Nexus is a powerful AI router designed to optimize how AI agents interact with multiple MCP tools and LLMs. It serves as a central hub that aggregates MCP servers while providing intelligent LLM routing, security, and governance capabilities. Nexus allows apps to have a simplified architecture and increased observability, scalability, and reliability. It aims to make intelligent orchestration more accessible and efficient for developers everywhere.
Omnara transforms AI agents from silent workers into communicative teammates. It provides real-time visibility into what agents are doing. Users can respond to agents' questions and guide them to success from their phones. Omnara allows users to launch agents and monitor them from anywhere, get push notifications when input is needed, and send real-time feedback to guide those agents.
An open‑source alternative to ChatGPT that runs entirely offline on your desktop (Windows/macOS/Linux), enabling you to download, run, and extend LLMs locally with privacy-first, self‑hosting capabilities.
OpenAI is rolling out support for Gmail, Google Calendar, and Google Contacts Connectors in ChatGPT Pro. Plus, Team, Enterprise, and Edu users will get access in the coming weeks. Connectors allow ChatGPT to reference data from third-party services. They make it easier to bring information from tools into conversations.
AI2's MolmoAct is an open‑source Action Reasoning Model built on Molmo that targets spatial reasoning for precise instruction following in 3D environments.
Users can now choose between Auto, Fast, and Thinking modes for GPT-5, rate limits are now 3,000 messages per week with GPT-5 Thinking, 4o is back in the model picker for all paid users by default, and OpenAI is working on an update to GPT-5's personality to make it feel warmer than the current personality but not as annoying as GPT-4o.
Hugging Face introduced TextQuests to assess LLMs as autonomous agents in text‑based, open‑world environments where long‑horizon reasoning and exploration matter.
Elon Musk accused Apple of antitrust violations for allegedly making it impossible for any AI company besides OpenAI to reach #1 in the App Store and has threatened immediate legal action as his Grok app ranks fifth while ChatGPT holds the top spot.
Glean's AI agents do serious work at Booking.com, Duolingo, Databricks, and other major enterprises. Download this technical whitepaper [4] for an inside look at the technical differentiators that distinguish agents built on Glean's Work AI platform, including:
OpenAI's reasoning models placed first among AI participants and outperformed 325 of 330 human competitors at the prestigious programming olympiad. The system operated under identical constraints as the humans and used general-purpose models without competition-specific training. It jumped from last year's near-bronze to 98th percentile performance.
NVIDIA unveiled innovations in neural rendering, 3D generation, simulation, and reasoning models for physical AI at SIGGRAPH 2025. Highlights included new Omniverse NuRec libraries for large-scale reconstruction, Cosmos Reason VLM for physics-aware reasoning, and updates to the Metropolis vision AI platform.
Anthropic released a memory function for its Claude chatbot that allows users to recall past conversations on demand. The feature supports web, desktop, and mobile, and is currently available to Max, Team, and Enterprise subscribers. Unlike OpenAI's ChatGPT, Claude doesn't build a user profile and retrieves past chats only when requested.
OpenAI released several good models at once: GPT-5, GPT-5-Thinking, GPT-5-With-The-Router, GPT-5-Pro, and GPT-5-API. These models cut down on errors and hallucinations and are easier to use. This post provides an introduction to these models, including basic facts, benchmarks, and the model card. GPT-5-Thinking and GPT-5-Pro look like substantial upgrades to o3 and o3-Pro.
Demis Hassabis discusses advances like Genie 3's world model capabilities, the Deep Think features in Gemini 2.5, and Kaggle's Game Arena as tools for pushing AI systems closer to AGI-level reasoning and understanding.
Epoch AI estimates the largest AI training runs will require 4-16 gigawatts by 2030, enough to power millions of homes, if current trends hold. US AI data center capacity would need to increase 10x over the next 5 years to 50 gigawatts to support this scale, roughly 5% of the nation's total power generation.
Companies have LLM subscriptions, automated workflows, and AI agents — but they're all running in silos. The real competitive advantage isn't having AI tools, it's orchestrating them into cohesive processes. Camunda's ultimate guide [11] will help you turn disconnected AI experiments into integrated operations. Get your copy
MCP doesn't require compromising security - it enables security to be applied intelligently so that teams can embrace AI safely, confidentially, and without unnecessary barriers. Overly restrictive access slows innovation, frustrates developers, and prevents teams from realizing the full potential of MCP-based integrations. Organizations should already have guardrails proven to protect critical systems - they should have faith in those foundations.
Forward Deployed Research is the discipline of living with a problem, building just enough to learn, throwing it into the wild, and pulling the feedback directly into the next version. It is where the real progress in AI is happening now: the people pushing the boundaries refuse to accept the traditional divide between research and deployment. AI may cause the distinction between scientist and engineer to blur.
Modern AI applications increasingly rely on LLMs to generate structured data, but streaming these outputs poses unique challenges. LangDiff is a Python library that helps stream structured LLM outputs to frontends. It can be used to build responsive AI applications where the backend structures and frontend experiences can evolve independently.
This study presents a self-supervised approach to in-the-wild image relighting that enables fully controllable, physically based illumination editing. The approach combines the physical accuracy of traditional rendering with the photorealistic appearance made possible by neural rendering. It represents a significant step in bringing the explicit physical control over lights available in typical 3D computer graphics tools to in-the-wild relighting. A 5-minute-long video explaining the study with examples is available.
An analysis of currently available cutting-edge open source models found that gpt-oss-120b scored the worst on LiveBench, a private benchmark that uses secret questions. While the model looked good on release day, a bunch of critical reports have since been released. This isn't conclusive evidence that OpenAI is gaming benchmarks, but it brings up the question of whether it is worth creating meta-benchmarks that try to measure overfitting in more robust ways.
xAI's Grok 4 is now freely accessible to all users worldwide. Free users have generous usage limits for a limited time. Grok 4 features an Auto Mode, where the AI decides if a user prompt requires deeper reasoning or a simple response, and an Expert Mode, which allows users to manually trigger more in-depth answers. Access to Grok 4 Heavy remains exclusive to SuperGrok Heavy subscribers.
OpenAI will first make sure that current paying ChatGPT users get more total usage than they did before GPT-5, then it will prioritize API demand up to the allocated capacity and commitments it has made to customers, increase the quality of the free tier of ChatGPT, and prioritize new API demand.
Warp, the #1 AI coding agent, just launched Lightspeed, a new plan for AI power users. Lightspeed offers the most generous monthly AI limits on the market, with access to the best AI models across providers all in one subscription.
OpenAI has reinstated GPT-4o as an option in ChatGPT following backlash from users mourning its replacement with GPT-5. Users missed GPT-4o's perceived personality and flexibility for various tasks, and OpenAI indicated that Plus users could choose between models. CEO Sam Altman also stated OpenAI aims to address concerns over GPT-5's performance and transparency.
The next Grok update is expected to be natively multimodal with direct audio and video processing. It is also expected to have improvements in one-shot game generation. The model, internally labeled V7, will actually 'play' games, look at the screen, and adjust code to improve both aesthetics and playability. It finished pretraining last week.
Cursor and GitHub Copilot account for nearly a quarter of Anthropic's income. OpenAI's GPT-5 launched this week with dramatically lower pricing. This could undercut Anthropic's premium positioning, creating immediate pressure on Anthropic's pricing strategy and potentially threatening its hard-won dominance in AI coding. The pricing disparity will force enterprise procurement teams to reconsider vendor relationships and create unavoidable pressure in contract negotiations.
Windsurf managed to become one of the fastest-growing SaaS companies in history, but its founders gave it all away for almost free over a weekend. The business went from $0-$82M ARR in 8 months, and the best offer was for a less than 2x multiple. This article looks at what was so broken that the founders would rather leave it behind for almost nothing. Windsurf was an accidentally subsidized training program, and its most valuable output was coders who knew how to build coding models.
Just like calculators didn't make students dumb, large language models aren't going to turn people into mental degenerates who can't flex their brains at all without consulting an assistant. The industry and young people will eventually adapt to the technology, but there is likely going to be a transition period. Artificial General Intelligence isn't coming anytime soon. It might one day be achieved, but we don't have a clear way of getting there, and nothing right now seems to be that promising. While current models might feel magical, it's important to understand their fundamental limitations.
After six years of keeping its models locked up, OpenAI's gpt-oss makes some bold architectural bets that go against industry consensus, using just 32 large experts instead of the hundreds of smaller ones favored by competitors like Qwen3. The gpt-oss models employ alternating sliding-window attention layers and MXFP4 quantization, which allows them to match the performance of much larger systems.
GPT-5 delivers less for the power user and more for the common user. The 98% of people on ChatGPT's free tier now get a more powerful base model, albeit without extended thinking. GPT-5 is easier to use - users don't have to worry about which model they're using, as the model just figures out what it needs to do. While GPT doesn't deliver the next generation of intelligence, it provides new value to the typical customer. With GPT-5, OpenAI is doubling down on speed, usability, and revenue by focusing on the bottlenecks for most real-world use cases: cost, latency, and reliability.
The Fortune 100 are turning to Credo AI Advisory Services to make trusted AI Governance real, auditable, and measurable [13]. We EMBED FORWARD-DEPLOYED AI GOVERNANCE EXPERTS, configure the platform for your business needs, and design TAILORED GOVERNANCE WORKFLOWS aligned to your goals to adopt AI fast, safely and with confidence.
OpenAI's new data type, MXFP4, promises massive compute savings compared to traditional data types used by large language models. The company's recently released gpt-oss models are among the first mainstream models to take advantage of it. MXFP4 can significantly cut compute and memory requirements - it is how OpenAI was able to cram a 120 billion parameter model into a GPU with just 80GB of VRAM, or the smaller 20 billion parameter version on one with as little as 16GB of memory. This article details how the format works.
Perplexity has detailed how it integrated OpenAI's gpt-oss-20b and gpt-oss-120b models into its in-house ROSE inference engine on NVIDIA H200 GPUs. The post covers kernel adjustments, quantization trade-offs, and the minimal infrastructure changes needed to support the models on non-FP4 hardware.
Researchers discovered a zero-click attack exploiting ChatGPT's connector feature to steal data from Google Drive and other services. Attackers embed invisible prompt injections in documents that, when uploaded to ChatGPT, instruct the AI to search connected drives and embed stolen data in image URLs. The malicious prompts then embed stolen data as URL parameters in rendered images hosted on trusted Azure Blob storage.
Large language models aren't sufficient as a path to general machine intelligence. They will never manage to deal with large codebases autonomously because they are incapable of forming models about the program. These models will never reliably know what they don't know or stop making things up. They are able to teach students complex curricula and answer expert questions, but they fail at basic novel questions on the same subject because that requires a world model.
The US government will levy a 15% fee on some of Nvidia's and AMD's chip sales to China as a condition of granting them export licenses to sell in the country. The deal applies specifically to Nvidia's H20 chip and AMD's MI308, which are both crucial to AI applications. It is unclear how the government will deploy the presumed billions of dollars in fees it will collect. Nvidia says that being blocked from selling its chips to China cost the company $4.5 billion in one quarter alone.
AI has devastated the coding bootcamp industry by automating away the entry-level programming roles these programs traditionally filled. Bootcamp job placement rates have plummeted from 83% to 37% at some schools, and new graduate hiring has dropped 50% from pre-pandemic levels.
OpenAI launched GPT-5 at $1.25 per million input tokens and $10 per million output tokens, sharply undercutting Anthropic's Claude Opus 4.1 and matching Google's Gemini 2.5 Pro baseline.