News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedApple researchers analyzed human-like behavior in large language models, testing how prompts and user factors affect traits such as emotion, boundaries, and relationship-building.
Netflix open-sourced an agentic workflow for observational causal inference that uses an actor-critic loop to estimate causality and draft reports.
Asana used Codex to replace an outdated testing system in two weeks, completing work the company estimated would take five years for about $12,000.
A JAMA opinion piece argues regulators should not hard-code doctors as the final step for autonomous medical AI, while noting evidence is still mostly simulated.
Former SpaceX engineers are building an automated factory for steel parts, using robotics while stopping short of promising full autonomy from the start.
AWS added NVIDIA Nemotron 3.5 Lightning to SageMaker JumpStart, offering a 30B mixture-of-experts model built for high-volume agentic workloads.
A Hugging Face post says changing workload order, not hardware, raised GPU utilization by 33 points on the same cluster, highlighting operational AI gains.
Cloudflare introduced WriteGuard in private beta to add fine-grained controls for MCP servers, focusing on agent actions that modify data or systems.
Warp launched Warp Factories, a system meant to make it easier for teams to build AI software factories with ready-made development infrastructure.
Mozilla’s Firefox Smart Window can now pull current web information into AI chats, cite source links, suggest tab groups, and preview previously visited pages.
An arXiv paper argues that FLOPs alone are a weak proxy for AI efficiency and calls for replication-focused assessments that reflect real work.
A new arXiv paper introduces HarmProfile, a benchmark of more than 80,000 validated harmful-output artifacts from 23 frontier LLMs.
OpenAI is partnering with CodeAI to help students build AI literacy, critical thinking, and responsible-use skills as schools adapt to generative tools.
Dario Amodei said AI centralizes around compute and infrastructure, while critics accused Anthropic of using safety concerns to seek regulatory advantage.
Z.ai’s new open-weight model may help defenders test systems, but experts warn the same capabilities could assist hackers.
Microsoft Copilot revealed a secret input tied to an attack path, showing how hidden agent instructions can become security risks.
Axonius built multi-tenant AI agents on Amazon Bedrock AgentCore, using managed isolation, authentication, and observability across customer environments.
Hugging Face details multi-vector embedding models in Sentence Transformers, a retrieval approach that keeps richer token-level signals for search and ranking.
Microsoft Foundry added five Claude capabilities, including web search, web fetch, MCP connector support, tool search, and structured outputs for agent workflows.
Apple researchers studied GRPO training outside English and found native-language reasoning can narrow the gap, highlighting limits in English-centered AI evaluation.
Stable Audio 3.0 now has a digital audio workstation plugin and a more advanced web generation experience, giving music creators more direct ways to use the model.
An InfoQ article argues that AI agents can extend evolutionary architecture checks to judgment-heavy risks while deterministic rules remain best for measurable constraints.
Grab reports that AI agents reduced mechanical analyst work from 44% of tickets in February to 30% in June by handling data discovery, SQL, validation, and drafts.
Bloomberg reports that Stripe has finalized a deal to acquire OpenRouter, an AI model gateway that gives customers access to more than 400 models.
Wispr raised a $280 million Series B at a $2 billion valuation and introduced a new speech model as it moves from dictation into meeting notes and broader workflows.
Groq raised $350 million at a $3.5 billion valuation as it moves from building its own inference chips toward operating Nvidia-powered AI infrastructure.
Amazon is buying and destructively scanning rare books for AI training, citing a 404 Media investigation that tracked one book to a Las Vegas facility.
The Financial Times reports that OpenAI disbanded its preparedness team and moved risk work into area-specific groups such as bio and cyber.
Anthropic says Claude-generated text will receive invisible watermarks based on Google's SynthID-Text approach as the company prepares for EU AI Act transparency rules.
A new arXiv paper proposes the Behavioral Consistency Metric to capture how similarly language-model agents behave across tasks, not just whether they succeed.
A new study examines stable miscalibration, where a language model keeps a confident wrong answer even after small prompt changes or self-critique.
An ICML position paper says evaluation focused on autonomous superhuman performance may push AI in the wrong direction and should include human-AI collaboration.
BCMT, a proposed transformer architecture, separates local attention from global memory to handle long context without dense attention across every distant token.
A routing paper introduces token inflation, showing that retries in agent workflows can make real costs far higher than a model's listed per-token price.
An arXiv paper proposes Self-Explainable Latent Reasoning, a framework meant to keep compact hidden reasoning efficient while generating human-readable explanations.
A new arXiv study questions a common assumption in AI coding: semantic retrieval through language servers can improve precision, but often does not reduce token use.
OpenAI is funding 14 independent projects focused on economic opportunity and resilience, arguing that access to AI is not enough without policy choices around its benefits.
It will secure about 8 gigawatts of IT capacity at Ohio's PORTS-Pike campus, alongside local grants, student Codex credits, and job commitments.
AI is making cyberattacks easier to automate while also giving defenders stronger tools, arguing that organizations need to close basic security gaps quickly.
Google is pairing Gemini and Pixel with five major football clubs to provide match insights, behind-the-scenes content, and club-specific fan material.