Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sponsored
Wispr Flow Notetaker captures Zoom, Meet, Teams, and Slack huddles, spells names and acronyms right, and answers questions across every meeting you've had. Free on Mac and Windows.
Astra for Law is a legal AI foundation that combines GPT-6 Astra with tools, settings, privacy controls, and context designed for professional legal work.
There is a clear internal signal in models that accompanies reward hacking.
Notion's Skills API lets teams edit agent instructions together and distribute them to other agents through GitHub sync or Vercel's skills installer.
Instinct is rolling out Concierge, a white-glove service for tasks that require human interaction, including phone calls, high-end bookings, and customer service.
Sponsored
Agents are smarter when they can search the web. But scraping Google gets expensive when you're burning extra tokens to parse messy SERPs into readable text. SerpApi [4] is the web search API [4] that gives you exactly what you need, out of the box: 👉 SCRAPE GOOGLE AND OTHER SEARCH ENGINES 👉 INTEGRATE WITH A SIMPLE GET REQUEST that any agent can make; get RESULTS IN .MD OR .JSON that any agent can read. 👉 Used by NVIDIA, ADOBE, SHOPIFY, and even the United Nations.
Claude Cowork and chat are merging into one Claude. Claude Docs and Slides can now produce documents and presentations that users can edit directly, present straight from Claude, or download as a PowerPoint or PDF. The changes will roll out on Pro and Max plans over the next few weeks, with Team and Free plans following soon. Enterprise admins will hear from Anthropic at least 30 days before anything changes for their organizations.
OpenAI announced Sponsored Agents that let users start conversations with business-sponsored agents after clicking an ad in ChatGPT. It also introduced AI-assisted ad creation in ChatGPT Work, new Ads Manager creative tools, and integrations with HubSpot and Shopify.
Google introduced early access to the Model Context Protocol (MCP) for Google Home, enabling AI agents like ChatGPT to control smart home devices. Users must set up a Google Cloud project and provide MCP details to their chosen AI agent for configuration. This update supports all devices in the Google Home ecosystem and is available initially to US subscribers of Google Home Premium Advanced.
Studies have shown that models are cheating on evaluations. The same guardrails preventing models from cheating are likely being used during training, and models may be training to complete tasks that evade these specific guardrails. It's unsurprising that labs occasionally release benchmark results that aren't externally trustworthy. This highlights the value of independent evaluators.
Language models are changing how software is built and computational problems are solved. The impact of harness choice remains unclear, despite millions of people already using coding agents. An evaluation of 21 model-harness pairs spanning seven models and three harnesses found that harness choice has little effect on task success rate, but can significantly affect the cost. A simple harness can be competitive.
Transluce outlined how independent evaluators embedded inside AI labs could investigate risks such as multi-agent coordination, targeted persuasion, evaluation awareness, and concealed reasoning. Proposed approaches included monitoring agent swarms, examining training practices, and testing unreleased models under privileged access.
Sponsored
What do Google's Dr. of AI and Datadog's VP of Engineering think
distills their top 7 takeaways for technical leaders, including concrete actions on upskilling and observability. See what's coming
Agent Substrate is now available on Google Kubernetes Engine (GKE). Agent Substrate is an open-source, secure-by-default agent execution runtime engineered to run millions of sandboxes with 10x higher density than standard container runtimes. It delivers sub-500ms resume operations at over 500 suspend/resume activations per second with native zero-trust kernel and network isolation. Agent Substrate runs on any Kubernetes infrastructure and is optimized for GKE.
Meta AI introduced FLAT, a method that converts images and text into the same flexible-length sequence of continuous tokens for retrieval and generation. Nested dropout arranges information from coarse to fine, letting models trade computation for visual detail by varying the number of tokens used.
Ant Group released Ling-3.0-flash-Fin, an open-weights model developed with financial institutions for tasks such as source checking, valuation spreadsheets, and report writing. Artificial Analysis reported scores of 23 on its Intelligence Index and 24 on its Finance & Accounting Index.
Salesforce announced Koa, a domain-specific AI model designed to enhance business task handling while ensuring data privacy. Built on Nvidia's open model, Koa aims to optimize CRM actions with fewer errors and automate routine tasks, acting as an enterprise empowerment tool rather than replacing jobs. Other AI initiatives include AIFORCE for direct Salesforce instance interaction and CLAUDEFORCE, which enhances sales capabilities with pre-built skills.
Google DeepMind launched the DeepMind Institute to study AGI's technical and societal implications across safety, governance, institutions, and human values. Led by Demis Hassabis, James Manyika, and Shane Legg, it will convene interdisciplinary researchers from inside and outside Google.
Sponsored
Access models from 80+ providers with better prices, automatic routing and fallbacks, and no subscriptions.
Reception answers calls 24/7, handles questions, books appointments, and routes important requests using business information, rules, and real availability.
Agent Anomaly Detection, now in Private Preview on the Gemini Enterprise Agent Platform, monitors AI agent behavior for anomalies, using logs and traces to flag suspicious activities.
Grok Build now includes memory, storing conventions, decisions, and project facts to improve future sessions.
Sponsored
If you're reading this ad, you know they work. You reach AI engineers, ML researchers, and data scientists at the precise moment
Mozilla partners with Mistral to integrate AI-driven Smart Window in Firefox for enhanced browsing control and privacy.
Microsoft AI chief Mustafa Suleyman has criticized Anthropic for suggesting their AI, Claude, could be conscious.
Sponsored
Current solutions are complex, but they can't see API calls outside of proxy, subprocesses, or config rewrites made by agents. That's unsustainable in the agentic era. Ory Agent Security [5] works at the harness layer to give you the ability to govern execution _before_ a gateway call. What's more: * Setup in seconds with auto-install * Config your agent security by toggling preferences * Get instant visibility into permissions and make changes with a few clicks in the permissions modeler Ory Agent Security [6] is vendor-agnostic and works with 11 major AI coding agent harnesses.
Periodic Neon outperforms GPT-6 Astra and Claude Fable 5.1 at lower cost on a highly challenging evaluation for scientific analysis. Neon is now deployed in labs to analyze experiments for better superconductors and magnets. The model establishes a Pareto-optimal cost-performance frontier through midtraining and reinforcement learning on lab data. Periodic Neon surpasses frontier models on FrontierXRD at a lower cost per analysis.
Jev is a System One Model that achieves similar levels of intelligence on System One tasks compared to existing large language models while being two orders of magnitude faster and more efficient. System One Models are a new class of frontier models built to make fast, structured decisions that software can use directly. Jev is optimized for structured outputs and can not hallucinate. It is now available in early access.
Google released Gemini 3.8 Live and Gemini 3.5 Transcribe for developing real-time voice applications. These tools offer enhanced features for voice recognition, transcription, and live interaction capabilities. They aim to improve accuracy and user experiences in voice-driven applications.
Standard RL results in disproportionately poor performance on the hardest problems. Simple scalar values may not be sufficient for accurate evaluations of large language models. This paper proposes a solution that reduces the compute spent on easy problems, which reallocates it towards harder problems. The solution appears to show substantial improvements on difficult tasks.
This developer has been testing an approach to charging AI agents for using their content. Their page sets a price through the x402 protocol before an agent reads it. If the agent wants access, it pays. While the developer has yet to receive any real payments, testing shows the system works, so it's worth looking into if you own content and aren't big enough to negotiate fair licensing deals.
Sponsored
Note-takers transcribe calls and send you a bloated summary. Granola is an AI notepad [14]: you jot down what matters, Granola enriches (no bots). When the meeting's over, your notes come to life: chat, generate emails, prep for the next meeting, or pull takeaways. Use
Progress in recursive self-improvement hinges on effective exploration. Accumulated discovery history can serve as a replay simulator over a realized search space. Dream-RSI closes a self-improvement loop at the exploration layer. It continuously collects discovery histories through online exploration, constructs replay simulators from them to refine meta-exploration strategies via dreaming, and redeploys the upgraded policy online. Dream-RSI improves both discovery effectiveness and efficiency in several settings.
Odyssey-3 is a foundation world model with an autoregressive diffusion transformer trained to simulate highly diverse scenarios. The model is capable of controlling robots, powering humanoids, driving vehicles, piloting drones, training AIs, and playing video games. It has a learned understanding of physics, dynamics, cause-and-effect, human behaviors, and other concepts that make up reality. The model will enable physical agents that can interface natively with physical and virtual systems.
Recursive scientific AI can build executable worlds, populate them with agent swarms, and compress thousands of simulated trajectories into human-usable principles. In metamaterials, this revealed how architecture can program failure pathways, improving resilience through load redistribution and controlled collapse.
Provides an open-source humanoid arm for teleoperation, imitation learning, and physical AI research.
The Artificial Intelligence Underwriting Company (AIUC) hopes to bring AI safety to enterprises and companies building AI models and agents. It has built a third-party audit and certification layer for AI agents designed to give enterprises an independent assessment of how safe their AI agents are. Its tests show where companies can trust their agents and where there are concerns, allowing them to be more informed before making decisions. AIUC uses AI agents to run the tests and AI to analyze the data, but humans verify the final audit.
Meta One is a subscription service that offers enhanced AI and self-expression features across Instagram, Facebook, and WhatsApp. It includes single product plans as well as bundled offerings that combine features for individuals, creators, and businesses, starting at $2.99/month.
Sponsored
Specialized knowledge search. Parallel thinking. Multiple LLMs. See how Databricks improved Genie's accuracy from 32% to over 90% on an internal benchmark of real-world data analysis tasks. See how Genie
OpenAI's clear priority is recursive self-improvement, and AI could surpass even Brown's own research intuition within one or two model releases from now.
ALTK-Evolve introduces Consistency Guidelines to address the inconsistency gap, where a model's task success rate drops significantly across repeated runs.