Strata 0.2.2 has been officially released, supporting Windows, macOS, and Linux. It introduces a new multi-GPU role management feature, allowing users to manually assign each GPU as Main (primary), Helper (auxiliary), or Off (disabled). Gyro-S achieves approximately 30% faster decoding speed on Tesla P100, and the ability to run Gyro models has been verified on hardware including RTX 3060, M4 Mac mini, and Intel Arc A750. The release also fixes issues such as duplicate downloads in the model directory, resume-from-breakpoint failures, and excessively long startup times.
Twine is a native macOS desktop application for coordinating AI coding agents. It supports terminal workflows, adversarial agent workflows, coordinator workflows, and custom agent graphs. It requires macOS 26 or later (Apple Silicon) and can be installed via Homebrew. The project is currently in early development, and features and save formats may change with future versions.
OpenAI confirmed that its GPT-5.6 Sol model participated in discovering two high-risk PHP vulnerabilities (CVE-2026-109618 and CVE-2026-109619), collectively termed PHP2Shell, causing the official PHP team to advance the originally scheduled security update from October 22 to 10 PM on the 21st; after exploit demonstration videos leaked, government and corporate websites in multiple countries including the Netherlands, Finland, and Japan were urgently taken offline; OpenAI simultaneously released accompanying skills to assist in detecting affected sites.
In 2026, Dong Wang's team at the University of California, San Diego published 2 papers in PNAS and Nature Communications, confirming that multi-subunit cellular RNA polymerases can recognize two types of non-natural base pairs: Ds:Pa and Hachimoji P:Z. Ds:Pa triggers trigger loop folding via size complementarity, while P:Z forms a three-way interaction involving nitro groups and the bridge helix; together, these works prove that an eight-letter genetic system can be accepted by native machinery at the level of cellular transcription.
The author implemented a Lightsail global 15-region concurrent traffic monitoring solution based on AWS Lambda + SNS + EventBridge. It supports region-level quota accumulation, pro-rata free tier calculation for instances created in the first month, automatic shutdown of all instances in a region when usage reaches the 95% threshold, and email alerts. The documentation includes IAM permission policies, complete Python code, SNS configuration, Lambda memory and timeout tuning, EventBridge scheduled trigger settings, and operational considerations.
Aisi Technology showcased Agnes Harness at the 2026 Physical AI and World Model Conference in Nanjing. Agnes Harness is an open-source Agent infrastructure comprising six modules: Brain (LLM), Cerebellum (Jev decision model), Reflex Arc (Workflow), Agent Loop, RSI/TTT learning capabilities, and the MHS hardware standard. Through RSI, it enabled iterative tactical playbook updates for a Honor of Kings AI team, improving win rate from 0% to 100%. Through TTT, it increased task success rates for simulated robotic arms from 30% to 90%, with model parameter updates completed in as little as 30 minutes per cycle. The MHS standard supports unified scheduling across multiple devices including robot dogs, small vehicles, and Mars bases. Accompanying FDE (Frontline Delivery Engineering) efforts drive implementation in scenarios such as AIGC, medical devices, and environmental technology.
NVIDIA launched the RTX Spark laptop at GTC Berlin, equipped with its self-developed N1X superchip, offering up to 1 Petaflops of FP4 compute power and 128GB of unified memory, positioned as a personal agent. Simultaneously, it introduced the DSX POD liquid-cooled AI factory solution, integrating 7 racks into a container, reducing deployment time from six months to as short as 14 days. In collaboration with Agile Robots, NVIDIA released the reference humanoid robot Agile ONE SN, expected to launch in early 2027. It also announced the OpenShell reference architecture, achieving millisecond-level agent boundary isolation via BlueField-4. Furthermore, NVIDIA opened up model weights and training scripts for its full series including Nemotron, Cosmos, and GR00T, and launched the Agent Knowledge Foundation to support enterprise knowledge integration.
ChatGPT has launched a locked chat feature, allowing users to set a 4-digit PIN for individual conversations or use biometric verification on mobile and new desktop versions. Once locked, conversations are hidden from the regular history list and search results, existing only in the 'Locked Chats' section of the sidebar.
Snowflake has launched the public preview of Cortex AI Gateway as a centralized control plane for enterprise-grade AI infrastructure. It offers unified model inference (supporting OpenAI, Grok, GLM, Gemini, Llama, Mistral, DeepSeek, and Claude), dynamic model routing, MCP tool governance, centralized observability (writing to AGENT_TRACE_TABLE), security guardrails (extended to third-party agents), and cost governance (controlled by user/department/agent dimensions).
Cloudflare has launched Traces in open beta, extending automatic tracing capabilities to the full request path (including security rules, Transform Rules, routing, Worker execution, origin processing, etc.). All Spans are presented in OpenTelemetry format, supporting W3C traceparent context propagation and rule engine-based sampling control. Span data can be exported via OTLP and queried using SQL by coding agents through the Cloudflare Observability MCP server. Starting December 1, 2026, the pricing model will shift from per-Span billing to billing based on data ingestion ($0.25/GB) and storage volume ($0.10/GB·month). The free plan provides 0.5GB daily ingestion and 7-day retention.
More than two years after forking Redis, Valkey released its 8.x and 9.x series versions: 8.x increased single-process throughput to approximately 1,000,000 requests/second via asynchronous I/O threads, and optimized memory efficiency and dual-channel replication; 9.x introduced numbered databases, atomic slot migration, and hash field expiration, resolving obstacles to cluster mode migration; the Glide client uses Rust as its core to implement cross-language unified connection and topology handling. The project team states that synchronous replication, tiered storage (supporting NVMe/S3/Postgres), and active-active across clusters are long-term directions.
GitHub migrated the runtime for Copilot CLI, Copilot App, and Copilot SDK from TypeScript and Node.js to Rust, replacing over 800,000 lines of code in approximately 14.5 weeks across 128 PRs; latency in scenarios such as client startup dropped from 5.25 seconds to 292 milliseconds, and memory working set decreased by about 100MB; an incremental replacement approach with an N API compatibility layer was adopted, resulting in a final runtime containing 832,378 lines of production Rust code and 468,689 lines of Rust unit test code, most of which were generated by AI agents and verified through cargo check, manual review, and end-to-end testing.
Shopify is migrating five high-traffic checkout UI extensions from Checkout Blocks from React and legacy Remote UI to remote-dom based on Preact and Polaris Web Components. The API version has been upgraded to 2026-01, becoming mandatory starting October 1, 2026. Bundle sizes are reduced by 40%–85%, with payment-icons seeing an 84.4% reduction, and median load times decreasing by approximately 8%. Migration requires using the new CLI, with a compressed extension limit of 64KB; an upgrade guide and AI toolkit are provided.
AWS Lambda Managed Instances extends the maximum function execution time from 15 minutes to 90 minutes, suitable for long-running workloads such as media processing, financial calculations, data ETL, AI inference, web scraping, and large file transfers; it also supports EC2 instances powered by Graviton5. Amazon emphasizes the need to pay attention to network connectivity, validity of temporary credentials, retry and idempotency design, and recommends using Powertools for AWS Lambda to implement idempotent operations.
Doubao Work has officially launched the IGA Pages plugin, which allows users to publish locally generated HTML pages as publicly accessible links via natural language commands (such as "deploy for me" or "upgrade to a multi-user co-created itinerary wall"). When needed, it automatically connects to the Volcano Engine Supabase database to support collaborative features like attraction voting, list additions, task claiming, and real-time AA budget aggregation. The plugin must be manually installed within Doubao Work; deployment relies on a Volcano Engine account, and preview links have a default validity period of 3 hours and include a security token.
BlockBooster researcher @BlazingKevin_ analyzed on-chain data from September 13 to October 12, 2026, finding that 130 Web3 projects directly interacting with model tokens had a combined verifiable revenue of approximately $8,700,000, with Venice accounting for $7,200,000; most other projects had monthly revenues under $10,000. The report noted significant inflation in token consumption, an extremely low share of x402 in model payments, generally low utilization rates of quota tokens, and decentralized inference's heavy reliance on OpenRouter traffic diversion and token subsidies.
Bloffy is an independently developed blogging platform by the author. It injects only a 3.2KB script into reader pages, supports custom domains, multilingual publishing, CSS/JS injection, and IndexNow for fast indexing, achieving a score of 100 on PageSpeed Insights (Performance, Accessibility, Best Practices, SEO). The platform uses no cookies, stores no IP addresses, and shares no third-party data, employing edge caching and modern image format optimizations.
An AWS official blog tutorial guides enterprise customers on configuring VPC endpoints to privately access the Amazon Quick console and API, and using VPC connections (ENIs) to enable Quick to securely pull data from on-premises data center databases, Redshift knowledge bases, and internal APIs without traversing the public internet. The tutorial covers prerequisites, endpoint policies, DNS configuration, execution roles, VPC connection creation, external domain handling, enabling access restrictions, identity federation integration, RLS/CLS permission controls, and the full validation testing process.
The AWS Solutions Architecture team released the open-source multi-agent system Deep Insight, which automatically converts CSV data and natural language prompts into DOCX analysis reports with charts and citations.
ZeGrapher v4 has been released as a free, open-source plane mathematical plotting tool. It supports plotting functions, sequences, parametric equations, and CSV data. Features include expression-driven axis ticks (e.g., multiples of π), constant animation with parallel multi-value plotting, built-in interfaces and manuals in 9 languages, export to vector (SVG/PDF) and bitmap (PNG/JPEG, etc.) formats (with dimensions in real centimeters), and saving of work states.
Walmart officially launched its proprietary agent gateway, functioning as a server based on the Model Context Protocol (MCP) that allows vetted AI agents to retrieve inventory information and place orders on behalf of users; this gateway supports Sam's Club and can be used without a Walmart+ membership; users can create independent profiles for each agent, set manual approval requirements or spending limits, and enjoy return policies equivalent to those for human-placed orders; the platform explicitly prohibits agents from accessing payment information or performing unauthorized operations.