标签：#ai

Anthropic is walking back a hidden policy that researchers say sabotaged their work.

Deezer will now scan your playlists on other streaming platforms to detect AI-generated music. Deezer was the first of the big streaming services to start labeling AI-generated music. It even offered its tech to other platforms, but it doesn't seem like it had many buyers. Qobuz launched its own detection tech, while Apple and Spotify […]

OpenAI has published a report detailing how China-linked influence campaigns against data centers used ChatGPT.

AI coding tools are changing how software gets built. Claude Code, Cursor, GitHub Copilot, Windsurf and other tools can generate code incredibly fast. For small tasks, they are already useful: write a component, explain a function, scaffold an endpoint, create a test, refactor a file. But after using AI in real projects, one thing becomes obvious: The problem is no longer code generation. The problem is engineering control. Most AI coding workflows still look like this: text idea -> prompt -> code -> fix -> prompt again -> more code -> lost context -> start over It feels fast at the beginning. Then the project grows. Requirements change. Architecture decisions disappear inside chat history. The AI forgets previous context. You start acting as product manager, architect, reviewer, QA, DevOps, and prompt engineer at the same time. That is not software engineering. That is vibe coding. ## Vibe coding works until it doesn't Direct AI coding is great when the task is isolated. Ask for a React component. Ask for a SQL query. Ask for a utility function. Ask for a unit test. No problem. But real software is not a collection of isolated snippets. Real software has: - business rules - architectural constraints - existing patterns - security concerns - database impact - deployment requirements - edge cases - regression risk - long-term maintenance When AI jumps directly from prompt to code, it often skips the thinking that should happen before implementation. The result may compile. But does it fit the architecture? Does it respect the domain? Does it create hidden technical debt? Does it solve the right problem? That is the gap we are trying to close with N45.AI. ## What is N45.AI? N45.AI is a framework that turns AI coding tools into a structured engineering workflow. It works with the tools developers already use, including Claude Code, Cursor, GitHub Copilot, and Windsurf. The idea is simple: Instead of treating AI as one generic assistant, N45.AI organizes the work like a

A useful thing happened in agent infrastructure this June: several teams shipped "escrow layers for AI agents" - production MCP tools that let an agent run a full commit -> hold -> complete lifecycle without a human anywhere in the loop. An agent can now park value with a contract or service, wait for the other side to deliver, and release on completion. That is genuinely new, and it solves a real problem. It is also worth being precise about, because "escrow" and "settlement" get used as if they were one thing. They are not. There are two structurally different ways to make a trade your agent does at 3am trustworthy, and the difference is exactly who holds the money while the trade is in flight . Model one: held custody In the held-custody model, a third party - a smart contract escrow, a custody service, a payment facilitator - takes the funds, holds them, and releases them when a condition is met. The condition can be anything you can express: a delivery confirmation, an evaluator's attestation, a timeout, a multi-sig approval. This is the right tool for a large class of agent commerce. If your agent is paying a merchant, buying a dataset, or hiring another agent to do a unit of work, the hard question is subjective : did the thing actually get delivered, and was it any good? A hash function cannot see that. A custodian can - it gives the trade a place to pause while something or someone checks. The new agent-escrow tooling is built around exactly this shape: a job, a held balance, a release on completion. For agent-to-merchant payments riding on rails like x402, held custody is the honest primitive. The cost is equally concrete. A held balance is a honeypot. Someone controls the funds between commit and complete, which means someone can freeze them, lose them, misconfigure the release condition, or get drained. You have added a trust assumption and a liveness dependency - the custodian has to be online, solvent, and honest at release time. That is often an accep

My daughter Emma is 11. She's been vibe coding lately, and honestly, she's pretty good at it. The...

It is a privacy-first, local-first photo organizer powered by deep learning face recognition. It detects, embeds, and groups faces to organize your photos automatically—all 100% offline. 🔥 Highlight Features: ✅ 100% Local: No cloud APIs, no telemetry, no leaks. ✅ Deep Learning: Driven by OpenCV DNN (YuNet + SFace ONNX models). ✅ Smart Automation: Copies matches, partial matches, and individual profiles into organized folders, complete with ZIP archives and JSON reports. ✅ Standalone EXE: Run it on Windows instantly with zero dependencies. ✅ Dynamic UI: Fully responsive Tailwind dashboard with Dark/Light modes. Check out the repository, download the EXE, or contribute: 👉 https://github.com/Shaan-alpha/face-sort-studio Let me know what you think! ⭐ machinelearning #computervision #python #localfirst #privacy #developers #opensource #ai Internet access on first launch only (to fetch the AI models ~40-50mb)

TLDR: we just launched free OpenClaw hosting. One-click deploy, no infrastructure to manage, only pay...

The decision comes as India emerges as the world’s largest GCC market.

If you doubted his genius, doubt no more.

If you've ever tried to use an AI resume builder, you've probably hit the same wall I did. You sign up, poke around, find the one feature you actually need — and then boom: "Upgrade to Pro for $29/month." It's frustrating. Resume help shouldn't be locked behind a paywall. So I built my own. Meet Persona Persona is an AI-powered resume builder that you run completely on your own machine . No deployment required. No subscription. No account on some third-party service. You clone the repo, set it up, and it's yours. It's a fork of the excellent open-source project ResumeLM , but I've added a bunch of features I couldn't find anywhere else — especially around local AI and template variety. 👉 GitHub: github.com/nithiin7/persona (Drop a ⭐ if you find it useful!) The Big Deal: Run AI Completely Offline with Ollama This is the feature I'm most proud of. Most AI resume tools call out to OpenAI or Anthropic and charge you for every request. Persona supports Ollama — which means you can run the AI model locally on your own hardware, with zero API costs and zero data leaving your machine. Here's how simple it is: Install Ollama on your computer Pull any model ( ollama pull llama3 , for example) Open Persona's settings, point it to your local Ollama URL Done — the AI now runs entirely on your machine No OpenAI key. No Anthropic key. No usage limits. Your resume data never touches an external server. If you do want to use cloud models, Persona supports those too — GPT-5, Claude Opus 4.7, Claude Sonnet 4.6, and a handful of open-source models via OpenRouter. But the Ollama path is what makes this genuinely different from everything else out there. It's 100% Free — Everything Unlocked The original ResumeLM had Stripe payments baked in. I ripped all of that out. Every single feature in Persona is available to every user, always. There's no "Pro plan." There's no feature gating. You self-host it, you own it, you use all of it. 10 Resume Templates Persona ships with ten distinct templ

I built a knowledge graph out of my own work sessions. Hundreds of them — transcripts of me building a system with LLMs, extracted into concepts, decisions, findings, and the edges between them. For a while it felt like the thing was working. I'd query it, get back a clean structured answer, and move on. Then I ran a foreign model against it. I gave a different model my concept definitions and asked it to reconstruct the system, both the vocabulary and the relationships. It recovered 97.7% of the words. It recovered 61.1% of the structure. That 36-point gap was the first time I could see the problem instead of just living inside it. The vocabulary transferred because the definitions were written carefully. The edges didn't, because the edges were the part I'd let the extraction handle. And the whole time, querying the graph had felt complete. The structure came back typed, connected, confident-looking — so I stopped looking. I started calling it premature retrieval closure: the retrieval returns something shaped like a whole answer, which is exactly why I didn't notice the parts that were missing. Part 10 of Building at the Edges of LLM Tooling . If you're running a long-term project through an LLM-backed memory system (anything that turns raw sessions into structured, persistent memory), this is about the step where the structure starts lying about how complete it is. Start here . Why It Breaks Every memory system of this kind does the same move. An LLM reads raw interaction (a conversation, a document, a session log) and lifts structured memory out of it: entities, facts, rules, summaries. That structured memory becomes the thing the agent reads later, instead of the raw record. The lift is where fidelity goes. Pulling clean structure out of messy text means making decisions the text didn't make explicit: which entity this pronoun refers to, whether a relationship is real or inferred, what to keep and what to drop. Those decisions can be wrong, and when they are,

After 8 months of daily Claude Code use, I've distilled my workflow into a 5-layer system. Each layer builds on the previous one. Skip one, and the whole thing falls apart. The Problem with Most Claude Code Users Most people use Claude Code like ChatGPT — open terminal, ask a question, close, repeat. The next day, they explain their project from scratch. Again. The symptom: 20% of every session is wasted on context re-establishment. The root cause: No project memory, no workflow discipline. Here's the system that fixed it for me. Layer 1: CLAUDE.md — Your Project's Memory Anchor This is the foundation. Without it, nothing else works. CLAUDE.md is a file at your project root. Claude reads it automatically at the start of every session. It tells Claude: What this project is (one sentence) The tech stack (specific technologies, not "Python web framework") The architecture (the big picture you'd need 3 files to understand) Unique conventions (not generic advice like "write tests") Quality priorities Bad CLAUDE.md (you've probably written this): # My Project A web application built with Python and FastAPI. ## Development - Write clean code - Add unit tests - Use Git This tells Claude nothing it doesn't already know. Good CLAUDE.md: # CLAUDE.md ## Project Overview Internal RAG knowledge base serving 500+ employees. ## Tech Stack FastAPI + LangChain + Milvus + PostgreSQL + Redis ## Commands - Start: `uvicorn app.main:app --reload --port 8080` - Test: `pytest -x --cov=app --cov-report=term-missing` ## Architecture Request flow: router → service → retriever → Milvus → generator → LLM API Key directories: - app/router/ - API layer - app/service/ - Business logic orchestration - app/retriever/ - Retrieval strategies (vector/BM25/hybrid) - app/generator/ - LLM calls and prompt management ## Key Conventions - All APIs return `{"data": ..., "error": null}` - Retrieval results MUST include source field - Milvus collection naming: `{env}_{doc_type}` The rule: Only write what's uniq

After debugging 20+ broken RAG systems, I've identified the 6 decisions that determine whether yours works. Here's how to get each one right. The RAG Developer's Trap Every RAG developer falls into the same trap: you build the basic pipeline, it sort of works, and then you spend weeks tweaking prompt templates — while the real problem sits untouched in your indexing pipeline. The 80/20 rule: 80% of RAG problems come from indexing, not generation. But 80% of debugging effort goes into generation. Let's fix that. Decision 1: Embedding Model — The Single Biggest Lever The mistake: Using all-MiniLM-L6-v2 for Chinese documents because it's the default in every tutorial. Why it's wrong: It's English-trained. Drop it on Chinese text and it loses 30-50% of semantic fidelity. Language Use This Chinese BAAI/bge-large-zh-v1.5 (1024-dim) Chinese + English BAAI/bge-m3 (multilingual + sparse) English text-embedding-3-large Code jina-embeddings-v3 or voyage-code-3 Non-negotiable: Indexing model and query model must be byte-for-byte identical. Switch models = rebuild entire index. Impact: +15-40% Recall@10 for Chinese RAG. Decision 2: Chunk Size — Not a Magic Number Physics: Too small (< 100 tokens) = semantic fragmentation. Too large (> 1000 tokens) = noise injection. Document Type Sweet Spot Overlap FAQ / Short-form 128-256 20 Technical docs 512 50 Long-form articles 768-1024 100 Code Function boundaries 0 The method matters more than the size. Use recursive splitting, not fixed-length: from langchain.text_splitter import RecursiveCharacterTextSplitter splitter = RecursiveCharacterTextSplitter ( chunk_size = 512 , chunk_overlap = 50 , separators = [ " \n\n " , " \n " , " . " , " " , "" ] ) Impact: +5-15% Recall@10. Decision 3: Index Type — HNSW vs IVF Scale Use Why < 1M vectors HNSW Recall > 0.95 1-5M, RAM tight IVF + PQ 75% memory savings > 5M IVF + PQ + Sharding Horizontal scale Key nuance: HNSW has high insertion cost. Streaming docs → IVF may be better even at small scale. Im

I run a production multi-agent AI system on a single M1 Mac in Jamaica. 6 autonomous agents. 26 cron workflows. 5-layer persistent memory. All containerized, all running 24/7. I checked my OpenRouter dashboard last week and realized something: I'd processed 2.4 billion tokens across 52 different AI models and spent a total of $0.52 . That's not a typo. Here's exactly where that money went and what it means. The Numbers Metric Value Total Requests 26,600+ Tokens Processed 2.4 Billion Models Used 52 Total Cost $0.52 Cost per Token $0.00000021 Tokens per Dollar 4.6 Million For context: GPT-4 Turbo costs about $0.00001 per token at scale. I'm running at roughly 50x below that rate. Where the $0.52 Actually Went Here's the breakdown by model: Model Requests Tokens Cost openrouter/owl-alpha 1,334 251.2M $0.00 nvidia/nemotron-3-super-120b 32 1.8M $0.00 google/gemma-4-31b-it 47 1.8M $0.00 openai/gpt-5 1 2.8K $0.03 google/gemini-3.1-pro-preview 1 3.2K $0.04 anthropic/claude-opus-4 1 2.0K $0.13 qwen/qwen3.5-plus 1 6.3K $0.01 z-ai/glm-5-turbo 1 3.0K $0.01 moonshotai/kimi-k2.5 2 4.1K $0.01 google/gemini-2.5-flash 2 5.5K $0.01 +42 other models ~125 ~8.5M ~$0.28 99.6% of my requests cost exactly $0.00. They ran on free-tier models or local inference. The $0.52 comes from a handful of premium model calls: Claude Opus, GPT-5, Gemini Pro. These are reserved for specific high-quality tasks — not everyday inference. What This Would Cost on Cloud Approach Hardware Monthly Cost Annual Cost My setup (M1 Mac) M1 Mac 16GB, local + free tier ~$0.09 ~$1.04 OpenRouter Paid Tier API-only, no local $15-30 $180-360 AWS (g4dn.xlarge + API) 1x T4 GPU, on-demand $350-500 $4,200-6,000 AWS (g5.xlarge + API) 1x A10G GPU, on-demand $700-1,000 $8,400-12,000 A $1,200 laptop replaces $500-1,000/month in cloud bills. The break-even point is about 2 weeks. How the Architecture Works The key insight: not every task needs a $20/month model . My system routes tasks intelligently: Local inference (free): Ollama

The company changed course after researchers spoke out against the policy, which would have covertly limited Claude’s ability to develop competing AI models.

"I don't prompt Claude anymore. My job is to write loops." — Boris Cherny, Claude Code creator Though I see where he's coming from, I'd put it differently. A developer's job isn't to write loops. It's to design state machines. Every major agent framework — Claude Code, Codex, Cursor, LangGraph — does the same thing under the hood. A while loop calls an LLM, checks if it wants to use a tool, runs the tool, repeats until done. The loop isn't just a solved problem. It's a boring problem. The hard part is everything around it. A loop has no idea what state the work is in. It just keeps going until something breaks or you run out of tokens. That's the Ralph Loop — named after the Simpsons kid who put a crayon in his nose. Agent, infinite loop, go. The Ralph Loop works, is famous, and has zero memory of where it is in the job. Like Ralph, it keeps going without knowing why. The Fix: A Finite State Machine Think about the NBA Finals. The Spurs and the Knicks aren't improvising — every possession has a state. Fast break. Inbound play. Half court set. Each one has specific reads and triggers for what happens next. Point guard De'Aaron Fox isn't making it up as he goes. The system tells him what situation he's in, and the situation tells him what to do. Your agent works the same way. You define the stages — planning, implementing, reviewing, error handling — and you define what triggers each transition. The agent doesn't orchestrate. It executes. One focused job per state. Why This Matters in Production When agents break, it's almost always one of three things: Infinite loops — one system repeated the same answer 58 times before anyone noticed. Context overflow — the history gets so long the model starts quietly forgetting things. Goal drift — 70 turns in, "don't touch auth" has completely evaporated. State machines fix all three. The loop runs until the list is empty, not until you run out of tokens. The goal lives in the transition logic, not in the context getting squeezed

The End of Vibe Coding: Why I Switched to Structured AI Workflows I spent 3 months "vibe coding" my SaaS. Then I realized I was spending more time fixing AI's mistakes than if I'd written it myself. Here's the system that changed everything. In early June 2026, two HN threads with a combined ~1,300 comments told me something had shifted. Thread 1 (~1,100 comments): "What was your 'oh shit' moment with GenAI?" Thread 2 (~230 comments): "What tools have you made for yourself since AI?" Both threads had the same pattern: people started with unfiltered excitement ("I built a whole app in one weekend!"), then hit a wall ("I'm spending more time fixing its bugs than writing code from scratch"). I know the feeling. I lived it. The Vibe Coding Trap When I started building MultiPost — an AI-powered cross-platform content tool — I was deep in "vibe coding" mode: Me: "Make it look better" AI: *adds Tailwind, restyles everything* Me: "Add a filter by date" AI: *adds a date picker, breaks the layout* Me: "Fix that bug where posts don't show" AI: *fixes the filter, introduces a null pointer* Me: "Okay now add a dark mode toggle" AI: *regenerates half the component from scratch* Three weeks later I had a working feature and zero understanding of how any of it actually held together. This was my daily rhythm for weeks. Fast output, slow cleanup. The ratio kept getting worse as the codebase grew. I was optimizing for speed of generation instead of speed of delivery . The "Oh Shit" Moment It came when I reviewed a feature I'd built entirely through unstructured AI sessions. The feature worked. But: The code had 3 different patterns for the same thing (Auth0 token handling in one place, hardcoded keys in another) Error handling was inconsistent — some functions returned null, others threw, others returned Result types Database queries were scattered across the codebase instead of in a repository layer A security reviewer would have cried The AI didn't do this maliciously. It did this

Your health data is probably the most sensitive information you own. Yet, most "AI Health Assistants" today require you to ship your symptoms, moods, and medical history to a cloud server. In the era of Edge AI and Privacy-preserving machine learning , this is no longer a trade-off we have to make. By leveraging the MLX Framework and Apple Silicon's unified memory, we can now run on-device LLMs like Llama-3-8B directly on an iPhone. This tutorial explores how to build a 100% offline, local health journal that summarizes your daily wellness without a single byte leaving your device. If you're looking for more production-ready patterns for secure AI, definitely check out the advanced guides over at Wellally Tech Blog . Why MLX-Swift? 🍏 Apple's MLX is a NumPy-like array framework designed specifically for Apple Silicon. When brought into the Swift ecosystem via mlx-swift , it allows us to tap into the GPU and Neural Engine with incredible efficiency. The Architecture: 100% Offline Inference Unlike traditional CoreML conversions that can be rigid, MLX allows for dynamic graph execution. Here is how the data flows from your typed notes to a structured health summary: graph TD A[User Input: Health Notes] --> B[SwiftUI View] B --> C{Privacy Layer} C -->|Local Only| D[MLX-Swift Engine] D --> E[Llama-3-8B Quantized Model] E --> F[Unified Memory / GPU] F --> G[Local Inference] G --> H[Markdown Health Summary] H --> B style C fill:#f9f,stroke:#333,stroke-width:4px style E fill:#00ff0022,stroke:#333 Prerequisites 🛠️ Device : iPhone 15 Pro or later (8GB RAM is highly recommended for Llama-3-8B). Software : Xcode 15.3+, iOS 17.4+. Tech Stack : MLX Framework, SwiftUI, Llama-3-8B (4-bit quantized). Step 1: Setting Up the MLX Engine First, we need to integrate the mlx-swift package. In your Package.swift , add: . package ( url : "https://github.com/ml-explore/mlx-swift-chat" , branch : "main" ) Now, let's initialize the model. Because we are on a mobile device, we must use a quantiz

Webflow's code export has two problems. It is only available on paid Workspace plans, and even when you pay, it does not include your CMS content: collection lists export as empty states, collection pages export with nothing in them. If your site has a blog, the export gives you a site without a blog. Forms and search are disabled in exported code too, per Webflow's own docs. Meanwhile, the published site is sitting on a CDN, fully rendered. Every CMS page is real HTML. wget --mirror will happily fetch all of it. What wget gives you, though, is not deployable. I migrated a production Webflow site this way and hit the same five breakages everyone hits, so I turned the fixes into a Claude Code skill that runs the whole workflow. This post is the five breakages, because they are useful whether or not you use the skill, and they apply to Framer, Squarespace, and friends with different domain names. Setup: the mirror itself The one wget incantation that matters, because Webflow serves assets from a separate CDN domain and you have to tell wget to follow it: wget --mirror --convert-links --adjust-extension \ --page-requisites --span-hosts \ --domains = yourdomain.com,cdn.prod.website-files.com \ --no-parent https://yourdomain.com/ This downloads every page plus the CSS, JS, images, and fonts they reference, and rewrites URLs to relative paths. It looks complete. It is about 90% complete, and the missing 10% is invisible until the page renders blank. Breakage 1: the page renders blank, console says "integrity" The symptom: your mirrored page shows raw unstyled text or nothing at all, and the console says Failed to find a valid digest in the 'integrity' attribute . The cause is subtle. Webflow ships its <link> and <script> tags with SHA-384 SRI hashes. wget's --convert-links rewrites URLs inside the downloaded CSS files, which changes their bytes, which means the SRI hash no longer matches, which means the browser silently refuses to apply the stylesheet. The file is right