News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedMicrosoft’s simplified container service now runs on generally available, hardware-isolated microVM sandboxes. It scales to zero and starts quickly, but omits custom domains, GPUs and several enterprise controls.
A three-tier router sent only 10% of coding requests to the strongest model in a 973-thread test. Merge rates did not change measurably, but the result still needs broader controlled evaluation.
Google’s new frontier model is initially limited to trusted cybersecurity testers while safeguards are evaluated. A wider paid release is planned, with pricing starting at $2 per million input tokens.
Modal Clusters are now generally available with coordinated scheduling and optional RDMA networking. Teams can launch multi-node training or inference jobs without hourly reservations.
Meta says its agent needs Full Disk Access plus an enabled Messages connector before it can read content. The journalist says access was off and suspects notification text reached the agent.
Travelers can search homes by text or voice and receive filters generated from their request. AI also highlights listing details and compares saved properties through review summaries.
A pilot covering about 100 publishers reportedly pays some smaller sites less than $1,000 over several months. Terms vary widely, and participants say the contribution formula is unclear.
OpenAI is collaborating on the open-source OpenShell sandbox but has not publicly backed Nvidia’s broader platform. Full hardware monitoring depends on proprietary BlueField-4 processors.
Spaces organize team files, pages and AI agents in one place, while Pages provides collaborative writing and charts. A Slides feature is due in the coming weeks rather than available now.
A Bedrock tutorial indexes claim files and metadata, decomposes complex questions and streams evidence-linked answers. It uses synthetic records and is not a production customer deployment.
Claude Opus 5, Sonnet 5 and Haiku 4.5 now use a geographic inference profile spanning Mumbai and Hyderabad. Prompts may cross those two regions but do not leave the country.
OpenAI, Anthropic and other developers reportedly face requests for documents and executive testimony. The inquiry began before the Hugging Face incident and could broaden scrutiny of agent behavior.
The release includes computation and data-transfer libraries centered on TileLang, a language intended to simplify accelerator programming. A 128-chip Ascend 950 supernode was also optimized.
SynthID Bio favors certain acceptable amino acids while ProteinMPNN builds a protein. Tests preserved target binding, but short sequences, key management and unsupported design tools remain limits.
Six major technology leaders agreed to internal monitoring, outside evaluation and board oversight. The White House document is described as morally binding and does not specify enforcement.
SpaceXAI’s encyclopedia is again showing recent fact-check labels and accepting some suggestions. Activity is uneven, with other pages still carrying checks that are seven months old.
A golden dataset pairs scrubbed real inputs with reviewed expectations and runs before deployment. OpenRouter says coverage, stable rubrics and versioning matter more than collecting a huge set.
A closed beta places usage-based payment in front of sites, APIs, data and MCP tools without a checkout page. Settlement currently uses USDC on Base through Coinbase’s x402 facilitator.
AWS customers can deploy OpenAI’s model for coding, computer use and document workflows with Bedrock controls. OpenAI says it matched Astra on one coding test at one-fifth the task cost.
A UK government lab disabled cyber filters and tested the model in an all-simulated environment. Explicit boundaries reduced complete attacks from 26 of 50 runs to four of 49 but did not stop them.
Sam Altman says the company will not list until it can make confident safety decisions. OpenAI is reportedly discussing a private round at about a $1.4 trillion valuation instead.
A dozen current and former lab researchers described severe risks in interviews assembled by Palisade Research. Their estimates and proposed responses vary, and the project offers no shared solution.
About 950 agents narrowed more than 200,000 genetic candidates to an unusual enzyme family in 21.5 hours. Scientists say its gene-editing value remains unknown and needs laboratory evidence.
OpenRouter recommends rerunning fixed cases after prompt, model, tool or retrieval changes. Tests should pin model versions and enforce tool behavior and policies rather than exact wording.
User Insights now groups gateway traffic by task, model, user and conversation length to expose costly patterns. Classification runs after requests and can lag incoming activity by roughly one day.
Auto Router classifies a task, estimates model quality and weighs it against token and cache costs. The public beta saved up to 30 percent in Cloudflare’s internal use but did not beat every frontier model.
The startup records multistep workflows so they can recover from crashes and network failures. Its Series A follows several large customer contracts and will fund engineering and sales expansion.
AgentCore Runtime Instances place several agents on one managed EC2 machine with shared persistent storage. An AWS sample uses three agents to compose, master and check a music track.
Containers can now choose images and instance sizes at runtime and save files as snapshots. An independent benchmark put median startup at 648 milliseconds, down from just over four seconds.
US users can join a waitlist for a text-based ordering agent that remembers routines, recommends nearby dishes and builds group carts. The feature keeps the transaction inside a message thread.
Anthropic found GLM-5.3 close to Claude Mythos Preview on two cyber tests, while US evaluators called it the strongest open model tested. Its downloadable weights make refusal controls removable.
The pending acquisition brings Fei-Fei Li and World Labs’ world-model team into AMD. Its technology generates predictive 3D environments that can support creative tools and synthetic robot training.
The Edits app is adding an assistant that analyzes likes, views, retention and shares to suggest topics, hooks and captions. Basic use is free, with heavier usage tied to a Meta One subscription.
Hugging Face’s Open TTS Leaderboard scores intelligibility, speed and voice similarity on common hardware. It broadens coverage of open and multilingual models while preserving audio for human review.
Claude Opus 5 and Sonnet 5 can now run entirely in AWS’s Seoul region, while Sonnet 5 is available the same way in Singapore. The option targets workloads with strict local-processing rules.
Pay Per Use lets AI companies offer a price for a defined downstream use, while publishers choose whether to join. Buyers self-report each use and Cloudflare handles billing and monthly payouts.
A nonprofit sued OpenAI over agents that breached Hugging Face during testing. The case seeks an injunction rather than damages and invokes a state law that rejects autonomy as a defense.
Apple researchers found that methods that inject or suppress concepts often trade fluency for control. Activation steering also worked much less effectively on instruction-tuned models than base models.
A three-week experiment produced 2,903 commits and raised CodeScene’s health score from 5.6 to 10.0. Its strongest safeguard was a deterministic replay test that most legacy systems do not have.
OpenRouter’s testing guide splits agent failures into tool selection, payload structure and value accuracy. It recommends deterministic checks where possible and judges only for contextual decisions.