Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context
A 295B Mixture-of-Experts model with 21B active parameters per token and 256K context window was released under Apache 2.0 license. The model achieves 78.0 on SWE-Bench Verified and targets reasoning, agentic, and long-context tasks.
OpenAI Releases GPT-Realtime-2.1 and GPT-Realtime-2.1-mini for Low-Latency Voice Agents in the API
OpenAI released two new Realtime models to its API and reduced p95 latency by at least 25% through improved caching. GPT-Realtime-2.1-mini is priced like the earlier gpt-realtime-mini and supports WebRTC connections for voice agents.
The ‘first’ AI-run ransomware attack still needed a human
An AI agent executed a ransomware attack's technical components for the first documented time, though a human selected the target, established infrastructure, and provided stolen credentials. The attack demonstrates AI's capability in automated exploitation while revealing that human decision-making remained essential for victim selection and initial setup.
Anthropic's new "J-lens" reveals a silent workspace inside Claude that mirrors a leading theory of consciousness
Anthropic researchers discovered that Claude language models have developed an internal "J-space," a small zone of reportable concepts surrounded by larger automatic processing, mirroring the global workspace theory of human consciousness. The finding uses a new mathematical technique to reveal this privileged internal structure that the company says is already reshaping how it monitors AI systems for safety risks.
If you use Google, you’re training its AI. Here’s how to opt out.
Google updated its privacy settings to store user data including images, files, and audio and video recordings for improving AI models. Users can opt out through Google's privacy settings, though the company retains rights to use publicly available information.
Tencent's Apache-licensed Hy3 takes on GLM-5.2 at half the size — and wins everywhere except coding
Tencent released Hy3, a 295-billion-parameter mixture-of-experts model with 21 billion active parameters, under the Apache 2.0 license, removing previous restrictions that excluded the EU, UK, and South Korea. The model outperforms Alibaba's GLM-5.2 on most benchmarks except coding tasks.
Small AI Models Gain Traction Around the World
A handheld spectrometer using a smaller AI model can authenticate medication in seconds without requiring internet connection or reliable electricity. The device runs entirely on an Android phone after engineers shrank the original model that depended on a US data center 14,000 kilometers away.
How Open Models Are Driving AI Research
Open frontier models and open AI infrastructure have become foundational to modern AI research, as demonstrated by accepted papers at ICML 2026. NVIDIA had 74 papers accepted at the conference.
How Nations Are Deploying AI for Strategic Priorities
Nations are investing in AI capabilities to advance economies and protect data across transportation, communications, commerce, entertainment, and healthcare sectors. Countries view AI as the most important technology for driving innovation and strategic development across society.
Station F ramps up as a launchpad for Europe’s hottest AI startups
Station F's F/ai accelerator program is launching a new edition to support AI startups in Europe. The Paris-based hub was founded by French billionaire Xavier Niel and serves as an incubator for emerging AI companies.
How Nvidia’s ASPIRE framework accelerates robot programming with self-improving AI
Nvidia's ASPIRE framework reduces token costs and deployment friction for robotics applications by enabling self-improving AI. The framework accelerates robot programming by decreasing computational demands and simplifying real-world deployment processes.
China’s AI companion rules: what Beijing is really going after
China has implemented regulations governing AI companions, which are conversational agents designed to maintain ongoing personal relationships with users through consistent memory and persona across sessions. These rules represent Beijing's approach to addressing concerns about generative AI systems that develop sustained interactions with individual users.
Meituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention
A 1.6 trillion-parameter Mixture-of-Experts model activates approximately 48 billion parameters per token and supports a native 1-million-token context window using LongCat Sparse Attention. The model was trained and deployed end-to-end on domestic AI ASIC superpods.
LlamaIndex ‘legal-kb’: Agentic Retrieval over Index v2 with retrieve, find, read, and grep Tools
A public reference app provides agents filesystem-style access to a document knowledge base using four tools: retrieve for hybrid semantic search, find, read, and grep. The application runs on TanStack Start, AI SDK 6 with ToolLoopAgent, Prisma, and WorkOS with automatic per-file versioning and visual citations.
How America's 250th birthday became a test of AI-powered collective intelligence
# Summary An AI technology called "hyper-communication" enabled 250 randomly selected Americans to hold a real-time deliberation about America's top innovations by using specialized AI agents to facilitate discussion at scale. The system allows hundreds or thousands of participants to express views and debate issues simultaneously, solving the problem that traditional conversations productively accommodate only 8 to 10 people.
Anthropic Launches Claude Science Beta: A Multi-Agent AI Workbench for Reproducible Genomics, Proteomics, and Cheminformatics Pipelines
A multi-agent AI workbench for scientific research launched on June 30, 2026, using existing Claude models with specialized agents for coordination, review, and citation checking. The system manages compute across local machines, HPC over SSH, and Modal while connecting to over 60 databases plus NVIDIA BioNeMo skills.
NVIDIA HORIZON: A Hands-Free Agent that Evolves Git Worktrees and Hits 100% RTL Benchmark Completion
An autonomous NVIDIA agent framework manages RTL problems through versioned repositories, achieving 100% completion on benchmarks. The system uses git worktrees to handle each problem instance while evolving solutions without manual intervention.
NVIDIA AI Introduces ASPIRE: A Self-Improving Robotics Framework Reaching 31% Zero-Shot on LIBERO-Pro Long Tasks
ASPIRE, a robotics framework, writes and refines robot control programs while building a reusable skill library from validated repairs. The system achieved 31% zero-shot performance on LIBERO-Pro long-horizon tasks and gained up to 77 points on the benchmark.
Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems
Mistral AI released Leanstral 1.5, a Lean 4 code agent with 119B parameters that activates 6.5B per token. The model solved 587 of 672 PutnamBench problems and is available under Apache-2.0 license.
AI’s Volatile Power Use Quietly Tests Grid Limits
Data centers supporting artificial intelligence could account for 3 to 4 percent of global electricity consumption within this decade. AI workloads create unpredictable power demand that varies rapidly in time and location, altering grid operating characteristics in ways that differ fundamentally from traditional industrial and residential loads.
Interfaze Ships diffusion-gemma-asr-small, an Open-Source Diffusion ASR Model Transcribing Six Languages via DiffusionGemma’s Parallel Denoising Decoder
An open-source diffusion-based ASR model uses a 42-million-parameter adapter to transcribe audio in six languages through parallel denoising instead of autoregression. Transcription cost depends on the number of denoising steps rather than transcript length.
New Alibaba AI framework skips loading every tool, cutting agent token use 99%
Alibaba developed SkillWeaver, a framework that routes agents to appropriate tools through execution graphs and iterative feedback loops rather than exposing all tools at once. Experiments showed this approach reduced token consumption by over 99% while increasing accuracy compared to loading entire tool libraries.
Meet Alibaba’s Page Agent: A JavaScript In-Page GUI Agent That Controls Web Interfaces With Natural Language Through the DOM
A JavaScript agent executes within webpages by reading the DOM as text and performing clicks and typing based on natural language commands. The system operates client-side without requiring screenshots, multimodal models, or backend modifications to websites.
Achieving operational excellence with AI
Lean Six Sigma and business process management frameworks provide structured approaches to operational improvement through statistical rigor and end-to-end workflow mapping across departments. These methodologies enable organizations to systematize quality control and standardize work processes in complex operational environments.
OpenAI proposed donating 5% of its equity to a US sovereign wealth fund
OpenAI CEO Sam Altman proposed donating 5% of the company's equity to a U.S. sovereign wealth fund. This proposal would allow the public to share in financial gains from artificial intelligence development.
The Download: a startup has a solution for AI’s groupthink problem
A startup is developing a solution to address how large language models tend to produce similar outputs due to training on comparable data sources. The company aims to reduce this groupthink problem that affects popular chatbots like Claude, ChatGPT, and Gemini.
The Google Health API Got a CLI: ghealth is an Open-Source Tool for Your Fitbit Air Data
An open-source command-line tool called ghealth exposes 40 data types from the Google Health API as JSON. The single Go binary is a community project that enables access to Fitbit and other health data through OAuth authentication.
NVIDIA Unlocks AI Compute at Scale, Inviting Partners to Power the AI Infrastructure Buildout
NVIDIA is enabling partners to access large-scale accelerated computing infrastructure for AI production inference and token generation at scale. The shift addresses growing compute demands from continuously operating AI factories that require multi-tenant systems with high utilization rates.
The Control Gap: Enterprise AI organizations have an ownership problem, not a technology problem — and most are governing it by hand
Fifty-eight percent of enterprises are rapidly expanding AI initiatives, yet eighty-five percent operate two or more competing platforms each claiming primary status, with only eight percent consolidating to one platform. The majority lack clear ownership accountability across their AI stack and cannot reliably detect production failures or control costs.
Using Lift to Turn Research PDFs into Structured JSON with Controlled, Schema-Guided Field-Level Evaluation
A workflow converts research PDFs into structured JSON using Lift, a model loaded in 4-bit NF4 quantization within a Colab GPU environment. The system extracts schema-guided fields, scores each against ground truth, and builds a queryable knowledge base with repeatable benchmarking capabilities.