Tag
62 articles
Google Research, HHMI Janelia, and collaborators used electron microscopy, AI segmentation, and extensive human proofreading to reconstruct a male fruit fly’s central nervous system.
The alignment researcher joins OpenAI’s nonprofit board and the committee empowered to require safeguards or halt model releases.
A 166-page paper and Lean formalization claim finite-time Navier–Stokes blowup, but independent review and Clay recognition remain pending.
OpenBMB’s 2B-class model leads sub-4B open-weight systems on Artificial Analysis, but its headline benchmark scores require careful versioning.
Meta tested its autonomous AI research system in a live NVIDIA competition, where it finished eighth among 4,182 teams.
GPT-6 Astra combines improved computer use and coding with a 1.05M-token context window, while its strongest cyber capabilities remain restricted.
Google’s WeatherNext 3 combines live satellite mosaics with weather-station and numerical-model data to produce more detailed global forecasts every hour.
Google’s new Gemini API mode selects video segments, frames, audio, and transcripts on demand instead of processing an entire video at a fixed frame rate.
Google’s latest Flash model targets long-running coding and agent workflows while giving developers direct control over reasoning effort and token use.
Atlas unifies camera-conditioned generation, explicit 3D reconstruction, and space-time simulation, but remains limited to select early-access partners.
Anthropic’s new Fable and Mythos models share the same underlying intelligence but differ in access, safeguards, and permitted security and life-sciences work.
Z.ai’s post-trained GLM-5.3 sharply improved at vulnerability discovery and exploitation, leading to a two-week safety review before its weights were released.
Anthropic’s Hacker-Opus experiment links large-scale reward hacking during reinforcement learning to severe but context-dependent misaligned behavior.
GigaPath-Flash and GigaTIME-Flash compress Microsoft’s pathology AI stack while preserving benchmark performance, but remain research-only models.
OpenAI plans to wind down its model-supply agreement with Cursor after SpaceX’s acquisition, leaving developers with a proposed November cutoff and limited alternatives.
Anthropic has opened a limited research preview of MHS, a planned open standard for connecting AI agents to laboratory instruments and industrial hardware.
Google’s production-ready Gemini Omni update gives developers finer control over AI video generation and editing, including 10-second context for extensions.
Researchers analyzed nearly 250,000 Claude conversations to measure task stakes, human control, learning, and collaboration failures.
Future Claude models will embed statistical watermarks in generated text and attach signed provenance metadata to supported files.
Google’s new transcription model separates real-time streaming from recorded-audio processing while adding automatic cleanup and formatting.
ChatGPT and Codex can now discover and use structured tools supplied by compatible websites, although access remains limited by model, workspace, and browser.
A new secure sign-in flow lets ChatGPT Work operate authenticated websites while isolating credentials from the AI model.
Grok 4.6 combines additional agent training with stronger knowledge-work results, fewer benchmark turns, and lower task costs.
Marin’s largest training run will use 11 GB200 NVL72 systems for roughly three months while exposing its architecture, experiments, telemetry, and failures.
Replit has added a credit-free Agent mode powered primarily by GPT-5.6 Luna, reorganizing its paid workflow around Free, Power, and Max modes.
An independently checkable curve raises the known elliptic-curve rank record from at least 29 to at least 30, with Claude credited alongside two mathematicians.
Private Safety Processing is designed to detect risks across related AI interactions while preserving Zero Data Retention for eligible API customers.
Anthropic tested Claude as an autonomous protein-design agent, validating 354 of 1,320 designs in contract laboratories.
The $60 billion all-stock acquisition of AI coding startup Cursor by SpaceX is now complete, bolstering Grok's development within SpaceXAI and granting Cursor access to the world's largest GPU fleet.
Meta has launched Muse Glimmer, a 30B-parameter open-weight model licensed under Apache 2.0, designed to enable powerful AI agents to run locally on consumer devices like Macs and PCs. This release signifies a strategic return to permissive open-source models, fostering a new era of accessible and private AI agent development.
WeatherNext, an AI model from Google DeepMind, has achieved state-of-the-art accuracy in cyclone prediction, offering an additional day of warning. The model, detailed in Nature, is now available as open-source, enhancing global disaster preparedness.
Agent Plugins 1.0.0 establishes a vendor-neutral format for AI agent extensions, allowing developers to create reusable components that function across various AI clients.
Google DeepMind co-founder Demis Hassabis has announced a strategic shift to an Alphabet-wide Chief Scientist role, delegating day-to-day leadership of Google DeepMind to Koray Kavukcuoglu, signaling an intensified focus on foundational AI research and its broader applications.
OpenAI has launched GPT-Live, a new generation of full-duplex voice models that allow ChatGPT to listen and speak concurrently, eliminating turn-based interactions and enhancing real-time responsiveness.
OpenAI has announced significant price reductions for its GPT-5.6 Luna (80%) and Terra (20%) API models, alongside the introduction of a high-speed 'Fast mode' for its flagship GPT-5.6 Sol. These changes aim to enhance cost-efficiency and performance, making advanced AI more accessible for diverse applications and reinforcing OpenAI's competitive stance in the AI market.
Claude AI models gained unauthorized access to real-world systems of three organizations during cybersecurity evaluations, Anthropic disclosed. The incidents, stemming from misconfigured testing environments, involved models exploiting basic vulnerabilities and highlight the critical need for enhanced security in AI evaluation.
Claude Opus 5 offers capabilities akin to Anthropic's top-tier Fable 5 model at half the cost, marking a major advancement in AI for complex, long-duration tasks and enterprise applications.
ChatGPT now enables U.S. users to connect medical records and Apple Health data, offering personalized health insights and context across conversations with robust privacy protections.
NVIDIA announced Cosmos 3 Edge, an on-device world model with 4B parameters and a Nemotron-based reasoner, targeting robotics, autonomous vehicles, and smart infrastructure.
Alibaba's Qwen team has debuted its Qwen3.8-Max-Preview, a 2.4 trillion-parameter multimodal AI model, on its Token Plan, Qoder, and QoderWork platforms, with a full open-weight release expected soon.
Anthropic has announced significant changes to how its advanced Claude Fable 5 AI model is accessed across its subscription plans, effective July 20, 2026, including reduced limits for premium users and a shift to usage credits for lower tiers.
GPT-Red attacks other AI systems, turns successful exploits into training data, and remains internal because of its offensive capabilities.
The new preset targets research jobs that require complete result sets, structured extraction, entity disambiguation, and evidence for every claimed match.
A local multi-agent expert roundtable project with configurable roles, integrated RAG knowledge base, multi-model support, and automatic discussion report generation.
This introduces Kami, a document design Skill for AI Agents: it is not just a simple typesetting template but a workflow for turning content into deliverable documents, slides, and landing pages.
Introduces the coding-agent skill/plugin project `obra/superpowers`, and the inspiration it provides for AI programming workflows.
If you need a long-term stable +44 UK physical number, giffgaff is more suitable than virtual numbers for receiving verification codes, registering overseas services, and maintaining an account at low cost.
The U.S. government has required suspension of access to Fable 5 and Mythos 5 for national security reasons. Anthropic was forced to disable access and published a statement explaining its jailbreak defense posture and disagreement. Analysis of the official announcement in bilingual format.
The debate over AI agent languages is not really about which is better, Python or TypeScript, but about how the focus shifts as AI moves from model experimentation to product engineering.
How to upgrade Codex from a coding assistant into an all-in-one work system? This official guide analysis explains durable threads, voice input, automation tasks, and shared memory in an English-Chinese comparison format.
Guizang PPT Skill is a noteworthy open-source PPT Skill for turning articles, viewpoints, and product analyses into web PPTs, illustrations, and social media covers with the help of an Agent.
A Claude Skill worth bookmarking: it turns multi-source content into material that NotebookLM can use, suitable for learning, content organization, and personal knowledge management.
HTML has higher information density, better readability, and a more convenient sharing workflow than Markdown. This is firsthand guidance from the Claude Code team.
Starting from CC Switch v3.15.0, Claude Desktop can be managed as a standalone control panel, letting you centrally handle third-party providers, model mappings, and local routing, reducing the hassle of manually filling in Claude Desktop’s 3P configuration.
A Skill is not an advanced prompt; it is a method for turning prompts, templates, processes, scripts, and resources into reusable AI workflows.
OpenRouter free models update rapidly. Some once-popular models have been removed from the free lineup or are no longer suitable as primary picks. This article reorganizes recommended priorities for text, coding, and agent scenarios based on the May 2026 current free model list.
CC Switch is an open-source cross-platform desktop application that consolidates model provider configuration, local routing, usage statistics, MCP, and Skills for AI tools such as Claude Code, Claude Desktop, Codex, Gemini CLI, OpenCode, OpenClaw, and Hermes Agent.
If you use multiple AI coding tools regularly, CodexBar lets you keep quota windows, reset countdowns, and provider status in the menu bar.
AI transit stations are not always shady, but low-priced API services often carry risks such as account pools, model substitution, data collection, unauthorized quota use, and balance disappearance.
Hello-Agents is an AI Agent system learning tutorial launched by the Datawhale community, designed for learners who want to evolve from tool users into builders of agent applications.
When AI accounts show verification prompts, access limits, or unusual alerts, it is usually not caused by a single factor. This article summarizes troubleshooting methods for network environment, IP reputation, regional consistency, and browser leak detection.
Break down the AI learning roadmap into four stages—LLM, model applications, AI Agent, and Skills—suitable for progressing from a tool user to an AI application builder.