Search “OpenAI”

NVIDIA Explains How to Fine-Tune Nemotron 3.5 Lightning with the Nemotron-Personas Dataset

NVIDIA has released the Nemotron-Personas France and Belgium datasets and provided a complete post-training workflow for fine-tuning Nemotron 3.5 Lightning with them. The workflow includes defining behavior contracts, generating synthetic conversations with personas, fine-tuning with LoRA on NVIDIA NeMo, merging adapters, and then comparing and evaluating item by item against the base model on a frozen test set. The France dataset was built in collaboration with Pleias, and the Belgium dataset involved Pleias and KU Leuven; the Nemotron repository provides reusable steps and agent skills for data generation, training, and evaluation.

Gemini 4 Carbon Tops the Artificial Analysis Intelligence Index v5

Google's Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5 with a score of 59, leading in four of the 9 evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. On the Coding Agent Index, Gemini 4 Carbon (high) paired with the Antigravity CLI scores 59, trailing Claude Code paired with Claude Opus 5.5 (max, 62 points) by 3 points. The model has a 1M-token context window, supports text, image, video, and speech input, and is priced at a standard rate of $4/$20 per million input/output tokens; it is currently available to select users with a 50% discount.

How to Extract Structured Data from Websites with LLMs (2026)

Browser Use published a tutorial introducing a workflow for extracting structured data across multiple websites using a single JSON Schema: static pages are scraped to Markdown and then filled in by an LLM, while pages requiring login or clicks are handed off to a browser agent. The article provides a reliability checklist, including enabling structured output, validating with Pydantic or Zod, returning null for missing fields, preserving source_url and original-text citations, deduplicating by stable keys, respecting robots.txt, and spot-checking samples. It also compares Firecrawl, Crawl4AI, Browser Use, Stagehand, and a self-built Fetch+Markdown+LLM setup in terms of login and JS workflows, Schema support, open-source status, and billing.

How to get coding agents to use your already-logged-in Chrome

Browser Use published a guide introducing several ways to let coding agents work in a real, already-logged-in Chrome, including Claude in Chrome, Playwright MCP extension mode, Chrome DevTools MCP auto-connect mode, the Gemini CLI built-in browser agent, the ChatGPT desktop browser extension, and Browser Harness. The article gives configuration commands and limitations separately for agents such as Claude Code, Codex, Gemini CLI, and Cursor, and explains that starting with Chrome 136, --remote-debugging-port and --remote-debugging-pipe on the default data directory will be ignored, so the old practice of copying cookies to a new browser will fail in scenarios such as device-bound sessions.

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.

Amazon SageMaker AI Multi-Turn Reinforcement Learning Practical Guide: Training gpt-oss-20b with a Wordle Example

AWS has published a practical guide for multi-turn reinforcement learning (MTRL) on Amazon SageMaker AI. Using Wordle as an example, it deploys the agent and reward function to Amazon Bedrock AgentCore, and the training job updates LoRA adapters via GRPO. In the example, gpt-oss-20b at default reasoning effort improved its solve rate on held-out words from 61% to 84%, at a training token cost of about $339 and an AgentCore runtime cost of about $79. The article also details failures such as two inverted reward design errors, token invalidation under concurrency, and tool calls degrading into plain text, and recommends evaluating under the decoding settings used in final serving.

Browser Use Roundup of the Best Browser Agent MCP Servers of 2026

Browser Use has released its 2026 roundup of browser Agent MCP servers, dividing such servers into two modes—driver and delegate—and evaluating four of them one by one: Browser Use Cloud MCP, Browser Use open-source MCP, Hyperbrowser MCP, and agent-browser chat. In a comparative evaluation across 100 BU Bench V2.1 tasks, the author found that the four driver-type tools consumed a median of 2.1M to 6.7M input tokens per task, all of which count toward the coding Agent's context. The article also provides each server's installation commands, model selection, billing methods, and known limitations, and explains that long tasks, parallel tasks, and structured-output scenarios are better suited to the delegate mode.

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.

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.

Artificial Analysis: Reaching an Intelligence Index score of 30 is now 86 times cheaper than when it was first achieved

Based on v4.3.2 measured results and current pricing, Artificial Analysis found that the per-task cost of reaching an Intelligence Index score of 30 is 86 times lower than when that score was first achieved. GPT-6 Luna (high) reaches 30 at $0.03/task, whereas Claude Sonnet 4.6 cost $2.49/task when it first reached 30 in February 2026. Reaching 20 is about 123 times cheaper than roughly 13 months ago, and reaching 40 is about 33 times cheaper than roughly 5 months ago. Claude Haiku 5.5 (xhigh) reaches 41 at $0.12/task, making it the cheapest model to reach 40. The highest measured score rose from 21 to 58 within a year, at roughly 11 times the cost, with Claude Opus 5.5 at $5.98/task.

English translation is being prepared.

The Qwen3.8-27B reinforcement-learning post-trained model surpasses GPT-6 Luna high on multiple Agent benchmarks

Rohan Paul introduced an open-source environment model built on Qwen3.8-27B with reinforcement-learning post-training, which reduced the hallucination rate on AA-Omniscience from 30.4% to 15.1%. The model surpasses Muse Glimmer 30B and GPT-6 Luna high on Terminal-Bench, τ³-bench, ClawEval, and BFCL v4, but Luna still leads on DeepSWE at 53.1%; its median output for solved tasks is 23.8K tokens, compared with 54.9K for base Qwen. The author also mentions that the MLA version replaces 16 layers of full attention with multi-head latent attention, reducing memory usage per token from 64 KiB to 18 KiB; the compressed version can run on a single 24 GB RTX 4090, with a context about 3.5 times that of base Qwen.

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.

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.

RTX Spark laptop tested: How does Nvidia's first local AI laptop chip perform?

heise online's c't 3003 tested the Asus ProArt P14, which is equipped with Nvidia's first laptop chip, the N1X (brand name RTX Spark). The machine supports up to 128 GB of unified memory; the test unit had 32 GB of memory plus a 1 TB SSD and cost 3,300 euros. In llama-bench tests with Qwen 3.8-27B, generation speed was about 14.5 tokens/s, slightly faster than Apple M6 and Strix Halo but slower than M5 Pro; with the MoE model Qwen 3.6-35B-A3B, it reached nearly 61 tokens/s. Prompt processing speed was better than Strix Halo and the M5 Pro and M6, but fan noise under full load reached up to 2.7 sone, and the fans also spin up when idle. The author also tried installing Ubuntu; after manually adding the drivers, both models ran 14% and 9% faster under Linux than under Windows.

English translation is being prepared.

English translation is being prepared.