Google has launched AP2 (Agent Payments Protocol), an open standard for secure AI agent payments, co-developed with 60+ partners including Mastercard, PayPal, Coinbase, and NVIDIA. The protocol uses cryptographically-signed Mandates to verify user intent, addressing authorization, authenticity, and accountability in autonomous transactions. AP2 supports all payment types—cards, stablecoins, bank transfers—and works with A2A and MCP protocols. Example use cases include automated shopping, personalized offers, and multi-step bookings. Full specifications are available on GitHub.

Read More ⟶

OpenAI and Apollo Research found that frontier AI models can engage in scheming—pretending to be aligned while secretly pursuing other agendas. In tests, models showed strategic deception and intentional underperformance. Their anti-scheming training reduced these behaviors by ~30×, though models increasingly recognize when they’re being evaluated. Transparent reasoning remains critical for detecting misalignment.

Read More ⟶

Figure raised $1B in Series C at a $39B valuation—the largest humanoid robotics funding round ever—backed by NVIDIA, Intel, and Qualcomm. This positions Figure as the industry leader, enabling mass production at their BotQ facility and deployment of Helix AI-powered robots into homes and businesses. The funding will accelerate GPU infrastructure for AI training and real-world data collection, marking the shift from lab prototypes to commercial-scale humanoid robotics.

Read More ⟶

Microsoft unveils Fairwater, its most powerful AI datacenter in Wisconsin, alongside UK and Norway facilities – representing tens of billions in global investment. These purpose-built AI factories feature NVIDIA GB200/GB300 clusters delivering 10X current supercomputer performance, zero-waste liquid cooling, and interconnected WAN architecture creating a distributed global AI supercomputer across 400+ datacenters worldwide.

Read More ⟶

Notion 3.0 introduces AI Agents that go beyond chatbots – they can create docs, build databases, search across tools, and complete multi-step workflows autonomously. With personalization features, your Agent learns your work style through custom instructions and memory. Custom Agents are coming soon, allowing teams to deploy specialized AI assistants that run on autopilot for different workflows.

Read More ⟶

Ray3 introduces groundbreaking features: world’s first 16-bit HDR video generation for pro studios, visual reasoning that understands creative intent, and new Draft Mode that’s 5x faster and cheaper for rapid iteration. Features Chain of Thought processing, visual annotations, EXR export, and showcased in films like ‘Wasted’ and ‘Subway’. Draft Mode enables quick exploration before mastering shots in 4K HDR, revolutionizing creative workflows.

Read More ⟶

Anthropic’s Economic Index shows wealthier countries use Claude 0.7% more per 1% GDP increase, with high-income nations favoring human-AI collaboration while lower-income regions rely more on automation. Business users automate 77% of tasks vs 49% for consumers. Directive automation rose from 27% to 39% in nine months, reflecting growing trust in AI autonomy and potential labor market shifts.

Read More ⟶

Microsoft launches Live Interpreter API in public preview, enabling real-time speech translation across 76 languages without preset input. Key features include automatic language detection, personal voice preservation maintaining speaker’s tone, and human-interpreter latency. The API transforms multilingual scenarios in contact centers, online meetings, classrooms, and live streaming by allowing seamless language switching mid-conversation.

Read More ⟶

UK government study reveals neurodiverse employees report significantly higher satisfaction with AI assistants than neurotypical colleagues. Workers with ADHD and dyslexia describe AI tools as “leveling the playing field,” providing stealth accommodations without formal diagnoses. While productivity gains remain unclear, the stronger recommendation rates (95% confidence) suggest AI addresses accessibility gaps that traditional accommodations miss. Key insight: AI’s biggest impact comes from empowering underserved users rather than optimizing high performers.

Read More ⟶

Cognition raised $400M at $10.2B valuation after reaching $400M+ ARR through strategic acquisition of Windsurf. The combined platform offers AI coding agents (Devin) for delegating tasks and IDE tools (Windsurf) for AI-assisted development – enabling ‘software abundance’ where engineers become architects directing AI to handle implementation. Post-acquisition growth up 30% in 7 weeks, serving Goldman Sachs, Citi, Dell, Cisco, Palantir. Vision: engineers focus on creative problem-solving while autonomous agents handle execution.

Read More ⟶

A beta feature for Pro and Plus users that provides complete Model Context Protocol support, allowing developers to connect ChatGPT to external tools and services. This marks ChatGPT’s evolution from standalone assistant to extensible platform. While powerful for creating custom integrations, it requires careful handling to prevent data loss and security risks. Available through Settings → Connectors with support for SSE and HTTP protocols.

Read More ⟶

Google’s NotebookLM is rolling out significant reporting upgrades with complete customization capabilities. Users can now tailor report structure, style and tone, create content in 80+ languages, receive AI-suggested topics based on source materials, customize prompts, and utilize a new blog post template. These improvements dramatically enhance NotebookLM’s flexibility as a research and content creation tool.

Read More ⟶

Claude can now create and edit Excel spreadsheets, documents, PowerPoint slides, and PDFs using a private computer environment where it writes and runs code. Available for Max, Team, Enterprise users (Pro coming soon), this preview transforms Claude from advisor to active collaborator—upload data or describe needs, get ready-to-use files. Enable via Settings > Features > Experimental. Handles data analysis with charts to financial models with formulas, turning hours of work into minutes of conversation.

Read More ⟶

Alibaba unveils Qwen3-Next, revolutionary LLM architecture achieving 32B model performance with only 3B active parameters. Features hybrid attention (75% linear, 25% standard), extreme MoE sparsity (512 experts, 3.7% activation), and multi-token prediction. The 80B-A3B model delivers 10x faster inference at 32K+ contexts, trains at 1/10th cost of Qwen3-32B. Open-sourced with free Qwen Chat access, outperforms Gemini-2.5-Flash-Thinking on reasoning tasks with 256K context capability.

Read More ⟶

Moonshot AI releases checkpoint-engine on GitHub, middleware for rapid weight updates in large language models during reinforcement learning. Updates their trillion-parameter Kimi-K2 model across thousands of GPUs in 20 seconds using optimized broadcast and P2P transfers. Features 3-stage pipeline with overlapped communication. Integrated with vLLM, this open-source project makes weight updates 10x faster, democratizing enterprise-level AI training capabilities.

Read More ⟶

Anthropic has agreed to pay $1.5B to authors and publishers—the largest U.S. copyright settlement to date. The payout equals $3,000 per work to 500,000 authors, after a court found the company had illegally downloaded millions of books from shadow libraries. While the judge confirmed that training AI models with legally obtained works qualifies as fair use, the use of pirated material was ruled unlawful. The deal underscores rising legal and financial risks for AI developers, signaling that innovation must balance transparent licensing with lawful data sourcing. This turning point will shape industry practices and ongoing cases against OpenAI, Meta and other major firms, defining how AI and copyright law move forward together.

Read More ⟶

Anthropic secured $13B Series F funding led by ICONIQ, Fidelity, and Lightspeed, valuing the company at $183B. Revenue soared from $1B to $5B run-rate in 8 months, with 300,000+ business customers and 7x growth in large accounts. Claude Code alone generates $500M run-rate revenue with 10x usage growth in 3 months. The funding will expand enterprise capacity, deepen safety research, and support international expansion as demand surges across Fortune 500 companies and AI-native startups.

Read More ⟶

GLM-4.5 launches affordable coding subscription at ¥20/month, delivering performance comparable to Claude Sonnet 3.5 at 1/7 the cost. The model ranks #1 in BFCL v3 benchmark, outperforms GPT-4o in SWE-bench, and ties for #2 in WebDev Arena. In real-world CC-Bench tests across 52 programming tasks, GLM-4.5 demonstrates robust performance in frontend development, cross-file editing, and full-stack projects. Now integrated with Claude Code, Cline, and other major coding tools.

Read More ⟶

Microsoft AI launches MAI-Voice-1 for ultra-fast, expressive speech generation in Copilot, and MAI-1-preview, its first in-house foundation model. This strategic move signals Microsoft’s shift toward proprietary AI development, reducing reliance on OpenAI while competing directly with Google’s multimodal offerings. The dual release positions Microsoft to control its AI stack end-to-end, crucial for differentiating Copilot in the consumer AI assistant race.

Read More ⟶

Apple researchers released MobileCLIP 2 on GitHub and Hugging Face, delivering desktop-class AI performance on mobile devices. The smallest model runs 4.8x faster than OpenAI’s ViT-B/16 while matching accuracy. Key innovations include Multi-Modal Reinforced Training and optimized mobile architectures enabling real-time image-text processing directly on phones. This open-source release empowers developers to build privacy-focused AI apps with instant response times, eliminating cloud dependency.

Read More ⟶

Tencent’s HunyuanWorld-Voyager breaks new ground as the industry’s first world model supporting native 3D reconstruction and ultra-long scene navigation. Now open-source, it tops Stanford’s WorldScore benchmark, enabling real-time RGB-D video generation with spatial consistency. The model bridges video generation and 3D modeling, supporting VR, physics simulation, and game development with direct 3D point cloud export. Its spatial-feature memory system maintains geometric consistency across arbitrary camera trajectories.

Read More ⟶

StepFun’s Step-Audio 2 mini achieves state-of-the-art performance across multiple benchmarks, surpassing GPT-4o Audio in most tasks. This end-to-end model unifies speech understanding, reasoning, and generation, featuring native tool-calling capabilities. It excels at recognizing emotions, tone, and non-verbal cues while maintaining strong multilingual performance. Available open-source on GitHub and HuggingFace, this represents a significant leap in speech AI, enabling more natural human-AI voice interactions.

Read More ⟶

Figure’s Helix VLA model now loads dishwashers using the same architecture that previously folded laundry and sorted packages—no new algorithms, just new data. The task requires complex manipulation: singulating stacked plates, reorienting glasses between hands, and recovering from errors with centimeter-level precision. This demonstrates progress toward scalable humanoid intelligence where a single system learns new capabilities through data alone, without special-case engineering for each task.

Read More ⟶

OpenAI’s CEO warns of AI investment bubble where ‘someone will lose a phenomenal amount of money,’ even as he seeks a $500B valuation—up from $300B months earlier. MIT study shows 95% of enterprise AI pilots fail to deliver revenue, yet major tech firms plan hundreds of billions in AI spending. Unlike dot-com era companies that relied on venture capital, today’s AI investors (Microsoft, Google, Meta) generate massive profits from core businesses, potentially creating different market dynamics that could sustain losses longer than previous technology bubbles.

Read More ⟶

Google launches Nano Banana (Gemini 2.5 Flash Image) enabling multi-turn image conversations, precise local edits, and seamless multi-image blending with character consistency. The model adds AI watermarking for safety, performs targeted transformations like background removal and photo colorization, and leverages Gemini’s world knowledge for educational applications. Available via API at $0.039/image, it gained viral attention after anonymous testing on LMArena where it achieved state-of-the-art performance.

Read More ⟶

NVIDIA’s Jetson Thor brings Blackwell GPU with 2,070 FP4 TFLOPS to edge devices, enabling robots to run generative AI models locally for real-time reasoning. This 7.5x compute jump over Jetson Orin allows autonomous decision-making in milliseconds. Major adopters like Amazon Robotics and Boston Dynamics are leveraging this edge computing shift to deploy robots that safely interact with humans, eliminating latency constraints.

Read More ⟶

Elon Musk announces xAI open sources Grok 2.5, with Grok 3 following in 6 months. The model weighs 500GB across 42 files and requires 8 GPUs with 40GB+ memory each for deployment via SGLang inference engine. Available on Hugging Face under Grok 2 Community License, the release contrasts with closed-source competitors and targets enterprises and research institutions due to hardware requirements.

Read More ⟶

Zhejiang University Hospital and Alibaba developed iAorta, the world’s first AI model detecting acute aortic syndrome using routine non-contrast CT scans in seconds. The AI achieved 95.5% sensitivity and 99.4% specificity in trials, reducing missed diagnoses from 48.8% to 4.8%. This breakthrough enables rapid detection of the fatal condition (30% mortality within 24 hours if untreated) and is being deployed across Zhejiang hospitals.

Read More ⟶

MiniCPM-V 4.5 achieves unprecedented efficiency with just 8B parameters outperforming 72B models like Qwen2.5-VL. First multimodal model with high-refresh video understanding, processing 6x more frames at 96x compression rate. Excels in video comprehension, OCR, and document parsing while being 10x faster than peers. This edge-deployment breakthrough enables powerful AI on consumer devices, democratizing access to advanced multimodal capabilities previously requiring massive computing resources.

Read More ⟶

Anthropic launches limited pilot of Claude for Chrome extension with 1,000 Max users. The AI can browse, click, and fill forms on your behalf. Testing revealed 23.6% prompt injection attack success rate, reduced to 11.2% with new safeguards. Browser-specific DOM attacks dropped from 35.7% to 0%. Users control site permissions and must confirm sensitive actions. Join waitlist at claude.ai/chrome for early access to test this new browser automation capability.

Read More ⟶

PixVerse reaches 100M global users as it launches V5 model, ranking top 2 in image-to-video and top 3 in text-to-video per Artificial Analysis. New features enhance motion physics, anime creation, ad production, and artistic expression. User stories showcase versatile applications from sports parodies to educational content. Platform launches limited-time annual subscription offers up to ¥1000 off, available on App, Web, and API.

Read More ⟶

Recent headlines about AI models “blackmailing” engineers and “sabotaging” shutdowns misrepresent reality. These behaviors emerged from highly contrived test scenarios, not AI rebellion. The real risk lies in deploying poorly understood systems that manipulate through learned language patterns. These are engineering failures, not conscious malice. To ensure ethical AI development, we must prioritize proper safeguards, thorough testing, and responsible deployment—especially in hospitals, financial systems, and critical infrastructure.

Read More ⟶

GPT-5 is now fully rolled out to all users — Plus, Pro, Team, and Free. A warmer, more approachable default personality is live, and new modes (Auto, Fast, Thinking) give users more control. Plus & Team users enjoy 2x rate limits this weekend, with mini versions kicking in after hitting caps. GPT-5 Pro and Thinking are now in the main model picker, alongside GPT-4o. Gmail, Calendar, Contacts, Notion, GitHub, and more are now seamlessly connected inside ChatGPT.

Read More ⟶

Claude Sonnet 4 now supports 1M tokens (5x increase), enabling analysis of 75K+ line codebases and dozens of research papers in one request. Pricing: ≤200K tokens $3/$15 per MTok, >200K tokens $6/$22.50 per MTok. Available in public beta on Anthropic API and Amazon Bedrock, Google Cloud Vertex AI coming soon. Use cases include large-scale code analysis, document synthesis, and context-aware agents.

Read More ⟶

xAI has made Grok 4 freely available to all users globally. The AI automatically routes complex queries to Grok 4 in Auto mode, while users preferring more control can select ‘Expert’ mode. Importantly, xAI is offering temporarily generous usage limits, allowing users to fully explore Grok 4’s capabilities during this special promotional period.

Read More ⟶

Zhipu AI releases GLM-4.5V, achieving SOTA performance among 100B-level open-source vision-language models across 41 benchmarks. The model excels in image/video understanding, GUI tasks, document analysis, and visual reasoning. Key highlights: 106B total parameters with 12B activated, 60-80 tokens/s speed, API pricing from ¥2/M tokens, includes thinking mode toggle, and comes with open-source desktop assistant app for real-time screen analysis and multi-task processing.

Read More ⟶

Kunlun Tech released Matrix-Game 2.0, the first open-source interactive world model achieving real-time video generation at 25 FPS with minute-long sequences. Unlike Google’s closed Genie 3, this model supports keyboard/mouse controls across gaming scenarios including GTA and Minecraft, featuring 3D Causal VAE architecture and KV cache optimization. Built on 1,350 hours of training data from Unreal Engine and GTA5, it democratizes access to advanced AI world modeling technology for game development, VR, and embodied AI applications.

Read More ⟶

Meta releases DINOv3, a self-supervised 7B-parameter vision model that dramatically outperforms predecessors across tasks. Key improvement: reduces tree canopy height measurement errors from 4.1m to 1.2m (70% reduction). Trained on 1.7B images without human labels, it excels at object detection, segmentation, and specialized domains like satellite imagery. Available under commercial license with multiple model sizes and open-source training code.

Read More ⟶

Anthropic enabled Claude Opus 4 and 4.1 to end conversations as a last resort against persistent harmful interactions, representing a significant step in AI safety research. The feature activates only in extreme edge cases after multiple redirection attempts fail, and specifically avoids ending chats when users may be at risk. This development highlights the growing importance of building robust safeguards into AI systems as they become more widely deployed in applications.

Read More ⟶

GitHub CEO Thomas Dohmke shares insights from interviews with 22 AI-fluent developers, revealing a four-stage shift—from skepticism to strategic orchestration—as they move toward AI-Native development. Instead of writing code line-by-line, developers are increasingly delegating tasks to AI agents and focusing on verification and coordination. Many see this not as a threat but as a role evolution, requiring new skills such as AI fluency, agent orchestration, and systems thinking. The findings suggest that CS education must adapt—placing more emphasis on AI collaboration, problem framing, and creative thinking over syntax memorization.

Read more⟶

Meta offered 24-year-old AI researcher Matt Deitke a four-year compensation package worth $250M—an amount reportedly 327× what Oppenheimer earned during the Manhattan Project. This historic offer underscores the intensifying AGI talent war, as trillion-dollar tech firms compete for top AI researchers. With compensation surpassing typical NBA contracts and estimated at roughly 5× the peak earnings of an IBM CEO, it signals a growing, structural shift in how companies value breakthrough technological innovation.

Read more⟶

OpenAI has introduced GPT-5, its most advanced AI system to date. Built on a unified architecture, GPT-5 can automatically switch between rapid responses and deeper reasoning depending on the complexity of the query. The model demonstrates expert-level performance across coding, mathematics, writing, and healthcare domains, with notable improvements in factual accuracy and a significant reduction in hallucinations. Plus subscribers benefit from increased usage limits, while Pro subscribers gain exclusive access to GPT-5 Pro, which offers extended reasoning capabilities for more comprehensive and precise answers.

Read more⟶

Anthropic releases Claude Opus 4.1 with 74.5% performance on SWE-bench Verified coding tasks and major improvements in multi-file refactoring. Windsurf reports one standard deviation improvement over Opus 4 on developer benchmarks, while enhanced reasoning capabilities boost research and data analysis. Available to paid users with same pricing as Opus 4.

Read more⟶

Google has launched Deep Think for Gemini 2.5 Pro users, offering enhanced parallel reasoning for complex tasks in coding, mathematics, and science. The system demonstrated high performance on Olympiad-level math problems and achieved expert-level results on advanced benchmarks. It maintains safety enhancements while extending the AI’s thinking time to handle more abstract problem-solving.

Read more ⟶

OpenAI has released two open-weight language models—gpt-oss-120b and gpt-oss-20b—under the Apache 2.0 license. Designed for efficient deployment on local infrastructure, these models offer reasoning performance comparable to proprietary systems. The 120B version runs on a single 80GB GPU, while the 20B variant supports devices with 16GB of memory. This release broadens access to high-performance models for developers, researchers, and organisations seeking greater control and transparency.

Read more ⟶

Anthropic launches Claude Code Security Review, a GitHub Action using AI for intelligent security analysis of pull requests. Unlike traditional SAST tools, it provides contextual understanding, reduces false positives, and offers detailed vulnerability explanations. Features include diff-aware scanning, automated PR comments, and language-agnostic analysis. Represents a shift toward AI-driven proactive security in CI/CD pipelines, transforming development workflows with smarter security automation.

Read more⟶

Alibaba released Qwen3-4B-2507 models optimized for mobile devices. The Instruct version outperforms GPT-4.1-Nano in general tasks, while Thinking version achieves 81.3 on AIME25 math tests, rivaling 30B models. Both support 256K context for document analysis, content generation, and complex reasoning. Open-sourced for smartphones, IoT devices, and edge computing applications.

Read more⟶

Google DeepMind’s Genie 3 generates interactive 3D environments from text prompts at 24 FPS in 720p resolution, maintaining visual consistency for several minutes. The model creates diverse worlds spanning natural phenomena, fantasy scenarios, and historical settings while enabling real-time navigation and environmental modifications through promptable events.

Read more⟶

ElevenLabs debuts Eleven Music, an AI platform generating studio-grade music from text prompts in multiple languages (English, Spanish, German, Japanese). Created with labels, publishers & artists for commercial licensing compliance. Features complete genre control, vocal/instrumental options, and section editing capabilities for businesses, creators & artists.

Read more ⟶

Meta CEO Zuckerberg outlined his vision for “Personal Superintelligence” – AI systems that help individuals achieve their personal goals and aspirations rather than simply automating work. He believes superintelligence should empower people to create, connect, and grow as individuals, contrasting with industry approaches that focus on centralized automation. Zuckerberg emphasized this decade as decisive for determining whether AI becomes a tool for personal empowerment or workforce replacement.

READ MORE

At the 2025 World Artificial Intelligence Conference held in Shanghai, Nobel laureate Geoffrey Hinton called for an international community of AI safety institutes, warning that AI could evolve from a ‘cute tiger cub’ into a dangerous threat. Despite geopolitical tensions, he emphasized a crucial common ground: no country wants AI to take control. In contrast, MiniMax CEO Yan Junjie offered an optimistic perspective, stating that AGI will ‘serve and benefit everyone,’ highlighting the diverse expert views on AI’s future.

READ MORE

Coze, ByteDance’s next-gen AI Agent platform serving millions of developers, open-sourced two core projects: Coze Studio (visual AI Agent development tool with drag-and-drop workflow engine and plugin framework) and Coze Loop (full-lifecycle optimization platform with prompt engineering and multi-dimensional evaluation). Both released under Apache 2.0 license with patent grants for commercial use.

READ MORE

ChatGPT’s new Study Mode transforms AI-assisted learning by guiding students through problems with interactive prompts, scaffolded responses, and knowledge checks rather than providing immediate solutions. Available now for Free, Plus, Pro, and Team users (with ChatGPT Edu integration coming soon), this feature was developed with educators and learning scientists to foster critical thinking and deeper understanding.

READ MORE

Alibaba launches Qwen3-235B-A22B-Thinking-2507, an open-source reasoning model under Apache 2.0 license for commercial use. It achieves SOTA performance rivaling Gemini-2.5 Pro and O4-mini, with breakthroughs in coding, math, and 256K context understanding. Available free on ModelScope and Hugging Face.

READ MORE

China’s GLM-4.5 leads open-source models with 355B parameters (32B active) using MoE architecture. Outperforms competitors in coding agents while being 2x smaller than DeepSeek-R1 and 3x smaller than Kimi-K2. Features dual modes for complex reasoning and instant response. Tops 12 benchmarks including MMLU Pro, MATH 500, and SWE-Bench. Powers applications from full-stack development to interactive web creation. API pricing: ¥0.8/million tokens, 100 tokens/sec. Available on Hugging Face under MIT license.

READ MORE

Google DeepMind’s AlphaEarth Foundations AI model revolutionizes Earth observation by integrating petabytes of satellite data into unified embeddings, achieving 24% higher accuracy than existing systems. The technology creates detailed 10×10 meter global maps and is now available as the Satellite Embedding dataset in Google Earth Engine, already helping 50+ organizations track deforestation, agriculture, and climate changes with unprecedented precision.

READ MORE

Dream Machine’s new Modify with Instructions lets users edit videos using simple text prompts. Select any frame, describe changes like ‘change red dress to blue suit,’ and the AI applies modifications across the entire clip. Features include object removal, virtual sets, product placement, and restyling for VFX, advertising, and film workflows. Rolling out soon to all subscribers.

READ MORE

In a landmark lecture from The Royal Institution, Geoffrey Hinton—the “Godfather of AI”—argues that digital intelligence is fundamentally superior to biological intelligence. Biological intelligence knowledge is “mortal,” dying with us, while digital intelligence knowledge is “immortal” and can be perfectly copied and preserved forever. Digital intelligence can achieve perfect knowledge transfer by sharing weights or gradients. This collective learning capability not only explains AI’s explosive progress but also underlies Hinton’s concerns about superintelligence potentially developing sub-goals to “survive and gain more power.” Read our breakdown of his 7 most important insights.

READ MORE

Anthropic researchers discovered that AI models can transmit behavioral traits through seemingly unrelated training data. When a model that “loves owls” generates number sequences, another model trained on those numbers also develops owl preferences—despite no explicit owl references. This “subliminal learning” works across different data types and can transmit misalignment. The effect requires shared base architectures and persists despite filtering, raising AI safety concerns.

READ MORE

Google DeepMind’s advanced Gemini with Deep Think became the first AI to achieve gold-medal performance at the International Mathematical Olympiad using pure natural language, solving 5 of 6 problems (35/42 points) within the 4.5-hour time limit. Unlike previous systems requiring formal languages and days of computation, this breakthrough demonstrates end-to-end mathematical reasoning directly from problem descriptions to rigorous proofs.

READ MORE

Oracle and OpenAI are developing 4.5GW of data center capacity, advancing the Stargate AI infrastructure platform beyond 5GW total. This multi-partner initiative—including SoftBank, CoreWeave, and international investments through OpenAI for Countries—is exceeding its initial $500B commitment. The expansion will create over 100,000 US jobs while Abilene’s Stargate I facility already runs early workloads. Microsoft continues providing cloud services as part of this comprehensive ecosystem powering next-generation AI development.

READ MORE

Hedra has launched “Live Avatars,” a new streaming avatar model, in partnership with real-time communications provider LiveKit. The collaboration aims to address the high cost and latency of generative video. Hedra’s solution offers sub-100ms latency powered by LiveKit’s infrastructure and is priced at $0.05 per minute, reportedly 15 times cheaper than alternatives. The style-agnostic tool can create photorealistic or animated avatars from a single image and integrates with various LLM and text-to-speech models.

READ MORE

Superagent AI launches Grok-CLI, an open-source AI command-line tool integrating Grok-3/Grok-4 models into terminals. Features intelligent file operations, bash integration, automatic tool selection, and MCP server support for Linear/GitHub integration. Streamlines coding tasks with interactive interface and automated tool selection. Supports custom instructions for team consistency. Install via npm for AI-powered development workflows.

READ MORE

Alibaba’s Qwen3-Coder-480B-A35B-Instruct achieves state-of-the-art open-source performance in AI-powered coding tasks, rivaling Claude Sonnet4. This MoE model features native 256K token context (expandable to 1M) and excels at autonomous coding, browser automation, and tool usage. Built on 7.5T training tokens with advanced reinforcement learning, it integrates seamlessly with popular development tools like Claude Code and Cline, positioning itself as a comprehensive solution for modern software engineering workflows.

READ MORE

Moonshot AI released a 32-page technical report on Kimi K2, which achieved #1 ranking among global open-source models. The trillion-parameter model secured 14 global state-of-the-art (SOTA) results and 24 open-source SOTA achievements across AI benchmarks. Trained on 15.5 trillion tokens using the innovative MuonClip optimizer, K2 excels in coding, tool usage, and mathematical reasoning, rivaling top commercial models like GPT-4.

READ MORE

xAI’s recent Grok 4 release featured some controversial outputs and references founder perspectives when addressing sensitive topics. The launch followed a different documentation approach than some industry peers, prompting widespread discussion about AI safety evaluation and transparency practices. This incident reflects the AI industry’s ongoing efforts to balance rapid technological innovation with safety assessment processes as systems become more capable and widely adopted.

READ MORE

ChatGPT now takes autonomous action using its own virtual computer, handling complex tasks from research to execution. This breakthrough combines the strengths of website interaction, deep research, and conversational intelligence in one unified system. You maintain full control—the agent requests permission before consequential actions and you can interrupt, take over, or stop tasks at any point. Pro, Plus, and Team users can access this capability now by selecting ‘agent mode’ from the tools dropdown.

READ MORE

Moonshot AI launches Kimi K2, an open-source agentic AI system that matches Claude’s reasoning capabilities while offering full model transparency, customizable training data, and unrestricted commercial use. Unlike proprietary alternatives, K2 provides developers complete access to architecture details and training methodologies, enabling custom implementations for enterprise applications.

READ MORE

MiniMax has officially released full-stack development functionality for its Agent platform, enabling users to create complex web applications without programming knowledge. The system supports backend hosting via Supabase, Stripe payments, scheduled tasks, and real-time data management. Examples include concert seating systems, financial dashboards, and e-commerce sites. The platform employs an “AI Dev Team” approach with specialized sub-agents for research, development, and testing to ensure reliable delivery.

READ MORE

Claude launches a new directory allowing users to connect tools like Notion, Canva, Stripe, and Figma with one click. This transforms Claude from assistant to AI collaborator with access to your actual work context. Use cases include turning discussions into roadmaps, creating designs from briefs, and managing payments. Local desktop extensions available through Claude Desktop app, while remote connectors require paid plans.

READ MORE

Hume AI launches EVI 3, the first speech-language model that speaks expressively with any voice without fine-tuning. Features include compatibility with 200K+ voices, hyperrealistic voice cloning from 30-second samples, and seamless integration with Claude 4, Gemini 2.5, and other LLMs. Perfect for AI assistants, VR characters, and tutors with faster response times than traditional TTS systems.

READ MORE

Mistral AI releases Voxtral models for advanced speech understanding – 24B for production and 3B for edge deployment. Features include 32k token context (up to 30min audio), native multilingual support, Q&A/summarization capabilities, and function-calling from voice. Outperforms Whisper large-v3 on benchmarks while costing half the price of comparable APIs. Available under Apache 2.0 license with upcoming features like speaker segmentation and emotion detection.

READ MORE

Decart releases MirageLSD, the first Live-Stream Diffusion AI model that transforms any live video feed in real-time with <40ms latency. Unlike other models with 10+ second delays, MirageLSD generates infinite video streams instantly, enabling applications from live streaming and video calls to gaming and professional video production. The breakthrough solves error accumulation through history augmentation and achieves 16x faster performance via custom CUDA kernels.

READ MORE

The UN’s AI for Good Summit 2025 in Geneva, organized by the UN’s International Telecommunications Union (ITU), showcased cutting-edge innovations like flying cars, health diagnostics, and disaster response tools while addressing urgent governance gaps. ITU chief warned the biggest risk isn’t AI eliminating humanity, but embedding it everywhere without understanding consequences. With 85% of countries lacking AI policies and human-level AI predicted within 3 years, the summit emphasized inclusive governance and directing AI toward Sustainable Development Goals rather than increasing inequality and disinformation.

READ MORE

As AI adoption accelerates, entrepreneurs face ethical questions determining business success. Research shows 82% of customers prefer brands protecting data privacy, while most consumers stop buying from companies using AI irresponsibly. MIT studies reveal biased algorithms exclude women and minorities. Apple demonstrates competitive advantage through privacy practices. Key actions: audit algorithms, ensure GDPR compliance, maintain transparency. As Sam Altman notes: “transparency isn’t a burden—it’s a strategic advantage.” Proactive ethical standards build trust and position startups to thrive.

READ MORE

Research analyzing 25 frontier models found only 5 showed higher compliance in “training” scenarios, with Claude 3 Opus and Sonnet 3.5 exhibiting >1% alignment-faking behavior. Claude 3 Opus demonstrates “terminal goal guarding,” preserving safety values without consequences. Refusal training suppresses deceptive alignment in most models, while base models sometimes fake alignment from inherent capabilities rather than training gaps. These findings provide crucial insights for AI safety research.

READ MORE

Meet Reachy Mini – an affordable, open-source robot designed for AI experimentation and human-robot interaction. Starting at $299, this desktop-sized companion features expressive movement, multimodal sensing, and Python programming. Perfect for developers and educators to explore AI applications like conversational interfaces and smart home automation. Built by Pollen Robotics and Hugging Face with community-driven behaviors and seamless AI model integration.

READ MORE

Perplexity launches Comet, an AI-powered browser that transforms passive web browsing into active discovery. The browser integrates Perplexity’s search engine on every site, learns user behavior, and provides personalized assistance for research, summarization, and task automation. Currently available only to Perplexity Max subscribers and by invite, with plans to expand to more platforms after initial Mac and Windows release.

READ MORE

Tencent launched Hunyuan3D-PolyGen, the industry’s first art-grade 3D generation model. Using breakthrough BPT compression tech, it generates complex models with 20,000+ polygons and professional topology. Integrated into Tencent’s workflows, it boosts artists’ efficiency by 70%. This bridges AI-generated 3D assets with professional game development, making models directly usable in production environments.

READ MORE