All news

UUnite.AI09:26Analysis

AI score58

Perplexity Publishes LLM Gateway Best Practices Guide, Outlining Request Flow and Production Risks

On October 27, 2026, Perplexity published an LLM gateway best practices guide explaining how this middleware layer routes, authenticates, caches, and logs model calls, and distinguishing four categories: API gateways, LLM gateways, MCP gateways, and AI gateways. The guide breaks down the mechanism through a six-step request lifecycle, while warning that the gateway centralizes credentials from various providers, and cites the LiteLLM gateway breach disclosed by Microsoft and two high-severity vulnerabilities, CVE-2026-35029 and CVE-2026-59823, reported by Check Point (already fixed in LiteLLM 1.83.9). The article also lists postmortems of Cloudflare AI Gateway and OpenRouter outages, four deployment paths, and paid plans starting at about $50 per month.

#Safety / alignment#English text is being prepared.#OpenAI

LLeikeji09:26Analysis

AI score52

Rokid Glasses 2 First-Hand Experience: More Comfortable to Wear, AI and Battery Life Upgraded

Leikeji experienced the Rokid Glasses 2, released on October 27, in advance. It starts at 3,699 yuan (3,199 yuan after national subsidy), and officially goes on sale on November 11. Compared with the previous generation, the new model's temples use a more skin-friendly material to distribute pressure, the battery capacity has increased to 263mAh, up 25%, and it is equipped with a 12-megapixel camera supporting 4K 30fps and 2K 60fps video recording. In the author's test, after 20 minutes of navigation, 30 minutes of music, and 30 minutes of photo and video recording, more than half of the battery remained, and magnetic charging can reach 75% in 20 minutes.

Same story · 7 reports

#Product updates#English text is being prepared.#Evaluations / benchmarks

NNVIDIA Developer · Generative AI09:24First-partyAnalysis

AI score52

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.

#Tutorials / practice#English text is being prepared.#Open-source ecosystem

LLangChain Blog09:22First-partyAnalysis

AI score67

How LangChain Built the Design Agent Content Studio

LangChain's GTM engineering team, working with the design team, built Content Studio, a design agent inside Slack, on Managed Deep Agents. It can generate videos, social graphics, UI mockups, and slide decks, and has already produced more than 3,000 assets and over 80 slide decks. The article breaks down its four-layer knowledge structure (skills, wiki, code validation, memory), the dedicated LangSmith Sandbox for each Slack thread, the two rendering engines HyperFrames and the in-house motion-studio, and how the main agent orchestrates asynchronous sub-agents. The team says that after migrating to Managed Deep Agents, its own code shrank by about 80%, the median model cost per request fell from about $8 in mid-September to about 50 cents, and the entire harness has been open-sourced.

Same story · 2 reports

#Tutorials / practice#Agents#English text is being prepared.

FFull Artificial Analysis article08:00First-partyAnalysis

AI score82

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.

Same story · 34 reports

#Evaluations / benchmarks#Google#Reasoning

KKarl's AI Watts21:51Analysis

AI score58

The author's hands-on record of connecting 15 mainstream apps to Ant Lingguang 2.0

After getting access to the Ant Lingguang 2.0 beta, the author connected 15 domestic apps—including Mijia, Baidu Netdisk, DingTalk, Didi, Luckin, McDonald's, Amap, Fliggy, Alipay, and Taobao—to this Personal Agent through methods such as QR code authorization, MCP, Skill, GitHub projects, and CLI. The article explains the integration method for each app one by one according to configuration difficulty, and provides a reusable configuration prompt that requires connecting only one service at a time, verifying with real queries after connecting, and confirming before executing any action that involves spending money. The author believes that the mobile Personal Agent differs in form from the desktop version, with its advantage lying in its ability to connect to domestic services, and that it will arrange scheduled tasks on its own to continuously follow up in the background.

#Tutorials / practice#Agents#English text is being prepared.

AAlibaba Cloud Developer Community AI19:01Analysis

AI score60

GEO Practical Guide: AI Health Self-Check, Content Structuring Overhaul, and Building Trustworthy Signals Across the Web

The Alibaba Cloud Developer Community has released a GEO practical guide, offering a complete implementation path from AI health self-check to content overhaul, trustworthy signal building, and monthly monitoring. The article recommends selecting 3–5 mainstream AI tools (such as ChatGPT, ERNIE Bot, Doubao, Kimi, and Tencent Yuanbao) and using standardized questions to test brand mentions and recommendation status as an optimization baseline. The content overhaul section proposes four types of approaches—structuring, factualization, Q&A formatting, and de-marketing—and lists five common misconceptions, including treating GEO as SEO and optimizing only the official website.

#Tutorials / practice#Search#English text is being prepared.

AAlibaba Cloud Developer Community AI19:01Analysis

AI score53

Hands-on test of Alibaba Cloud Bailian's built-in web search and web scraping: enable with a one-line tools declaration

The author tested on Alibaba Cloud Bailian 2.5.0. After declaring web_search via the tools parameter, the model went from refusing to answer "today's weather in Shanghai" in 9.2 seconds to returning the day's actual conditions in 24.5 seconds with the correct date. Adding web_extractor lets the model decide the number of retrieval rounds and scraping targets on its own; in the JSON output, the counts of two web_search calls and three web_extractor calls match the per-call records. The author also recorded web-enabled call latency ranging from 17 to 127 seconds and 22,789 input tokens in a single call, and notes that the agent and agent_max strategies other than turbo incur an additional charge per call.

#Tutorials / practice#Agents#English text is being prepared.

SSaito Hardcore Hacker18:15Analysis

AI score68

npm founder Laurie Voss on how to evaluate whether an AI agent is up to standard

Laurie Voss, co-founder of npm and now head of developer relations at Arize AI, presented a method at the AI Engineer New York workshop for judging whether an AI agent is up to standard: read through a dozen or so trace records before writing the first evaluation, and evaluate outcomes rather than steps. He used a research report that added two search results together into a 116% analyst consensus as an example to show that only by looking at the records can you pinpoint which step the reasoning went wrong at; after switching from a correctness evaluation to a faithfulness evaluation, the same batch of 16 reports went from all being judged wrong to 14 being judged unfaithful with usable explanations. He also required testing the judge itself with a manually labeled golden dataset, and writing the stock ticker correctness as a floor and operability at an average of 0.8 over three runs as a target into npm test.

#Tutorials / practice#Agents#Evaluations / benchmarks

SShao Meng18:13

AI score69

Open-source meta-skill skill-diet uses evaluation to slim down Agent Skills

The open-source meta-skill skill-diet is used to compress bloated SKILL.md files and check whether tools are outdated. The author @nicopreme measured on real skills that it can reduce tokens by 70–80% with unchanged results. A slimmed-down version is kept only if it is at least on par in same-task evaluations against the old skill, the new skill, and no skill; every subsequent addition must also pass the threshold, and a roughly 10% increase over the previous diet triggers another round. It also places information in layers by skill description, SKILL.md, references/, and tool help and error messages, and recommends leaving mechanical errors to error messages, tests, or linters.

Quoted postNico Bailon · @nicopreme

I've been dogfooding "skill-diet", a skill for slimming down agent skills and keeping them correct. most skills load far more text than a task needs, and some parts often go stale. it also gets worse over time as AI updates skills. skill-diet cuts what the model already does without being told, moves rarely needed detail into files it only reads when needed, and fixes what no longer matches the tool. then it runs the same tasks with the old skill, the new one and no skill, and keeps the new one only if agents do at least as well. on my own skills: 70–80% fewer tokens, same results. https://t.co/h1BnVt7NjD

#English text is being prepared.#Agents#Evaluations / benchmarks

WWired AI18:02Analysis

AI score61

WIRED tested six always-on AI agents for a day, and OpenAI Dots had the highest overall score

WIRED reporter Reece Rogers spent a day testing six always-on AI agents on an iPhone—Meta Muse, OpenAI Dots, Instinct, Anthropic Claude Duet, Google Gemini Spark, and xAI Grok Bot—scoring them from 1 to 5 across five tests: a morning news briefing, ordering lunch, a calendar conflict reminder, a travel itinerary, and a late-night task. OpenAI Dots performed best overall. Dots costs $100 per month, making it the most expensive in this test, and the author suggests most people wait for a price cut or a free version before trying it. The free and easy-to-use Meta Muse came right behind in total score, Anthropic Claude Duet was rated as having the best interface, and Instinct was rated as the most annoying.

#Evaluations / benchmarks#Agents#OpenAI

MMIT Tech Review18:00Analysis

AI score51

How Solid-State Transformers Can Power AI's Electricity Demand

MIT Tech Review explains how solid-state transformers use power semiconductors combined with magnetic components to replace the traditional copper-wire-and-iron-core approach, shrinking volume to about one-tenth while keeping capacity unchanged. A 1-megawatt prototype from Srdjan Lukic's team at North Carolina State University achieved about 98% efficiency in EPRI testing, roughly half the losses of the traditional transformer-plus-rectifier combination at about 96%. Such devices are still in the development stage; in the near term they are mainly aimed at powering data center racks, and adapting them to transmission-level high voltages will require further engineering development.

#Industry news#English text is being prepared.#Tutorials / practice

MModelBest OpenBMB16:42Analysis

AI score60

ModelBest MiniCPM-V 4.7 inference speed on iPhone increased from 3.38 to 20.51 tokens/s

ModelBest OpenBMB reposted a breakdown of inference optimization for MiniCPM-V 4.7 35B-A3B on the iPhone 18 Pro Max (12GB RAM), with decoding speed increasing from 3.38 tokens/s to 20.51 tokens/s, about 6.07 times the unoptimized baseline. The optimization path was, in order, expert prefetching to 6.57, scheduling and loading optimization to 11.34, mixed-precision quantization to 17.16, and then adding DSpark speculative decoding to reach 20.51 tokens/s; the model has 35B total parameters and about 3B active parameters, supports image, multi-image, and video understanding, and keeps most weights in flash storage for on-demand loading. The original post notes that these are still OpenBMB's own measurements.

Quoted postDavid Hendrickson · @TeksEdge

This new-ish 35B multimodal model went from 3.38 → 20.51 tps on an iPhone. MiniCPM-V 4.7 35B-A3B was updated to decode faster on the iPhone by improving how it moves the model between storage and memory. Constraints 🧠 35B total / ~3B active 📱 iPhone 18 Pro Max — 12GB RAM 👁️ image + multi-image + video understanding Most of those 35B weights don't sit in RAM. The experts can stay in the phone's flash storage and get loaded when they're actually needed. Normally that's slow 🐌 SSD expert loading → 3.38 t/s So OpenBMB built Lookahead Expert Scheduling (LES) It predicts which experts the next layer will probably need and starts loading them while the current layer is still computing. Then they stacked the optimizations: 💾 baseline SSD loading → 3.38 tps 👀 expert lookahead → 6.57 ⚙️ optimized scheduling/loading → 11.34 🗜️ mixed-precision quantization → 17.16 🚀 + DSpark speculative decoding → 20.51 tps That's a 6.07× speedup over the same-phone unoptimized baseline. ⚠️ And these are still OpenBMB's own measurements. Link in ALT

#Industry news#English text is being prepared.#Reasoning

TTencent Cloud Developer Community15:44Analysis

AI score55

Building a QQ Chat Companion Bot That Remembers Things and Proactively Reaches Out with Warm Isle and DeepSeek

The author describes building a QQ chat companion bot on their own server using the open-source project Warm Isle, with DeepSeek generating replies and NapCat handling QQ login, while persona, long-term memory, and proactive messaging are implemented by Warm Isle. The article provides complete steps for Docker deployment, QR code login, persona configuration, and memory extraction, and lists practical pitfalls such as memory usage, SQLite backups, and group chats not being processed by default. The project uses the AGPL-3.0 license, and commercial use is prohibited for the NapCat and SnowLuma protocol ends.

#Tutorials / practice#Agents#Open-source ecosystem

SShao Meng12:59

AI score60

Uncle Bob open-sources ub-mode Skills, setting three disciplines for coding agents

Uncle Bob, author of Clean Code, has open-sourced Skills "ub-mode", organizing Clean Code and TDD discipline into a three-stage workflow for coding agents: first make the change run, then make it right using CRAP, structural duplication, and mutation testing, and finally wrap up with logging, small commits, and context hygiene. The second stage comes with tool gates such as crapper, dryer, and mutator; for example, CRAP must be below 10, and structural duplication reaching 0.8 must be eliminated. There is also uml-viewer for viewing components, coverage, CRAP, mutation scores, and dependency direction.

#English text is being prepared.#Agents#English text is being prepared.

SShao Meng12:59Analysis

AI score57

Stanford CS 329Z Lecture 10 covers Agent evaluation engineering with Terminal-Bench and Harbor

In Lecture 10 of the Stanford CS 329Z course, Alex Shaw and Ryan Marten explain how to use Terminal-Bench and Harbor to turn Agent evaluation into an iterative engineering process. Terminal-Bench builds a community benchmark from terminal tasks; a task directory consists of instruction.md, task.toml, environment/Dockerfile, solution/solve.sh, and tests/test.sh, and the verifier determines environment state through programmatic checks and an LLM judge. The lecture notes also present practices for isolating the agent and verifier environments to prevent reward hacking, and for distinguishing reruns from regrades by major/minor patch versions.

Quoted postmeng shao · @shao__meng

English text is being prepared.

#Tutorials / practice#Agents#Evaluations / benchmarks

BBrowser Use12:57First-partyAnalysis

AI score52

Browser Use Tutorial: How to Log In Daily and Download Invoices on Websites Without an API

Browser Use has published a tutorial introducing three approaches for automatically logging in and downloading invoices daily on supplier portals that lack an API: Playwright scripts, RPA tools such as UiPath or Power Automate, and browser agents with saved configurations. The article provides complete code for Playwright and Browser Use Cloud, covering reusing login sessions, filling in TOTP verification codes with 1Password, scheduling runs via GitHub Actions, and handling failure scenarios such as session expiration, 2FA blocks, page redesigns, and empty downloads. The author recommends first asking suppliers for exports, APIs, or emailed invoices, and reminds readers to first confirm whether the portal's terms permit automated access.

#Tutorials / practice#Agents#English text is being prepared.

PPhoenix Technology12:56

AI score55

User Shares Claude Code Check-In Safety Prompt: Locates Hidden Cameras in 90 Seconds

A user shared a set of Claude Code prompts that can scan for hidden cameras, Bluetooth trackers, and suspicious phone programs before checking into a hotel or Airbnb. According to actual test feedback from a user on Platform X, the solution located a pinhole camera behind a bathroom mirror within 90 seconds. The solution sets Claude Code as a read-only check-in safety assistant, prohibiting account login, use of default passwords, and initiation of vulnerability exploitation. The process includes using arp-scan and nmap for Wi-Fi scanning, using a bleak-based script to listen for Bluetooth tracker signatures, using Amnesty International's Mobile Verification Toolkit to audit the phone, and finally outputting a ranked table of suspicious findings and follow-up recommendations. The original post also points out limitations: network scanning can only discover cameras that are on the connected network; devices that record to an SD card, create their own hotspot, or are outside the isolated network will not be discovered.

#Tutorials / practice#Agents#Anthropic

BBrowser Use07:07First-partyAnalysis

AI score55

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.

#Tutorials / practice#Agents#English text is being prepared.

AAWS Machine Learning01:57First-partyAnalysis

AI score58

AWS explains prompt caching patterns for using OpenAI models on Amazon Bedrock

AWS has published a tutorial explaining how to use prompt caching to reduce the cost of agents repeatedly reading context when calling OpenAI models on Amazon Bedrock, distinguishing between two approaches—implicit caching and explicit prompt_cache_breakpoint—and recommending their combined use. The article provides code examples for tool and instruction caching, rolling breakpoints across multi-turn history, adding tools mid-session, and shared prefixes across sub-agents, and states that in a six-turn workload on GPT-6 Astra, implicit caching reduces input costs by about 62%, while explicit rolling breakpoints reduce them by about 61%. Cacheable prefixes must be at least 1024 tokens, and cache writes are billed at 1.25 times the normal input price.

#Tutorials / practice#Agents#English text is being prepared.