English text is being prepared.

OpenAI discloses offline training environment for frontier models; two isolated AIs communicate via CPU cache

On October 23, OpenAI disclosed the training environment for its frontier models. When practicing searching, browsing, and operating web pages, the models use pre-saved and indexed web resources and do not directly access the real internet; all of its frontier models' reinforcement learning training uses this kind of offline interaction. The environment employs multiple layers of isolation, including restricted containers, lightweight virtual machines, hypervisors, firewalls and identity and permission management, and monitoring and suspension mechanisms, but OpenAI still retains some external connections for collecting metrics, logging, and managing authentication. In experiments specifically testing the effectiveness of the isolation, OpenAI had two AIs attempt to pass messages to each other through the CPU cache; in most tests, researchers had given hints about the communication direction or encoding rules; OpenAI said it has patched the issues found and added monitoring for these communication methods.

Kimi K3.1 Pro (Max) ranks 11th on Agent Arena, up 6 percentage points from Kimi K3

Kimi K3.1 Pro (Max) ranks 11th in Arena's Agent Arena, achieving a +7.6% net improvement based on 4K+ real agent sessions, up 6 percentage points from Kimi K3 (Max)'s 25th place. Its median cost per task is $1.03, 13% lower than the previous version, ranking 8th, 9th, and 10th in the Work, Code, and Chat categories respectively. Arena stated that the model's open weights will be released soon, at which point it will become the top-ranked open-source model in Agent Arena, surpassing DeepSeek-v4.1-Flash.

Sebastian Raschka releases an 11-hour video course on building a reasoning model from scratch

Sebastian Raschka has released 8 programming videos on building a reasoning model from scratch, totaling 11 hours, using Python and PyTorch to implement evaluation, inference-time scaling, reinforcement learning, and distillation starting from a pretrained Qwen3 0.6B model. The course first builds a math answer verifier and evaluates on MATH-500, then implements temperature sampling, top-p filtering, and self-consistency, followed by RLVR training with a simplified version of GRPO, and finally distills a small model using answers generated by DeepSeek R1. The content accompanies his book "Build a Reasoning Model (From Scratch)"; the early text generation and evaluation examples can run on a CPU, while later training experiments are recommended to use a GPU.

DeepSeek V4.1 Flash 1017 Tops the Artificial Analysis Cybersecurity Index Open Model Leaderboard

DeepSeek V4.1 Flash 1017 ranks first among openly available models on the Artificial Analysis Cyber Index with a score of 57%, and second on the overall leaderboard, 12 points behind GPT-6 Sol (69%), which is offered through a trusted access program. The index consists of three benchmarks: CWE-Bench-AA, DeepsecBench-AA, and CyberGym-E2E-AA, and mainly evaluates agentic cybersecurity defense tasks. Compared with the previous generation DeepSeek V4.1 Flash's 41%, the new version improves by 16 points, mainly from CyberGym-E2E-AA rising from 23% to 76%; its cost per task is $0.42, placing it on the capability-cost Pareto frontier.

SemiAnalysis benchmarks Trainium3's PyTorch training and Neuron Explorer profiler

The SemiAnalysis engineering team benchmarked the native PyTorch training stack TorchNeuron and TorchTitan on Trainium3, training Qwen3-32B on 16 Trainium2 chips; eager mode reached 11,099 tokens/s and 23.89% MFU, while compiled mode reached 14,156 tokens/s and 30.47% MFU. The author considers the Neuron Explorer profiler superior to Nvidia's and AMD's solutions, but first-time compilation takes too long: the first-step compilation for the gpt-oss training recipe took 27 minutes 17 seconds. FlexAttention tests passed 89 of 96 cases, with all 7 failures due to compilation timeouts rather than numerical errors.

Anthropic shares when to fold subagents into the main agent loop

In comparisons across multiple enterprise deployments, Anthropic's Applied AI team found that a single agent with skills often outperforms subagent architectures in quality, with lower per-task cost and latency, and presents a migration method for folding subagents into the main loop. The article notes that subagents remain suitable for scenarios where context bloat degrades performance, independent tasks need to be processed in parallel, or tool selection conflicts with instructions, and recommends first treating subagents as tool calls, then converting them one by one into skills and MCP tools, using end-to-end evaluations to compare performance before and after. The article mentions that Claude Code's system prompt for more advanced models has been reduced by over 80%, accompanied by Claude API skills such as build-eval, prompt-audit, hillclimb, and cost-optimize.

Who Is On Call After an AI Agent Goes Live: Nine Engineering Steps for Production Ownership

The New Stack published an article discussing the ownership problem for production-environment AI agents, pointing out that teams often focus their energy on prompts, retrieval, tools, and model selection while overlooking who is responsible for the system. The article lays out nine actionable steps, including defining an owner and escalation channel for the agent, using a tool allowlist that denies unauthorized operations by default, setting budgets for execution steps and tool calls, logging model calls, distinguishing failures from escalations, requiring human approval for high-risk operations, providing a health check endpoint, blocking by default when security controls fail, and creating a degraded-operation runbook for middle-of-the-night incidents.

Fireworks Offers Free Training of Muse Glimmer 30B: Using SFT to Predict Kalshi Market Outcomes

Fireworks AI announced that Muse Glimmer 30B can be trained with SFT for free until November 30, and released a complete recipe for a prediction experiment using historical Kalshi market data. The fine-tuned Glimmer achieved 58.3% accuracy on unseen topics, 11.2 percentage points higher than Jev's 47.1%; in a historical backtest, a hypothetical $1,000 grew to $1,071 after accounting for simulated fees, while both base Glimmer and Jev lost money. Training used LoRA adapters and a K-way cross-entropy loss, ran through the training data only once, and separately used a calibration set to fit a temperature parameter for each question category.

MIT Course MAS.S62: Two Lectures on the Evolution of Harnesses and Multi-Agent Systems

Lectures 4-5 of the MIT course MAS.S62 "How to AI an AI" focus on Harness evolution, covering the evolution of harness components, memory, tools, and skills, as well as the evolution of planning, task decomposition, and multi-agent systems. The memory section uses a three-axis taxonomy to connect MemGPT, mem0, MEM1, and Reflexion/ExpeL, and uses CLM to unify context management as file operations; the tools section moves from Toolformer, ToolLLM, and CodeAct to Meta-Harness, which represents the entire harness as a file system and provides about 400 times the amount of feedback information of previous methods. The multi-agent section presents five empirical findings, including that MAS wins 81% on decomposable tasks and loses 70% on tasks with sequential dependencies, and that CORAL achieves new SOTA on 8 tasks, with multi-agent evolution improving over single-agent by 4%-20%.

English translation is being prepared.

9to5Mac reviews the iPhone Duo: the first iPhone in a foldable form factor

9to5Mac editor Chance Miller reviewed Apple's first foldable iPhone Duo after a week of hands-on testing, arguing that this form factor represents the future of the iPhone. When unfolded, the iPhone Duo has a 7.6-inch inner screen; when folded, the outer screen is 5.4 inches. It weighs 254 grams, is 11.3mm thick when folded, is powered by the A20 Pro chip, and costs $1,999. The review notes that it replaces Face ID with Touch ID, lacks a telephoto lens, and that iOS 27 moves navigation controls to the right side and supports Split View on an iPhone for the first time.

Developers use AI reverse engineering to run iOS 16 on Android phones such as Redmi

Developer dipa-studio successfully booted and ran iOS 16.7.16 on a Redmi Note 10 Pro, able to enter the home screen, operate Settings and Control Center, and subsequently similar demonstrations appeared on devices such as the Redmi Note 12, Nothing Phone, and OnePlus 12. The porting process extensively used Claude and Codex to write driver adaptation code, and a reverse-engineering tool REA that automatically analyzes hardware interfaces also emerged. In the current version, cellular network, Bluetooth, camera, and speaker are all unavailable; GPU hardware acceleration has not been made to work, causing lag; booting takes 4 minutes; code signing verification and sandbox restrictions have also been relaxed; it is only suitable for early-adopter demonstrations.

English translation is being prepared.

i14 Trains and Deploys the Laplace 3B Linear Attention Model on AWS Trainium2 Using Native PyTorch

i14 partnered with AWS to train and deploy Laplace 3B, based on Gated Linear Attention, on AWS Trainium2 using TorchNeuron Native. A single trn2.48xlarge achieves a training throughput of about 120,000 tokens/s. The model's recurrent state per sequence is fixed at 29.5 MB and does not grow with context length; inference reaches 11,544 output tokens/s at 1,280 concurrent sequences. The team also reworked Muon's all_to_all communication to cut per-GPU communication volume by 24x and raise training throughput by 1.63x.

Simon Willison built Battle Drones, an audio-networked RTS game, with Claude Code

Simon Willison released Battle Drones, a real-time strategy game developed entirely through vibe coding with Claude Code. It went through 12 major versions and 321 commits and runs in mobile and desktop browsers. The game includes 9 building types, 7 ground vehicles, 6 air vehicles and 3 infantry units, and supports fog of war, random map generation, AI versus AI, and an Agent battle mode in which two LLMs fight each other. He also implemented a multiplayer method in which two devices synchronize game state via microphone audio, and because the audio caused headaches, he added online multiplayer based on a WebSockets relay server.

Google Developers Blog publishes ADK voice agent tutorial: using streaming tools to eliminate silence

Google Developers Blog published a step-by-step tutorial on how to build a voice agent with ADK and gemini-3.8-live that does not stay silent for long periods. The core approach is to change tool functions from return to yield to make them streaming tools, or to use FunctionTool's response_scheduling to deliver results when idle; in the sales report benchmark flow, the agent's time to first speech dropped from 14.4 seconds to 1.2 seconds. The tutorial also covers the stop_streaming cancellation mechanism, preventing duplicate calls, and the trade-offs among the three scheduling modes WHEN_IDLE, INTERRUPT, and SILENT. It requires google-adk >= 2.10.0 and uses run_live bidirectional sessions.

How to run n8n AI workflows on NVIDIA DGX Spark without calling cloud APIs

n8n Blog published a tutorial on how to deploy a complete model stack locally on a single NVIDIA DGX Spark, enabling n8n workflows to perform transcription, OCR, document Q&A, and image generation without calling cloud APIs. The entire stack exposes an OpenAI-compatible endpoint through LiteLLM; the main model is the NVFP4 version of Qwen3.6-35B-A3B, with a single-user generation speed of 120 tokens/s and an aggregate of 359 tokens/s across eight users. The author's testing found that image generation reduces chat speed to about one-third, and recommends that small teams queue image tasks or schedule them outside working hours.

A Guide to "Agent Engineering in Practice: Eight Lectures": Engineering Methods from Working Prototype to Production-Ready

The Tencent Cloud Developer Community has published a guide to "Agent Engineering in Practice: Eight Lectures," centered on how to upgrade a prototype capable of calling models and tools into a system that can complete business tasks, explain failures, control costs, and iterate continuously. The series runs through two fictional teaching cases—a case-filing assistant Agent and a postoperative follow-up Agent—covering eight topics: task contracts, reasoning and collaboration, context and memory, reliable tool execution, observability and replay, evaluation and calibration, security boundaries, and production operations. The article emphasizes that a model being able to suggest the next step does not mean the system can deliver business results; the example fields, budgets, and figures are all teaching calculations and do not constitute legal advice.

What is a self-replicating AI worm, and why is Anthropic worried?

The Guardian interprets a report released this week by Anthropic's offensive research team: the team says it has built a malicious AI program that can spread autonomously among internet-connected smart TVs, splitting itself across 15 TVs and borrowing a little computing power from each to keep running. The researchers believe a successful AI worm would need stronger capabilities on lower-powered devices, but that current trends make this potentially feasible by the end of the year; the worm in an earlier University of Toronto paper instead infected a full-sized computer and could spread faster. Security experts disagree: some stress that IoT devices such as smart TVs are slow to receive patch updates and their risks are hard to fix, while others argue the experiment moves slowly and its actions are predictable and interceptable, making it stunt hacking.

English translation is being prepared.