Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sponsored
Agent Bricks [12] helps you build, evaluate and optimize AI agents grounded in your unique data. It evaluates automatically, scores outputs against your goals and improves with human feedback — giving you a clearer path to production. Build agents that work in the real world.
The Universal Commerce Protocol (UCP) is an open-source standard that enables seamless commerce journeys between customer surfaces, businesses, and payment providers. It works with existing retail infrastructure and is compatible with the Agents Payments Protocol. The UCP provides businesses with flexible ways to integrate via APIs, Agent2Agent, and the Model Context Protocol. It was developed by Google in collaboration with industry leaders.
We are getting closer to long-running autonomous agents. However, models get worse as context grows. Effective agent design largely boils down to context management. This post explores common design patterns across agents.
Anthropic's practical approaches to agent evaluation emphasize pre-deployment tests that simulate real-world conditions and reduce failures in complex, multi-turn agent systems.
Meta AI researchers developed DeepThink with Confidence (DeepConf), a technique that uses internal confidence signals to cut LLM reasoning overhead by up to 84.7% while maintaining accuracy. The method monitors token-level confidence to terminate low-quality reasoning traces early, particularly when models generate uncertainty markers like "wait" or "think again."
Agent-native architectures replace step-by-step code with agents that decide how to achieve outcomes, while developers define only the desired results through prompts. This model makes software faster to build and adapt, since behavior changes come from modifying language instead of rewriting logic. The shift trades predictability for flexibility, pushing software design toward continuous pruning and observation rather than rigid control.
In AI agents, the code only orchestrates models and tools, while real decision-making happens inside the model at runtime and cannot be understood by reading the code alone. Traces capture the actual reasoning steps, tool calls, errors, costs, and outcomes, making them the primary artifact for debugging, testing, optimization, and monitoring.
Anthropic has cut access to xAI developers who were using its Claude models for coding via the Cursor AI coding tool.
Indonesia and Malaysia have temporarily blocked access to xAI's chatbot Grok after it generated non-consensual deepfakes, including content depicting minors and violence.
Experts predict significant economic and technological impacts.
Demand for AI compute is outstripping supply due to soaring token consumption and a surge in LLM users.
Deep Future is an AI stress-testing tool for strategies that simulates thousands of futures using a scenario generation engine.
Sponsored
If AI code reviewers are supposed to save time, why am I triaging 40 comments about trailing whitespace, "consider using an f-string," and "line exceeds 100 characters" on a 10-line change?
Gmail is adding AI Overviews using Gemini technology, summarizing email threads and allowing users to query their inbox with natural language. Its new features, Help Me Write, Suggested Replies, and Proofread, improve email drafting, while AI Inbox filters prioritize key messages. These updates, available for free, are rolling out now in the US.
OpenAI is acquiring the team behind Convogo, a business software platform that automates and improves leadership assessments and feedback reporting. The startup will not acquire Convogo's IP or technology. Convogo's team will join OpenAI to work on its AI cloud efforts. Its product will be wound down.
The poster for 'Lord of War', which depicts an image of Nicholas Cage comprised of bullets, is a masterclass in compositing. A team of designers likely had to manually source images of bullets, match the lighting of the individual bullets with the shadows on the face, and arrange thousands of shell casings to form a recognizable human face. That level of skill was rare in 2005. This article explores whether AI can redesign a similar 'Lord of Tokens' image with GPU dies and silicon wafers constituting Jensen Huang's face.
This post demonstrates how to implement long-term memory and stateful personalization in AI agents using OpenAI's Agents SDK. Structured memory objects and context-injection patterns enable agents to recall preferences, consolidate user notes, and evolve over time.
Increased efficiency in resource use can lead to higher total consumption, but only when demand is elastic. Cheaper production just changes who captures the surplus when demand is constrained by something other than production cost. AI coding will change how things are built without changing how much gets built or at what quality for a huge amount of economically significant software. This post looks at eight markets to see what impact AI coding will have.
Sponsored
Why is AI stuck in pilot mode? It's constrained by data infrastructure and enterprise context. CData surveyed 200+ data & AI experts [10] on what's working (and breaking) in AI data connectivity. The key finding? IF YOUR DATA ISN'T LIVE, CONTEXTUAL, AND GOVERNED, YOUR AI WON'T DELIVER. Read the 2026 State of AI Data
Warp specialization is a popular technique that improves kernel performance on GPUs out of the box. The key idea is to have specialized code paths for each warp instead of the same code. This reduces performance hits due to control flow divergence, improves latency hiding, and makes better use of hardware units on the GPU. This post outlines the current design of warp specialization in Triton, a compiler that aims to generate performance-portable code and runtime across hardware for AI kernels.
This guide walks readers through the process of building a code review agent from scratch that can read files, run commands, edit code, and figure out the steps to accomplish a task. The resulting agent can analyze a code base, find bugs and security issues, and return structured feedback. The guide aims to help readers understand how the Claude Code SDK works so they can build whatever they actually need.
Agentic Rubrics enable test-free verification of code changes by creating checklist-style evaluations from expert agents that interact with codebases. They outperform baseline methods on SWE-Bench Verified and offer a scalable, interpretable signal for reinforcement learning and inference-time feedback.
Modern AI systems have evolved past simply predicting the next word and now demonstrate enhanced capabilities, such as solving complex problems and reasoning. Systems like OpenAI and Google DeepMind have achieved impressive feats, including outperforming top mathematicians in competitions, reflecting significant advancements in cognitive tasks beyond autocomplete.
Coding agents like Gemini 3 Pro and Claude Opus 4.5 have automated tasks ranging from invoice creation to predicting US corn yields. These agents exhibit capabilities close to artificial general intelligence but require human understanding and oversight to avoid errors and maximize their potential. The rise of these tools signals a shift towards enabling individuals to independently explore complex software projects. They demonstrate the importance of foundational technical knowledge and creating opportunities and challenges for both new and incumbent firms.
Sponsored
Developers are shipping more code - with PRs getting bigger and denser. Read Greptile's cross-industry study [17] on the state of AI coding to see AI tool adoption, model growth trends, and model snapshots.
This GitHub gist contains a Claude Code hook that auto-approves read-only MCP tools to eliminate constant permission dialogs.
Third-party apps can act like a VPN for AI services, as they prevent service providers from being able to see the exact identity of users.
Ridding ourselves of all boredom could result in the loss of creativity.
Nvidia now requires Chinese customers to fully prepay for its H200 AI chips without refunds due to regulatory uncertainties from the US and China.
OpenAI has introduced ChatGPT Health, a new feature designed to assist users with health and wellness by securely integrating their medical information into a personalized ChatGPT experience.
OpenAI set aside an employee stock grant pool equivalent to 10% of the company last fall, when it was valued at $500 billion. The startup has already given $80 billion in vested equity. It is in preliminary talks with some investors about raising funds at a valuation of around $750 billion.
Anthropic plans to raise $10 billion at a valuation of $350 billion. The funding round is expected to close in the coming weeks, and the total amount of the deal could change. AI companies collected a record $222 billion in funding last year. This year is slated to be a big year for public offerings.
Chinese companies are making strong, open AI models that are applying increasing pressure on the US AI economy. Dethroning Qwen in adoption in 2026 looks impossible, but there are areas of opportunity. The US could very well have the smartest open models again in 2026, even if they're used far less across the ecosystem. This post contains 8 plots that show the impact of Qwen, DeepSeek, Llama, GPT-OSS, Nemotron, and other new entrants to the ecosystem.
Context graphs shed light on how work really gets done in the enterprise. AI agents are a major unlock with automation, but only if their reasoning is grounded in the right enterprise context. It's not enough to just understand the enterprise data, systems also need relationship knowledge. Context graphs enable agents to learn and automate distributed work by capturing processes and learning intent over time.
Recursive Language Model (RLM) is a flexible solution to long-context limitations in LLM agents. RLMs avoid summarization by delegating memory management to scripts and sub-models.
Sponsored
Most companies aren't getting the AI results they want. This Forrester study [11] shows how an overemphasis on individual productivity, instead of collaboration, could be the culprit. For a look at how collaboration-first AI lives up to the hype, join the Miro
PDFs have been largely ignored by the open source AI community, despite being a source of high-quality data for pretraining. Hugging Face has created a new pretraining dataset of more than 3T tokens spanning over 1,000 languages extracted from PDFs. This post details how the dataset was created.
NitroGen is a generalist gaming agent trained on internet gameplay videos capable of predicting controller actions from pixels and adapting to unseen games through post-training.
WebGym is a large-scale environment with nearly 300,000 real-world web tasks for training visual agents. It has a high-throughput RL system that boosts training speed and a fine-tuned Qwen-3-VL that outperforms GPT-4o and GPT-5-Thinking on unseen web navigation tasks.
Screenless and always-on AI devices will fail, while mini-apps enable UGC personal software and a US consumer super-app. By 2030, consumer AI will split into a reliable task assistant and an AI companion, performance marketing will collapse, and AI-generated virtual creators will explode. The biggest win will go to whoever solves AI discovery beyond chat and search.
Brendan Foody, the youngest unicorn founder and CEO of Mercor, is transforming AI training by hiring experts like poets and economists to create evaluation frameworks for frontier models. Mercor pays top talent to teach AI models that are rapidly improving in economically valuable tasks.
Sponsored
Making AI agents doesn't have to drag on. Read this eBook from Algolia to see how you can deploy branded conversational AI agents
Google now lets educators create podcast-style audio lessons using Gemini in Google Classroom.
The breakneck speed at which AI tools are being integrated into existing systems and workflows without proper threat modeling, testing, and security analysis is dangerous.
AI agents and MCP servers enhance threat modeling by automating the synthesis of organizational context, enabling security teams to map threats, identify detection gaps, and target high-priority threats efficiently.