The AWS Machine Learning blog published a tutorial introducing the Adaptive Agent Framework (AAIF) built using Strands Agents. This framework uses the llama3.2:1b small language model (SLM) running locally as an intelligent scheduling layer, routing routine requests (such as device telemetry and camera image analysis) to the local SLM hosted by Ollama, while complex requests invoke large models in Amazon Bedrock cloud. The tutorial includes complete Python code, tool definitions, configuration methods for three deployment modes (pure cloud, hybrid, and fully offline), as well as explanations of MCP server extension schemes and Bedrock Guardrails security integration.
At the Q3 earnings call, Meta confirmed that its AI agent Muse has over 10,000,000 users, with 25% of new users using it daily; more than a quarter of active shopping users have an average order value exceeding $200. Plaid stated that activity intensity on its connectors within Muse is at an all-time high. Muse has enabled Plaid for agentic shopping and supports productivity scenarios such as email/calendar, ensuring privacy via Muse Secure VM and the upcoming Confidential VM. Meta plans to achieve profitability through platform commissions, premium model subscriptions, and token fees, expecting compute costs to be optimized by more than 10x within one year, targeting per-user revenue exceeding current levels of Facebook/Instagram/WhatsApp/Messenger (hundreds of dollars/year). Muse will be integrated into Meta's smart glasses strategy and expanded into a 'Muse Network' supporting cross-user collaboration.
Bits is a personal agent tool that runs on the user's local machine, operating continuously as a background daemon. It supports scheduled wake-ups, task continuation after application closure, and automatic recovery after crashes; memory storage consists of readable and editable Markdown files; high-risk operations require user confirmation via text through the Sentinel review mechanism; it supports voice interaction, shortcut invocation (⌥Space), and any model (including local models), with output formats including real-time cards and mini-apps.
QAmcp is an open-source terminal testing tool used to automatically detect the compliance, security, and availability of MCP servers. It connects to servers from the perspective of AI applications, outputting clear pass/fail results, millisecond-level latency, and precise matching assertions, without relying on AI judgment or uploading any data. It supports HTTP/stdio modes, --demo demonstrations, CI integration (including GitHub Actions examples), Docker, and programmatic calls. The current version is 0.1.0, licensed under MIT.
This article demonstrates how to train Qwen3 on AWS Trainium accelerators using torchtitan on a single trn2.48xlarge instance: the first method uses Transformers v5 as the modeling backend, loading models from Hugging Face Hub via AutoModelForCausalLM and directly reading/writing safetensors; the second switches to torchtitan's native Qwen3 implementation, specifying the 4B configuration and SDPA via config.model_spec; the third integrates MoE kernels written with the Neuron Kernel Interface (NKI) and shows how to register custom NKI rotary embedding operators.
OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon participated in discovering a critical PHP vulnerability affecting millions of websites globally. The vulnerability was independently verified by Wiz Security and coordinated over two months through five stages with PHP developers, Cloudflare, the US CISA, and the White House. Wiz stated that Gemini 4 Argon was particularly effective in exploit development, attack detection construction, and patch verification, while also noting that large language models significantly lower the barrier for attackers to adapt to different PHP deployments.
Renor is a self-hosted memory engine for coding agents that shares private memories based on local Markdown files across tools such as Claude Code, Codex, and Cursor via the MCP protocol.
The author released the open-source project Reactive-LEDS, which includes ESP32-S3 firmware, CLI tools, a browser JavaScript client, WebSocket/UDP control protocols, and accompanying 3D printing models. It supports low-latency real-time control of 1-meter FCOB LED tubes; currently, the firmware can render frame-by-frame at 90–120 fps streams, whereas WLED merges updates to approximately 62 fps under equivalent conditions; the project uses dual licenses: GPL-3.0-only (firmware/CLI) and MIT (client/shared modules).
onichandame released the open-source tool den for fast semantic search of documents in the docs/ directory within projects, aiming to address issues of high maintenance costs and outdated documentation; project URL: https://github.com/onichandame/den.
OpenAI President Greg Brockman stated that the company, together with multiple parties, led the discovery and mitigation of the PHP2Shell vulnerability within two months. The vulnerability had existed in PHP for 14 years, causing government websites in Finland, the Netherlands, and France to go offline on Wednesday. Calif Security Company discovered partial versions using GPT-5.6 Sol in August; OpenAI discovered a more complete version in September; Wiz Company independently discovered two vulnerabilities. PHP released a fix on October 21, earlier than the originally planned date of October 22, due to premature leakage of details. Brockman proposed that authorized defenders (such as governments) should have remote scanning and patching capabilities, and announced an investment of $1 billion to subsidize the Daybreak project.
Google Quantum AI has launched a verifiable quantum circuit compilation competition on Kaggle, requiring participants to use AI to build smaller and more efficient kickmix circuits; the competition is based on the newly open-sourced C++20/Python library kickmix (integrated into Qualtran), accompanied by the published paper "Leveraging verifiable structure in quantum circuits to find more efficient compilations with agentic AI"; it sets up three major benchmark tasks in fusion energy, optimization, and chemistry, each providing reference circuits and Apache 2.0 open-source validators, with the goal of minimizing the number of qubits and gates under the premise of passing verification.
NVIDIA released Kumo Relational, a pre-trained relational foundation model that can perform predictions (such as churn, demand, and fraud) directly on enterprise databases without task fine-tuning, feature engineering, or table flattening; given a prediction query, it reads up to 10,000 historical examples as context and outputs results in a single forward pass; across 41 benchmarks (including RelBench and SAP SALT), it outperforms expert-tuned gradient boosting trees and AutoGluon ensemble models, leading by up to 16% in MRR; it supports deployment via the open-source structured-data-models library or NVIDIA NIM, with model weights hosted on Hugging Face, code under the Apache 2.0 license, and model weights under the OpenMDW License v1.1.
The U.S. National Security Agency (NSA) warned that PHP has two critical vulnerabilities affecting all versions since PHP 5 (released in 2004), which could lead to compromise of server file systems or databases; the PHP Foundation has released patches and recommends upgrading to PHP 8.5.12.
Google Cloud announced the launch of three OpenAI models: GPT-6 Astra (for high-demand tasks such as coding and research), GPT-6.1 Sol (similar performance but lower cost), and GPT-6 Luna (a fast, low-cost option for everyday use). Currently available only in private preview for enterprise customers, they support integration with Gemini for Business and Google Antigravity. Costs are consolidated into Google Cloud billing and existing contracts, and customer data is not used for training.
Dharma-AI previewed the upcoming VocabX paper on the Hugging Face blog, proposing a vocabulary specialization method that does not modify the base model but only rebuilds the tokenizer: it optimizes based on the actual inference cost of generated tokens (rather than frequency), jointly compressing at the subword and multi-word macro levels; in a Portuguese essay auto-grading system based on Qwen3-4B, the vocabulary size was reduced from approximately 152,000 to approximately 65,000, the average response token count dropped from 350 to 101, measured inference cost decreased by 77.0%, and benchmark scores (grading accuracy + off-topic screening) rose from 0.703 to 0.810; this method outperformed baselines such as FVT, MWT, and in-place BPE across five public domains—contracts, medical papers, scientific papers, news, and Python code—and three model families: Qwen3, Llama-3.1, and Gemma-3.
Omnara is an open-source hosted agent platform offering self-hosted (Apache 2.0) or cloud service options. It supports any model compatible with OpenAI/Anthropic APIs, local or sandboxed machines, built-in toolchains, permission controls, and event streams. It provides a complete development interface via REST API, TypeScript SDK, and CLI, with native support for the MCP protocol, and has been listed in the official MCP Registry. The project code is hosted on GitHub.
PHP officially released emergency patches on October 21, 2026, fixing two unauthorized remote code execution vulnerabilities (CVE-2026-109618 and CVE-2026-109619) with a CVSS v4.0 score of 9.2. The root causes were incomplete data initialization during memory exhaustion and use-after-free errors. Sites including the Netherlands NCSC, Finnish government, French CNIL, and Japan's pixiv implemented preventive shutdowns on the same day. The vulnerabilities affect all PHP runtime modes, including mod_php and PHP-FPM. The OpenAI GPT-5.6 Sol model assisted researcher Calif in discovering exploitable paths for the first time, marking the first instance where AI identified such vulnerabilities before attackers did. Over 70% of disclosed version servers run on PHP branches like 5.4–8.1 that have ended official support, receiving only minimal patches. Cloudflare WAF rules took effect nine hours early, and agencies including CISA recommended immediate shutdown for systems unable to upgrade.
Google Research published a paper in JAMA Network Open reporting that its AI model, after fine-tuning with only 120 cases from Chicago, achieved mean absolute errors (MAE) of 4.10 days and 4.27 days for gestational age assessment at two clinically distinct institutions: Northwestern University Feinberg School of Medicine in Chicago and Jacaranda Health in Nairobi. These results reached non-inferiority levels compared to expert sonographer clinical standards (MAE 4.48 days). The model supports Clarius C3 portable probes, maintained stable performance at the Kenya site without local data fine-tuning, and demonstrated robustness against extreme fetal weight values.
Ollama v0.40.8 has been released, officially integrating the open-source Flint model from Recursive, designed specifically for coding agent tasks. It supports understanding code, tracking bugs, implementing features, and adding tests; it possesses reasoning capabilities, multi-turn tool calling, and a 262K-token context window. It is an Apache 2.0 licensed 40B parameter MoE model, activating approximately 2.5B parameters per token. Currently, it only supports macOS and can be run via `ollama run flint`.
The author implemented a fully automated pipeline from generative video motion to skeletal animation of non-humanoid 3D models: using Tripo3D to generate models, TokenRig for automatic rigging and skinning, synthesizing 3,000 dual-view random pose rendering samples in Blender, fine-tuning a small visual model based on Meta's DINOv2 to recognize joint coordinates, then using MiniMax H3 to generate motion videos, extracting features frame-by-frame, and reconstructing them into 3D skeletal animations; the relevant training and inference code has been open-sourced to GrinZero/open-game-studio (monster package).
Hugging Face blog published the first article in the TorchNeuron performance analysis series, detailing how to perform system-level inference performance profiling on AWS Trainium chips using torch.profiler (in conjunction with NeuronConfig and NeuronProfiler). It compares the Neuron device time, call paths, and hardware execution traces of three RMSNorm implementations (naive PyTorch operations, built-in fused torch.rms_norm, and compiled torch.compile). The accompanying Neuron Explorer visualizes the complete lifecycle from framework scheduling to NeuronCore execution.
Recursive Superintelligence has open-sourced Flint, a 40B Mixture-of-Experts (MoE) coding model with 39.8 billion total parameters and approximately 2.5 billion active parameters per token, supporting a context window of up to 512,000 tokens. It achieved scores of 54.9%, 89.7%, and 74.2% on Terminal-Bench 2.1, SWE-Bench Pro v2, and SWE-Bench Verified, respectively; the BF16 version averaged 80.6 points across 6 benchmarks. It reached 131.22 tokens/second on an M4 Pro and 155.82 tokens/second on an RTX Spark. Licensed under Apache 2.0, it is available on Hugging Face and GitHub. The entire training process was driven by the proprietary automated system Eureka Machine, completed in 4 months.
Anthropic has launched the personal AI agent Claude Duet, based on the Claude Opus 5.5 model. It supports proactive discovery of to-do items, calls over 6000 connectors from the Claude Marketplace, runs continuously in the background within an isolated virtual machine, and is currently open for research preview waitlist applications for Pro and Max users in the US.
Meta released its Q3 2026 earnings report, showing revenue of $64.601 billion and net income of $9.860 billion, with advertising accounting for 97.6% of revenue; the AI agent Muse was configured by over 10 million users in the US and Canada within a month and a half of launch, with a daily active usage rate of approximately 25%, where shopping is the most common scenario; capital expenditures reached $34.91 billion, an 80% year-over-year increase, while free cash flow was -$2.120 billion; global headcount dropped to 66,775 employees, a 15% year-over-year decrease (approximately 12,000 people); losses at Reality Labs widened, and AI glasses are being integrated with Muse; the board approved the creation of non-voting Class C common stock.