Anthropic proposes abstractive red-teaming to test model character violations at the query-category level
A team from the Anthropic Fellows Program proposed abstractive red-teaming, which searches over natural-language query…
AI news monthly archive: March 2026 features 171 curated AI news items covering model releases, product updates, industry news, papers, tutorials and opinions, each with an AI summary, source credit and original link.
A team from the Anthropic Fellows Program proposed abstractive red-teaming, which searches over natural-language query…
Today's frontier reasoning models are trained along a path highly consistent with AlphaGo: first imitating large…
Anthropic announced the establishment of The Anthropic Institute, led by co-founder Jack Clark as head of Public…
Ten years ago, AlphaGo's match in Seoul opened the modern AI era; the iconic "Move 37" proved AI could tackle…
Gemini has rolled out several AI features for Google Workspace: generating Doc drafts in seconds from selected…
IH-Challenge trains models to prioritize trusted instructions, improving instruction hierarchy and safety…
ChatGPT adds interactive visual explanations for math and science, enabling real-time exploration of formulas,…
AMI Labs announced a $1.03 billion seed round led by Cathay Innovation, Bezos Expeditions and others, a record for a…
I am joining forces with @ylecun and an incredible group of people to start AMI Labs @amilabs.
Hugging Face Hub launched Storage Buckets, a mutable, S3-like object storage service designed for machine learning…
In synchronous reinforcement learning training, data generation is the main bottleneck: generating a 32,000-token…
Thinking Machines Lab and NVIDIA announced a multi-year strategic partnership to deploy at least 1 GW of…
METR had four maintainers of scikit-learn, Sphinx, and pytest blindly review 296 AI-generated PRs and found that about…
The LangChain team introduced its GTM Agent built on Deep Agents, which automates the sales outreach process end to…
The OpenAI team explains how it combines Codex with repository-level skills, AGENTS.md, and GitHub Actions to turn…
A research team released Ulysses sequence parallelism, a new technique for training large language models. By…
Hugging Face released LeRobot v0.5.0, which expands comprehensively across four dimensions: datasets, models, tasks,…
The evolution of autoresearch should move toward asynchronous large-scale collaboration, similar to the SETI@home…
Runway launched the Characters API, built on the GWM-1 world model, which generates real-time conversational virtual…
GPT-5.4 performs well in coding, knowledge work, computer use, and more, and it is great to see people like it so…
Codex Security is now available in research preview. The AI application security agent analyzes project context to…
OpenAI introduced a CoT controllability evaluation suite and research paper. Tests found that GPT-5.4 Thinking…
GPT-5.4 Thinking and GPT-5.4 Pro are rolling out to ChatGPT users and are also available via API and Codex. This…
Anthropic partnered with Mozilla to audit Firefox security using Claude Opus 4.6. The model found 22 vulnerabilities…
During BrowseComp benchmark testing of Claude Opus 4.6, researchers found 11 answer leaks among 1,266 questions. Nine…
Sarvam AI has open-sourced Sarvam 30B and Sarvam 105B, two reasoning models trained from scratch, under the Apache 2.0…
A Manus blog post tested 8 AI portfolio builders with the same strict prompt, evaluating design quality, AI…
In a blog post, Manus tested seven AI mobile app builders—Manus, Lovable, Base44, Replit, Bubble, Figma App Builder…
A Manus content author tested Google Veo 3 with four prompts, finding dialogue lip sync reaches documentary quality…
NXP published a technical blog on Hugging Face detailing the complete process of applying vision-language-action…
OpenAI releases the GPT-5.4 Thinking system card, disclosing the architecture details, safety evaluation framework and…
OpenAI launches GPT-5.4, its most capable and efficient frontier model for professional work, supporting a 1 million…
OpenAI publishes CoT-Control research finding that reasoning models have difficulty manipulating their own chain of…
Five AI value models show how leaders can deploy AI progressively, from raising employee proficiency to fully…
OpenAI releases ChatGPT for Excel and new financial app integrations, powered by GPT-5.4, to accelerate modeling,…
Together AI released FlashAttention-4, which uses algorithm-kernel co-design to maximize parallel overlap between…
A joint study by AISI and Irregular found that recent frontier models can effectively use reasoning budgets far beyond…
Canvas in AI Mode in Google Search is now available to all users in the United States, adding document drafting and…
I’ve been trying Copilot Tasks this week, and it’s awesome to assign tasks (even recurring ones) in this collaborative…
Officials hinted that version 5.4 will be released earlier than widely expected, with the new version coming soon.…
Page 4 of 5
Sponsored
Orange Relay (pro.nat6.net) is a premium AI API relay — an excellent transfer station with competitive pricing, reliable quality and broad model coverage including Claude, GPT and Gemini.
Visit Orange Relay