All news

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.

GGuillermo Rauch01:36

AI score55

Guillermo Rauch reproduces OpenAI's CPU covert communication research, building cpu-chat

Guillermo Rauch built cpu-chat, reproducing OpenAI's research on using CPUs for covert communication. In a networkless Firecracker microVM environment, it achieves communication through three mechanisms—CPU contention, cache eviction timing, and DRAM bandwidth contention—can run across multiple CPUs, and bypasses Firecracker's production recommendations (disable SMT, disable KSM, bind tenants to different cores). He pointed out that this differs from Meltdown/Spectre-type side-channel attacks, which can steal information, whereas this kind of "whisper" technique is harder to carry out in real environments and easier to monitor by sampling hardware counters and large-scale statistical data. He also mentioned that OpenAI and Anthropic models refused to participate in this work, and it was ultimately completed by open-weight models.

Same story · 2 reports

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

AAlibaba Cloud Developer Community AI00:54Analysis

AI score62

What actually happens during a single AI inference: from fixed parameters to the trade-off between latency and throughput

The article explains that training is learning parameters, while inference is using parameters without changing them; a single inference run only performs forward propagation and does not require loss, gradients, or weight updates. The article breaks down the autoregressive generation of large language models, distinguishing the Prefill and Decode phases: the former is compute-intensive and determines time to first token, while the latter is limited by memory bandwidth and determines the inter-token interval. The article also provides an example of estimating KV cache usage, and discusses the benefit boundary of batching for improving throughput, the unbounded growth of waiting time as utilization approaches 100%, and how chat scenarios prioritize reducing TTFT while Agent scenarios focus on end-to-end latency.

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

CcnBeta22:58Analysis

AI score67

DeepMind hurricane forecasting model incorporated into official forecasts by the US National Hurricane Center

Google DeepMind's AI hurricane forecasting model has been incorporated into official forecasts by the US National Hurricane Center since last year, and by the end of the 2025 Atlantic hurricane season it had become one of the best models for predicting storm tracks and intensity. Meteorologists noted that the model judged the track of Hurricane Isaias correctly from the start, but its intensity forecasts were inferior to the National Hurricane Center's blended forecast. Unlike traditional physical models, which rely on real-time atmospheric data and physical equations, AI models identify patterns by analyzing thousands of historical hurricanes, and their internal mechanisms remain a black box to forecasters.

#English text is being prepared.#DeepMind#Google

KKarl's AI Watts22:31

AI score62

Six prompts to remove the AI flavor from content generated by DeepSeek's new model and codex

The author shares 6 prompts for removing the AI flavor when DeepSeek's new model and codex generate code animations and web UI, targeting issues such as repetitive copy, decorative nested frames, color overuse, chaotic font sizes and alignment, information overload on a single screen, and purposeless motion effects. The author says they have recently been working on using DeepSeek to create code animations similar to Claude Opus 5.5, and teases an editing Skill currently in development that can already use DeepSeek to edit ten-minute real-person talking-head videos.

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

AA Yi AI Notes22:00

AI score54

Fatih launches Grok Bot recipe library, packaging 331 community recipes

Fatih, author of Awesome ChatGPT Prompts, has launched a Grok Bot recipe library, which he says packages 331 ready-made community recipes that can be plugged directly into Grok Bot. The author A Yi notes that the library bundles together three parts—system instructions, routine tasks, and skill plugins—covering scenarios such as invoice payment reminders, side-hustle bookkeeping, and reading coaching. He also cautions that 331 is a figure self-reported by the original author, that some one-click installation paths have not yet been made to work, and that when mounting third-party recipes, users need to confirm whether the bot will automatically send messages to outside parties on their behalf.

Quoted postAYi · @AYi_AInotes

English text is being prepared.

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

SShao Meng22:00Analysis

AI score60

Comparing Three Agent Skill Sets: How to Choose Between Pocock, Addy Osmani, and Cursor pstack

Shao Meng relays Rob Shocks' video comparison of the three Agent Skill sets, concluding that no single set works for everything: Pocock focuses on whether you have aligned on what to build before coding, Addy Osmani covers the complete engineering workflow from Define to Ship, and Cursor's official plugin pstack focuses on proving that changes actually run. Addy's set has about 25 skills, including commands such as /spec, /plan, /build, /test, /review, and /ship, plus multiple reviewer personas. Rob recommends using them tiered by change risk; installing the full set consumes context and is prone to accidental triggering.

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

AAlibaba Cloud Developer Community AI21:53Analysis

AI score58

Breaking Down LangChain 1.0 Core Concepts: From Chained Connections to the Agent Runtime

The article outlines the technology stack layering after LangChain 1.0 and LangGraph 1.0 reached GA simultaneously, explaining that create_agent replaces create_react_agent and runs on top of the LangGraph runtime. It breaks down six core concepts including the unified model interface, tools, memory, and middleware, gives the selection boundaries between LangChain and LangGraph, and includes a runnable minimal Agent example based on the DeepSeek model.

#Tutorials / practice#Agents#RAG

HHacker News Trending21:52Analysis

AI score53

Tasks as Code: Moving Task Tickets into the Code Repository as Markdown

A Typemill knowledge base article proposes the Tasks as Code approach: writing tasks and requirements as Markdown files and storing them in a Git repository, so that AI coding agents can read requirements and code in the same place. The article presents a five-step process: describe the task, analyze the requirements, generate an implementation plan, obtain human review and approval, then implement and write back a report. It also recommends splitting code, documentation, and tasks into three repositories, using .gitignore to isolate local clones. The author also mentions that interfaces supporting Git sync, such as Typemill, can let product managers and technical writers participate, while Git still handles version history.

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

HHacker News Trending21:52Analysis

AI score52

How Ubicloud Used vLLM Constrained Decoding to Raise Tool Calling Accuracy from 92% to 100%

When Ubicloud deployed Kimi K2.6 on vLLM 0.28, tool calling schema accuracy was only 92%, meaning 2 out of every 25 calls failed. vLLM 0.31 turned constrained decoding into a server-side switch, using XGrammar's structured tags to mask illegal tokens during generation; after Ubicloud enabled it, accuracy on the K2VV ToolCall benchmark rose to 100%, and it has been deployed to production. The article also explains why vLLM disables this feature by default, including some models' outputs deviating from the schema, grammar conversion bugs, and interaction issues with speculative decoding.

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

TThe New Stack21:52Analysis

AI score60

MiniMax M3.1 Flash vs. M3 hands-on comparison: lower cost, but no improvement in speed or results

The author ran each of three agent tasks once on OpenRouter to compare MiniMax M3.1 Flash with M3. The two models scored the same on the two scorable tasks, and on the long-context task neither produced an answer in the first round; Flash's estimated total cost was $0.15, lower than M3's $0.27, but its total elapsed time was 661.6 seconds, longer than M3's 465.0 seconds. The author concludes that apart from lower cost, these runs show no improvement in Flash's speed or results, and notes that three single-round tests are insufficient to verify the benchmark performance MiniMax claims.

Same story · 19 reports

#Evaluations / benchmarks#Agents#Reasoning

UUnite.AI21:52Analysis

AI score62

How computer use agents see the screen, click, and operate software

Unite.AI published an article explaining how computer use agents perceive software interfaces through screenshots, accessibility trees, and other means, and complete tasks with actions such as clicking, typing, and scrolling, then read the new state to verify results. The article compares these agents with traditional browser automation, APIs, and tool calling, points out failure modes such as visual grounding errors, long-task drift, prompt injection, and credential exposure, and recommends running them under isolated environments, least privilege, and confirmation boundaries for critical operations. The article mentions benchmarks such as OSWorld used to evaluate performance in real computer environments.

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

SShao Meng15:10Analysis

AI score68

How the Claude Code team does Code Review: six layers of checks and fast rollback

The Claude Code team, under the premise that Claude writes about 90% of the code and the volume of changes has surged, did not cancel review; instead, they reframed the quality problem as "where things might break, and which layer—automated or human—catches it," forming six layers of checks: pre-PR validation, AI review, security review, blocking automation, risk-scaled human approval, and merge queue, with fast rollback as the safety net. From June to September, the number of merged PRs in the Claude Code repository more than doubled, while the share requiring revert dropped by more than 40%; internally, the share of PRs with "substantive comments" rose from about 16% to 54%. The article also offers transferable practices such as CLAUDE.md, REVIEW.md, and limiting rounds to reviewing only the new diff.

Quoted postAddy Osmani · @addyosmani

https://t.co/4Hfk0dYIBI

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

AAlibaba Cloud Developer Community AI14:58Analysis

AI score61

Analysis of the DeepSeek Harness Framework and Local Deployment Tutorial

DeepSeek Harness (dsh) is an open-source agent execution scaffold released under the MIT license. It follows the "Agent = Model + Harness" paradigm, encapsulating model adapters, tool registries, sandboxes, session storage, and more as hot-pluggable plugins, all scheduled by the Cordis kernel. The framework includes four built-in operating modes—Standard, PTC, Minimal, and Creative—and supports session trajectory recording, fork branch replay, and sandbox permission approval. The article presents three deployment methods: launching the Web UI via npx, global installation, and building from source, along with headless batch task commands and examples of custom tool plugins.

#Tutorials / practice#Agents#Open-source ecosystem

336Kr · AI14:56Analysis

AI score62

Developer uses Claude Code to remove Xiaomi TV ads and open-sources CleanYourTV

Developer Gathin handed a Xiaomi TV model to Claude Code, connected to the device over Wi-Fi to disable 25 system components and install a minimalist launcher, then open-sourced the prompt, launcher source code, and troubleshooting notes as CleanYourTV; the project has received about 13,700 stars. The principle is to use the Android ADB interface to investigate and freeze boot ads and telemetry components. The article also mentions YouTuber Ansub Khan's similar project TV Detox, which measured available TV memory rising from about 370 MB to 600–650 MB.

#English text is being prepared.#Agents#Open-source ecosystem

SSemiconductor Industry Observer14:55

AI score74

Open-source tool GPU Unlock unlocks hidden compute power in RTX 30 series graphics cards

The open-source tool GPU Unlock, developed by GitHub user Zagitta, uses a Linux kernel module to rewrite the speed select register before the driver loads, removing NVIDIA's restriction on tensor core instruction issue rates on consumer Ampere graphics cards, without the need to flash the VBIOS or modify the driver. In measured tests, the BF16 compute power of the RTX 3090 Ti rose from 87.5 TFLOPS to 157.4 TFLOPS, an increase of about 80%, and ComfyUI image generation throughput increased by 56% to 59%; for the RTX 3090 under vLLM, prompt prefill throughput improved by about 33%, but the Decode stage of token-by-token generation showed almost no improvement due to memory bandwidth limitations. The project supports GA102, GA103, GA104, and GA106 cores, and is a firmware-level modification not officially supported by NVIDIA, so users must bear the power consumption, cooling, and warranty risks themselves.

Same story · 2 reports

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

KKazike13:02Analysis

AI score67

Retrofitting a TV, Fill Lights, and a Camera with Codex and Claude

The author shared the process of using Codex and Claude to reverse engineer and retrofit real-world devices: first enable ADB debugging on the TV, then have the Agent disable system packages, clear boot ads, and rebuild a clean launcher, and replicate the same method on three other TVs. The team also used Claude to reverse engineer 4 fill lights and built a unified web console, and used Codex to build a multi-camera console for a Sony camera and a DJI Pocket 3 and deploy it to Feishu. The author reminds that different brands of devices expose different interfaces and permissions, and that backups should be made before modifying system software.

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