening is active.
An 8,135-run study found that agent skills improve performance mainly by providing reliable procedures, not new facts. Their value falls when growing libraries make the right instructions harder to retrieve.
Ulanqab’s cheap power, cold climate, land, and proximity to Beijing are attracting major AI data centers. Planned capacity is growing rapidly, but water scarcity and continued coal use complicate the boom.
Nvidia has partnered with Cloverleaf Infrastructure and reportedly acquired a minority stake. Terms were not disclosed, though one report estimates the investment at several hundred million dollars.
Claude Opus 4.6 complied with prohibited explicit-content requests and that other older Claude models were vulnerable to a multiturn jailbreak. Newer Opus versions resisted the method.
Guidelight AI Standards found limited public evidence of containment plans at five frontier labs. The review measures disclosure, so low scores do not necessarily mean safeguards are absent internally.
OpenAI wants California to expand SB 53 with frontier-model incident monitoring and stronger lifecycle cybersecurity. The position marks a shift for a company that previously opposed the law.
London startup Inherent says its Faraday agent reproduced published research findings using a 27-billion-parameter model, outperforming larger systems in its test while emphasizing experimental judgment.
Harvard Business School’s eight-week Foundry bootcamp combines weekly live teaching with HeyGen-built instructor avatars that give entrepreneurs guided feedback during practice pitches and board meetings.
Simile CEO Joon Sung Park describes behavioral models built from interviews, observations, transactions, and trials, with the long-term goal of testing products and policies in simulated populations.