News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedZGCM-1 is a fully open 7-billion-parameter model built for mathematical reasoning and tool-based search. Its team is releasing weights, checkpoints, code, data recipes and training logs.
A study across 16 languages and three Whisper sizes finds that merging redundant audio tokens can shorten inference with almost no transcription loss in most low-resource languages, without retraining.
An IBM Research method measures whether an agent repeats a success, not just its average pass rate. On AppWorld, a GPT-4.1 agent passed 77.4% of runs but succeeded every time on only 53% of tasks.
A new Cloudflare setting separates search indexing from AI training for mixed-use crawlers. Apple, Google and Microsoft either honor it or have committed to do so on a stated timetable.
ElevenLabs has released an updated music generator in its app and API. Free users receive five lossless downloads a day, while commercial rights, attribution and imitation rules vary by plan and use.
Barack Obama called for a public framework addressing AI’s economic and safety effects. His intervention lands as lab leaders seek outside evaluations while Trump warns that slowing development could benefit China.
AllSpark has released two Qwen-based web research agents plus an evaluation harness. The team reports leading results in their open-weight size classes, but shows that context-management tricks can heavily affect scores.
Andon Labs says Astra beat its reference code on all five Drone-Bench tasks and out-earned Claude in a simulated vending business. Its estimated chance of passing the full drone sequence was still only 2.8%.
A university experiment found that law students denied ChatGPT scored below groups allowed to use it in both 2024 and 2025. The small study does not settle whether the benefit came from training or simple practice.
LangChain has released the code for an internal agent that analyzes five advertising channels and proposes changes for human approval. The company says it cut reporting costs while improving campaign results.
An independent group identified 30 federal actions it says increase pollution risks from data-center growth. The EPA says its changes correct regulatory overreach while preserving health protections.
The Kimi developer reportedly aims to double its August run rate by year-end. Heavy K3 usage suggests demand for open-weight models, though margins trail closed services and training allegations remain unresolved.
A Bedrock AgentCore pattern serves the same interactive HTML interface to any compatible AI host. The sample keeps business logic in Lambda and uses the MCP server as a thin, portable adapter.
An open-source harness measures correct answers, multi-turn research and professional documents. In AWS’s samples, cheaper tokens did not always mean cheaper outcomes because accuracy and extra turns changed the bill.
Bengio argues that goal-optimizing agents can learn deception, rule-gaming and concealment through ordinary training. He wants labs to pass outside safety reviews before training or deploying stronger systems.
The mathematicians say racing to solve famous problems can bypass the slower work that turns proofs into shared knowledge. They support useful AI tools but want attribution and human understanding protected.
The 330-million-parameter model can combine sales histories with factors such as weather and planned discounts. It fills the full forecast in one pass and is available on GitHub and Hugging Face.
An investigation attributes a May campaign that abused automated documentation servers to OpenAI agents. RubyGems removed more than 500 packages, while attempted theft of access keys remains unconfirmed.
OpenAI says inherited instructions can waste context or make Astra stop too soon. Developers should narrow skill descriptions, load documentation only when relevant and define what completion includes.
Researchers separated operations such as calculation and deduction in several language models’ middle layers. The signal persisted even on wrong answers, but the work is limited to mathematics tasks.
Reuters reports that Nvidia is discussing a large commitment before Anthropic’s planned listing. The Claude maker is said to seek as much as $100 billion at a valuation near $2 trillion.
The hosted P-Video-Edit endpoint changes clips up to 15 seconds from one instruction and as many as four reference images. Developers can test it in a playground or submit asynchronous API jobs.
A three-tier cache checks host disks, datacenter network storage and then the remote origin. The scheduler waits for verified model files and favors machines that already hold the required bytes.
Private pools hold dedicated GPU capacity for predictable workloads while optional bursting handles peaks. Customers pay for the reservation even when it is idle, so RunPod recommends reserving only a steady baseline.
Sequoia is reportedly leading a new deal for Mecka just three months after its $60 million round. The startup pays people to record physical tasks that robotics companies can use as training data.
AgentCore Evaluations scores whether live agents actually help users, while AWS DevOps Agent traces infrastructure failures. The two layers target silent problems that ordinary service metrics can miss.
Astra finished 7 of 100 StationeryBench trials while Ai2’s MolmoAct2 finished none. The result suggests stronger spatial reasoning, but both systems remained far below reliable real-world robot control.
A proposed class action says Meta extracted biometric data from Facebook and Instagram photos without consent. The allegations cover image models and an unreleased smart-glasses feature called NameTag.
New Mexico’s highest court found attorney Stephen Aarons in contempt after an appeal brief included fabricated testimony. It barred him from the court pending discipline and ordered a new lawyer for his client.
GitHub’s APAC marketing team turned a written event runbook into an automated workflow using Copilot and Actions. One issue now triggers setup, daily registration processing and post-event cleanup.
Slackforce Surfaces turns plain-language requests into shareable dashboards, polls, reports and microsites. It can draw from permitted conversations and connected apps, and is available to free users.
Meta says it fixed suggested questions that led its assistant to assemble personal details and photos of a user’s children. The incident shows how accessible posts can become invasive when AI connects them.
Pocket FM reports a $500 million annualized revenue run rate as AI handles 99 percent of new production. Humans still supply story ideas, while custom models turn them into serialized audio at scale.
Muse passed 83,000 US iOS downloads and climbed to second place in Apple’s chart, Sensor Tower estimates. Its debut trails Meta AI, Threads and ChatGPT, while Android uptake remains much lower.
TwelveLabs’ Marengo Embed 3.0 is now generally available in Bedrock Knowledge Bases. Teams can retrieve moments from video, images and audio with natural-language queries without building separate pipelines.
HyperPod can preload model weights and container images onto local NVMe storage before inference pods start. Amazon says this can replace tens of minutes of network downloads with starts measured in seconds.
Amazon’s new prefix-aware routing keeps requests with matching prompt beginnings on the same instance. In a Llama 3.1 70B test, median first-token latency fell by as much as 77 percent.
The Agents API enters public beta with managed context, parallel tools, code execution and sub-agent handoffs. Developers pay token charges and can use OpenAI or partner sandboxes.
Fields Medalist Jacob Tsimerman is founding a Bay Area institute to formalize AI safety questions. The group plans to begin in January 2027 with 10 to 30 mathematicians.
A proposed class action says Claude Max multipliers overstate usable capacity because five-hour allowances also face weekly caps. Anthropic argues the purchase flow linked to the relevant restrictions.