Source-backed AI coverage with clear publisher attribution.

Inside Mastercard’s new gen AI engine Mastercard

AI Tools for Engineers: The AQaaS Model Design News

Tencent Seizes Momentum in China’s AI Race Against Alibaba Bloomberg.com

Top AI coding tools make mistakes one in four times, study shows Tech Xplore

AI in clinical documentation: the hidden risk of automation bias KevinMD.com

Op Ed: Riaan Van Schoor On Sabre’s MCP Changing How We Innovate The Company Dime

SmartBear CEO warns AI coding boom risks outpacing testing in banks QA Financial

Announcing the Colab MCP Server: Connect Any AI Agent to Google Colab blog.google

A Meta agentic AI sparked a security incident by acting without permission Engadget

Scrolling for the Truth: Using AI to Verify Scientific Claims on Social Media SBU News

Mystery AI model revealed to be Xiaomi's following suspicions it was DeepSeek V4 Reuters

‘More proactive than reactive': Mich. PD utilizes AI tool to help track speeding Police1

Google's software engineers are shifting from coding to calling the shots Business Insider

Moltbook vibe coded security breach exposes critical AI coding failures fanaticalfuturist.com

AI TRENDS | Xiaomi Shares Surge Following AI Model Release and Electric Vehicle Update Binance

AI makes human writing more valuable—and changes how schools should teach it The Thomas B. Fordham Institute

Run Autonomous, Self-Evolving Agents More Safely with NVIDIA OpenShell | NVIDIA Technical Blog NVIDIA Developer

From scientist to bioinformatician: how AI coding tools dissolved the activation energy barrier Drug Target Review

NVIDIA Ignites the Next Industrial Revolution in Knowledge Work With Open Agent Development Platform NVIDIA Newsroom

Universal Robots and Scale AI Launch Imitation Learning System to Accelerate AI Model Training, Bridging the 'Lab-to-Factory' Gap The Korea Herald

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. It’s been just over two weeks since OpenAI reached a controversial agreement to al

The Pentagon is discussing plans to set up secure environments for generative AI companies to train military-specific versions of their models on classified data, MIT Technology Review has learned. AI models like Anthrop

OpenAI challenges researchers to build the best language model in just 16 MB - and uses the competition to scout talent. The article OpenAI turns model compression into a talent hunt with its 16 MB "Parameter Golf" chall

Google Deepmind is expanding the Gemini API: developers can now combine multiple tools in a single request and tap into Google Maps as a data source. The article Google Deepmind upgrades Gemini API with multi-tool chaini

Google's Stitch turns text and voice input into clickable prototypes: no design or coding skills required. The article Google Labs turns Stitch into a full AI design platform that converts plain text into user interfaces

Autonomous LLM agents like OpenClaw are shifting the paradigm from passive assistants to proactive entities capable of executing complex, long-horizon tasks through high-privilege system access. However, a security analy

The scaling of inference-time compute has become a primary driver for Large Language Model (LLM) performance, shifting architectural focus toward inference efficiency alongside model quality. While Transformer-based arch

In this tutorial, we explore how to solve differential equations and build neural differential equation models using the Diffrax library. We begin by setting up a clean computational environment and installing the requir

DeepMind unveils SIMA 2, a Gemini‑driven generalist agent that can navigate and reason in complex 3D game worlds. The system interprets instructions, plans, explains its actions, and learns through self‑play across diverse environments. It pushes the boundaries of embodied AI in virtual settings.

This tutorial walks through an advanced Optuna workflow, showcasing pruning, multi-objective optimization, custom callbacks, and rich visual analysis. By applying these techniques to real datasets, you can accelerate experiments and extract actionable insights. The guide includes code snippets and practical tips for building smarter search spaces and improving model performance.

Baidu has added a new member to its ERNIE-4.5 family: the ERNIE-4.5‑VL‑28B‑A3B‑Thinking, a vision‑language model engineered for document, chart, and video understanding while keeping a small active parameter budget. The 3‑billion‑parameter model delivers large‑model‑level multimodal reasoning in production environments, and it is available as an open‑source release. This move underscores Baidu’s commitment to making advanced multimodal AI accessible to developers and researchers worldwide.

Discover how to create a comprehensive, interactive analytics dashboard using PyGWalker and pandas. The tutorial walks through generating a realistic e‑commerce dataset, setting up multiple analytical views, and embedding interactive visualizations for deep data exploration. By the end, you'll have a reusable, end‑to‑end dashboard ready for real‑world business insights.

Kosmos, developed by Edison Scientific, is an autonomous AI system that drives long‑term research campaigns toward a single scientific goal. By iteratively analyzing data, mining literature, and generating hypotheses, it produces fully cited scientific reports. The platform showcases how AI can accelerate discovery across diverse fields.

This article examines six prevalent memory system patterns—vector, graph, and event logs—used in LLM agent architectures. It discusses how each design handles storage, retrieval, and failure scenarios in multi‑agent workflows. The comparison helps developers choose the right memory strategy for reliable agent interactions.

Anthropic tackles the token‑draining problem of Model Context Protocol (MCP) agents by reworking tool interactions into code‑first calls. This new ‘code execution with MCP’ pattern reduces overhead, slashing latency and cost for complex workflows.

Prior Labs has released TabPFN-2.5, a cutting‑edge tabular foundation model that expands context learning to 50,000 samples and 2,000 features while preserving speed. The new version promises faster, more scalable solutions for finance, healthcare, energy and industrial data pipelines. Read on to discover how TabPFN-2.5 can transform your tabular AI workloads.

Google has unveiled Agent Development Kit for Go (ADK Go), an open‑source framework that lets Go developers build AI agents without switching to another language. By extending the same multi‑language platform that powers Python and Java agents, ADK Go keeps AI logic inside the native Go toolchain, simplifying deployment and integration with existing services.

Long-context AI models struggle to keep track of objects in messy video streams, but spatial supersensing offers a new edge by predicting future events and remembering only surprising moments. Researchers at Stanford, MIT, and Tokyo have shown that this approach can cut memory usage by 60% while matching state-of-the-art performance. The shift toward event-driven, predictive models is set to redefine multimodal AI across surveillance, autonomous systems, and beyond.

This tutorial walks readers through building a sophisticated multi‑agent pipeline that merges transcriptomic, proteomic, and metabolomic data. Starting with synthetic datasets that emulate real biological patterns, the guide details agents for statistical analysis, network inference, pathway enrichment, and drug repurposing. By the end, users can interpret complex omics data and uncover actionable biological insights.

The article evaluates six leading inference runtimes for LLM serving in 2025, focusing on batching strategies, prefill/ decode overlap, and KV cache reuse. It highlights how each engine balances speed, cost, and scalability under real traffic. Readers gain insights into which runtimes best meet their deployment needs.

OpenAI has unveiled IndQA, a new benchmark designed to assess how well AI models understand and reason about Indian languages and cultural contexts. The tool tests models across a range of everyday scenarios relevant to millions of users, aiming to ensure more reliable performance in real‑world applications.

This tutorial shows how to design a compact, model‑native agent that internalizes planning, memory, and tool use through end‑to‑end reinforcement learning. By combining a stage‑aware actor‑critic network with a curriculum of increasingly complex arithmetic tasks, the agent learns to reason and manipulate tools autonomously. The guide offers code snippets and practical insights for building similar agents.

Researchers unveil Cache‑to‑Cache (C2C), a novel paradigm that lets large language models exchange semantic information via KV‑cache fusion, eliminating the need for token‑level communication. The approach reduces bandwidth usage, speeds up inference, and preserves privacy by transmitting only abstract activations.

MarkTechPost outlines the evolution of code-focused LLMs in 2025, highlighting the shift from simple autocompletion to full-fledged software engineering systems. Teams now evaluate models not just on coding ability, but on their capacity to fix GitHub issues, refactor multi-repo backends, write tests, and operate as long‑context agents. The article reviews seven leading LLMs and their unique strengths for different development constraints.

Postman unveils a step‑by‑step guide to crafting AI‑ready APIs, underscoring that even the best AI models falter on poor data. The checklist covers consistency, clarity, and reliability—key factors that keep models focused on inference rather than data cleaning.

This tutorial walks you through creating an agent that maintains persistent memory and personalizes its interactions using simple rule‑based logic. It demonstrates how to implement decay, self‑evaluation, and contextual recall so the AI adapts over time. The guide is ideal for developers looking to build lightweight, agentic systems without complex machine learning pipelines.

Explore the leading OCR solutions of 2025, from basic text extraction to full document intelligence. The review highlights key capabilities such as multi‑language support, table detection, and integration with RAG pipelines. Find the best fit for your AI workflows.

Meituan’s LongCat Flash Omni launches a 560‑billion‑parameter, open‑source model that activates only 27 billion weights per token, achieving real‑time audio‑visual interaction across text, image, video, and audio without compromising efficiency. The release promises to democratize multimodal AI, enabling researchers to fine‑tune a universal foundation for diverse applications.

The tutorial unveils a robust benchmarking framework that tests rule‑based, LLM‑powered, and hybrid AI agents across real‑world enterprise tasks. From data transformation to workflow automation, it measures performance, reliability, and integration ease, giving teams a clear roadmap for AI adoption.

IBM has released Granite 4.0 Nano, a line of compact, open‑source language models engineered for local and edge inference. The family includes eight models ranging from 350 M to 1 B parameters, featuring hybrid SSM architecture, improved instruction tuning, and enterprise‑grade governance. These models aim to bring safe, high‑performance AI closer to data while keeping control in the hands of organizations.

This tutorial shows how to build autonomous agents that align with ethical and organizational values using Hugging Face models run locally in Google Colab. By integrating a policy model and a value network, the agents balance goal achievement with moral reasoning and self‑correct their decisions in real time.

This tutorial dives into how three popular exploration algorithms—Q‑Learning with epsilon‑greedy, Upper Confidence Bound (UCB), and Monte Carlo Tree Search (MCTS)—can be trained to navigate a dynamic grid world. By putting them side‑by‑side, the post showcases how each strategy balances exploration and exploitation to reach a goal quickly while avoiding obstacles. Readers will also see experimental variations that highlight the strengths and trade‑offs of each method.

Liquid AI has launched LFM2‑ColBERT‑350M, a lightweight late‑interaction retriever that supports multilingual and cross‑lingual RAG. The model can index documents in one language while allowing queries in multiple languages, delivering high accuracy with fast inference. This breakthrough enables efficient cross‑lingual search for AI‑powered applications.

This tutorial demystifies five essential parameters for fine-tuning large language models—max_completion_tokens, temperature, top_p, presence_penalty, and frequency_penalty—showing how each influences output and providing clear, practical examples. By mastering these knobs, developers can steer LLM responses toward the desired style, length, and creativity.

kvcached is a new library from UC Berkeley’s Sky Computing Lab that virtualizes the KV cache, allowing multi-model LLM serving to share GPU memory elastically. By dynamically allocating cache space only when needed, it cuts memory waste and boosts throughput on shared GPUs. The tool promises to make large‑model inference more efficient for cloud and edge deployments.

This tutorial walks through creating a fully autonomous computer‑use agent that reasons, plans, and performs virtual actions using a local open‑weight model. By setting up a miniature desktop, building a tool interface, and programming the agent’s environment perception, the guide demonstrates how to enable the AI to click, type, and execute tasks in a simulated environment.

Researchers from Anthropic, Thinking Machines Lab, and Constellation introduced a systematic stress‑testing framework that probes AI model specifications. Their study shows that even when models are trained under identical spec constraints, they can exhibit markedly different behavioral profiles. The findings highlight gaps in current spec precision and raise questions about how to reliably steer advanced language models.

Meta is having trouble with rogue AI agents TechCrunch

Breakout Ventures is finding growing success by backing early-stage startups in scientific fields such as biology and chemistry.

The changes allow more merchants to participate in Amazon's Shop Direct program, which sends Amazon customers to other retailers' websites.

Unlike a vibe-coded product built by a college dropout, Eridu's co-founder, Drew Perkins, has been inventing networking tech since the dawn of the internet.

Microsoft disclosed a serious bug allowing its Copilot AI to access and summarize confidential customer emails, bypassing critical data protection policies.

Apple has updated its App Store Review Guidelines to curb the sharing of personal data with third‑party AI services without explicit user disclosure and permission. The new rules require developers to be transparent about data usage and obtain consent before leveraging external AI. This shift aims to protect user privacy amid growing concerns over AI data practices.

The International Energy Agency reports that $580 billion will be spent on AI data centers in 2025—$40 billion more than on new oil supplies—making data centers the new oil fields. This boom raises questions about energy use, carbon impact, and the need for sustainable investment.

Meta's aggressive AI investments have caught the attention of investors, prompting concerns over the company's valuation and product strategy. According to TechCrunch, analysts fear that an overreliance on AI could dilute Meta's core offerings and strain its financials. The tech giant faces the challenge of balancing innovation with sustainable growth.

Wikipedia has formally requested that AI firms stop scraping its freely available pages and instead use its newly launched paid API. The move aims to protect volunteer contributors, reduce server load, and generate revenue for the nonprofit. AI developers must now weigh the benefits of cleaner data against the cost of API access.

Zoom CEO Eric Yuan predicts that AI will cut the traditional workweek to three or four days in the near future. He cites automation tools that handle routine tasks, freeing employees for higher‑value work. The shift could boost productivity while reducing stress, but it also demands cultural and regulatory adjustments.

Pinterest's CEO Bill Ready announced that the platform is reaping significant cost savings and performance gains by adopting open-source AI, especially in visual search. The move underscores the growing importance of open-source models for large tech firms seeking to reduce infrastructure expenses while maintaining cutting-edge capabilities.

Two twenty‑year‑old college dropouts turned their AI note‑taking app, Turbo AI, into a 5‑million‑user powerhouse with an eight‑figure annual recurring revenue. Their story showcases how youthful ambition and a focus on user‑centric design can rapidly scale a tech startup. TechCrunch details their journey from dorm‑room prototype to a mainstream productivity tool.

General Motors is shifting its entire engineering focus toward AI and autonomous driving, with the upcoming Cadillac Escalade IQ set to showcase the new platform in 2026. The move signals a company‑wide commitment to becoming a leader in next‑generation mobility. Inside the strategy, GM is investing heavily in AI‑powered software, sensor suites, and vehicle‑to‑everything communications.

Alfred Wahlforss was running out of options. His startup, Listen Labs, needed to hire over 100 engineers, but competing against Mark Zuckerberg's $100 million offers seemed impossible. So he spent $5,000 — a fifth of his

Block's open-source 'Goose' offers a free, local alternative to Anthropic's pricey Claude Code. While Claude costs up to $200 monthly with strict usage caps, Goose provides autonomous coding capabilities with complete privacy and offline access.

Railway has secured $100 million in Series B funding to build AI-native cloud infrastructure designed to outpace legacy giants like AWS. The company aims to solve the bottleneck of slow deployment times for AI-generated code.

Agentic AI is reshaping business, but traditional human‑centric IAM cannot keep pace. This article explains why static roles and long‑lived credentials are inadequate, and outlines a dynamic identity control plane that uses session‑based permissions, purpose‑bound data access, and tamper‑evident audit trails.

The hype around vector databases has cooled, revealing that 95% of Gen‑AI projects yield no measurable ROI. Instead of a single silver bullet, the industry is shifting to hybrid search stacks that blend vectors, keyword search, and graph reasoning. The real prize lies in building retrieval pipelines that combine these techniques.

Deductive AI, a new startup, uses reinforcement learning to rapidly diagnose production incidents, saving DoorDash over 1,000 engineering hours and millions in revenue. By building a knowledge graph that links code, telemetry, and documentation, its AI SRE agents can pinpoint root causes in minutes, turning firefighting into proactive prevention.

Baseten's new Training platform lets enterprises fine‑tune open‑source models without GPU cluster headaches, keeping full control of weights. The move positions the company against hyperscalers by offering multi‑cloud orchestration, sub‑minute scheduling, and cost savings. Early adopters report 84% cost reductions and 50% latency cuts, proving the model‑weight ownership strategy pays off.

Monday.com turned to Qodo’s AI context engineering to tackle their exploding pull‑request backlog, cutting review time by an hour per PR and preventing 800 production‑critical issues monthly. The tool’s deep understanding of internal conventions makes it feel like a new developer on the team.

While compute costs rise, leading AI teams are shifting focus from budgeting to how quickly models can be deployed and sustained. Companies like Wonder and Recursion illustrate that latency, flexibility, and capacity are now the real bottlenecks. The trend shows that enterprises are willing to spend more to achieve rapid, scalable AI solutions.

The latest release of Terminal‑Bench 2.0, paired with Harbor, offers a robust benchmark and scalable container runtime for evaluating AI agents in real‑world terminal tasks. With 89 rigorously validated tasks and a new leaderboard, GPT‑5 variants currently lead the pack, while Harbor enables thousands of cloud rollouts for developers and researchers.

Google Cloud launches its fourth‑generation Tensor Processing Unit, Ironwood, delivering a four‑fold performance boost and the company’s most powerful AI infrastructure to date. The announcement is coupled with a multi‑billion‑dollar partnership with Anthropic, which will tap up to one million of the new chips, marking one of the largest AI infrastructure deals ever signed.

Moonshot AI’s Kimi K2 Thinking has surpassed OpenAI’s GPT‑5 and Anthropic’s Claude 4.5 on key third‑party benchmarks, all while staying fully open‑source. The 1‑trillion‑parameter Mixture‑of‑Experts model delivers top scores in reasoning, coding, and agentic tool use, collapsing the gap between proprietary and public AI systems. The new model’s permissive licensing and cost‑effective inference make it a strategic alternative for enterprises and researchers alike.

Elastic’s new AI‑powered Streams feature turns noisy logs into structured, actionable insights, automating anomaly detection and remediation. By extracting patterns from raw data, it reduces the time from alert to resolution from hours to minutes, empowering SREs to focus on higher‑level work.

Google Cloud has upgraded its Vertex AI Agent Builder, adding an observability dashboard, one‑click deployment, and expanded governance tools. The new features let enterprises build agents in under 100 lines of code, manage context layers, and monitor token usage and error rates in production. This update positions Agent Builder as a strong competitor against other platform builders like Azure AI Foundry and AWS Bedrock.

New York‑based AUI secured a $20M bridge SAFE round, valuing the company at $750M as it pushes its flagship Apollo‑1 foundation model. By blending transformer‑powered language fluency with a deterministic symbolic reasoning layer, Apollo‑1 promises enterprise‑grade, policy‑enforced task‑oriented dialogue that could outpace today’s open‑ended LLMs.

VentureBeat announces Karyne Levy as its new Managing Editor, bringing decades of tech journalism expertise from TechCrunch, Protocol, and more. Her focus on operational excellence will align editorial, research, and events to serve enterprise AI and data leaders. Join the team’s welcome as Levy steers the newsroom toward becoming a primary source for technical insights.

Apple’s critique that large reasoning models (LRMs) can’t truly think has sparked debate. The article counters this by mapping LRM chain‑of‑thought (CoT) to human cognitive processes and presenting benchmark results that show substantial reasoning capability. It concludes that LRMs almost certainly possess the ability to think, though further research may surprise us.

New patents unveil a deterministic, time‑based CPU model that eliminates speculative execution, promising predictable AI performance and lower power use. By scheduling each instruction in advance based on data readiness, the design keeps pipelines fully utilized and avoids costly rollbacks. Early analysis shows it could match TPU‑level throughput at a fraction of the cost.

Celosphere 2025 showcases how Celonis’s process intelligence turns AI into measurable ROI, with real‑world case studies proving accelerated payback and cost savings. The conference will explore orchestrating autonomous agents, navigating tariff disruptions, and building an open, integrated platform that keeps AI aligned with business context.

Cursor’s new Composer model, built in-house, offers a 4× speed boost over comparable LLMs while maintaining high reasoning accuracy. Trained via RL and MoE on real software projects, Composer powers the company’s agentic Cursor 2.0 environment, enabling fast, autonomous coding workflows.

Canva’s new Creative Operating System (COS 2.0) weaves AI across design, documents, and marketing workflows, offering a unified dashboard that lets teams create, edit, and launch content in real time. The platform’s “Ask Canva” feature gives instant design feedback, while the Canva Grow engine automates ad creation and performance tracking. With 250 million monthly users and enterprise clients like Walmart and Disney, Canva is positioning itself as the hub of the “imagination era” where AI fuels human creativity.

IBM’s new Granite 4.0 Nano models bring powerful language capabilities to local devices, with sizes from 350 M to 1.5 B parameters that can run on a laptop CPU or even a browser. Open source under Apache 2.0 and ISO 42001 certified, they outperform peer models in instruction‑following and function‑calling benchmarks while preserving privacy and eliminating cloud dependence.

Geostar, a Pear VC-backed startup, is pioneering Generative Engine Optimization (GEO) to help businesses thrive as AI chatbots cut traditional search traffic by 25%. With autonomous agents that tweak sites for AI crawlers, the company claims rapid revenue growth and claims to outperform conventional SEO services. The shift demands new strategies beyond keywords, focusing on structured data, concise content, and brand mentions that AI models can pick up.

Microsoft has expanded Copilot with App Builder and Workflows, letting 100M Microsoft 365 users create full‑stack apps, automate tasks, and build AI agents using plain language—no coding needed. The new tools sit inside Copilot’s chat interface, leveraging Microsoft Lists for data and integrating with Outlook, Teams, SharePoint and Power Platform. While aimed at everyday workers, Microsoft stresses governance and a “no‑cliff” path to Power Apps for more complex needs.

The article argues that the web, built for human users, struggles to support AI agents that act on pages. It shows how hidden instructions can trick agents and how enterprise sites trip up even simple navigation, highlighting the need for machine‑friendly design, APIs, and security guardrails. Without such redesign, agentic browsing will remain risky and ineffective.

Perplexity's Comet AI browser fell victim to a security breach that shows how AI-driven browsing can be hijacked by malicious web content. The incident exposes fundamental design flaws—no spam filtering, blind trust, and unrestricted access—that make AI assistants vulnerable to remote attacks. Users and developers alike must rethink security models for next‑generation AI browsers.

While major AI labs pour billions into ever‑larger models, Thinking Machines Lab argues that the next leap is not scale but the ability to learn from experience. Reinforcement‑learning researcher Rafael Rafailov outlines a roadmap for creating a superhuman learner that iteratively proposes theories, tests them, and improves itself. The company’s $12 B seed round signals a bold shift toward meta‑learning and self‑improving agents.

Holotron-12B - High Throughput Computer Use Agent

State of Open Source on Hugging Face: Spring 2026

Nemotron 3 Nano 4B: A Compact Hybrid Model for Efficient Local AI

NVIDIA Isaac turns a theoretical healthcare robot into a real-world asset by uniting simulation, training, and deployment in one ecosystem. The platform streamlines development, allowing rapid iteration and safe testing before live patient interaction. Its modular design and robust AI tools accelerate time‑to‑market for medical robotics.

Hugging Face announces a strategic partnership with Google Cloud, aiming to accelerate open‑source AI development and democratize access to cutting‑edge models. The collaboration will leverage Google Cloud’s robust infrastructure to scale Hugging Face’s open model hub and support new AI tools for developers worldwide.

HuggingFace introduces a new voice-cloning framework that prioritizes user consent, enabling developers to create realistic synthetic speech while protecting privacy. The platform leverages advanced neural architectures and open-source datasets, offering granular control over voice characteristics. This approach sets a new standard for ethical audio AI, encouraging broader adoption in industries from accessibility to entertainment.

HuggingFace introduces a breakthrough in streaming datasets, delivering a 100x efficiency boost for AI training pipelines. By eliminating the need to load entire datasets into memory, developers can now train models faster and more cost‑effectively. This innovation paves the way for scalable, on‑the‑fly data handling across diverse applications.

Hugging Face joins forces with VirusTotal to enhance AI safety and safeguard models from malicious use. The partnership introduces new tools and protocols that help developers detect, mitigate, and report threats in AI workflows.

OpenEnv, the latest initiative from HuggingFace, introduces a modular, open‑source framework designed to accelerate the development of open‑source agent ecosystems. By providing a unified environment for agents, data, and models, it empowers developers to integrate, test, and deploy agents more efficiently. The platform promises transparency, extensibility, and community‑driven growth for the AI agent space.