标签：#ai

The most effective safety control for an email agent isn't a better model, a longer system prompt, or a stricter eval suite. It's a draft folder. Here's the setup. Nylas Agent Accounts — currently in beta — are hosted mailboxes your application creates and controls entirely through the API. Each one is a real address with a grant_id that works against the existing Messages, Drafts, Threads, and Folders endpoints, and each mailbox ships with six system folders: inbox , sent , drafts , trash , junk , and archive . That drafts folder is where your approval workflow lives. Full autonomy is a choice, not a default A common pattern for support mailboxes: an LLM drafts replies to common questions, and humans approve the sensitive ones via a webhook flow. The agent handles the boring 80% on its own — password reset instructions, shipping status, "where's the invoice" — and anything touching refunds, legal language, or an angry customer goes through a person first. The threat you're mitigating is mundane: a model that's confidently wrong. Hallucinated discounts, replies to the wrong thread, a tone-deaf response to a complaint. None of these are exotic attacks. They're the everyday failure modes of putting a probabilistic system on an outbound channel, and the mitigation is to put a deterministic gate between "the model wrote something" and "a customer received it." The gate is three API calls The flow: a message.created webhook fires when mail arrives, your classifier decides the risk level, and high-risk replies become drafts instead of sends. Drafts support full CRUD at /v3/grants/{grant_id}/drafts , so the agent creates one like this: curl --request POST \ --url "https://api.us.nylas.com/v3/grants/ $GRANT_ID /drafts" \ --header "Authorization: Bearer $NYLAS_API_KEY " \ --header "Content-Type: application/json" \ --data '{ "subject": "Re: Refund request for order 4821", "body": "Hi Sam, I have processed the refund...", "to": [{ "email": "sam@example.com" }], "reply_to_mess

Tech leaders debate whether the Anthropic episode is a wake-up call for India’s AI ambitions.

The story of AI for the last three years has been written in megawatts. Nvidia GPUs stacked in desert data centers . Models with trillion-parameter counts. APIs that pipe your prompts, photos, and personal data to the cloud, burn a forest of electricity to process them, and return an answer 800ms later. If you're building with AI in 2026, the default assumption is that intelligence lives somewhere else. Your device is just a glass terminal. Apple has been telling a different story. No press tour. No "AGI in your pocket" hype cycles. Instead, a decade of silicon releases where the Neural Engine number (FLOPS) quietly doubled, then doubled again. Core ML updates that casually added transformer support. Here is my thesis : Apple’s on-device AI strategy is a privacy-first, performance-oriented architectural break from cloud-centric AI. By co-designing silicon, models, and APIs to run locally, Apple is unlocking a new class of local-first applications where user data never leaves the device, latency is measured in milliseconds, and features work in airplane mode. This doesn’t kill cloud AI. But it forces every developer to answer a new question: what part of your product must be in the cloud, and what gets better when it stays in the user’s pocket? This post is a technical teardown of that shift. I’ll cover the hardware realities of the Neural Engine and unified memory, the brutal constraints of fitting LLMs on device, what Core ML actually gives developers in 2026, and where this architecture creates new product opportunities that cloud-first can’t touch. I’ll also be blunt about the limits. On-device AI won't replace GPT-5 training clusters. But it might replace 80% of the API calls you make to them. At this moment, cloud AI gets all the headlines, but the real transformation may already be running 24/7 in your pocket, without ever touching the internet. Why the On-Device Push? Apple's AI strategy looks slow only if you measure it in keynote superlatives. Measure it in

Check this out: i Cut Our Image Captioning Costs 60% — Here's the Backend Story Look, I'll be honest. Six months ago I didn't think twice about image captioning. We were a small team, traffic was low, and we just threw everything at GPT-4o because it was the path of least resistance. Then our infra bill came in, my manager did that thing where he just stares at the dashboard, and suddenly I was a "cost optimization" guy. fwiw, that was not in my job description. This is the story of how I went from "we just use GPT-4o for everything" to a multi-model setup that cut our spend by more than half, with quality that — imho — is actually better than what we had before. No, this is not a sponsored post. Yes, I am going to mention Global API at the end because they made my life easier. More on that in a bit. Why Image Captioning Was Even on My Radar Our product has a lot of user-uploaded images. Think: product photos, profile pictures, the usual suspects. For each one we need a short, accessible caption that we use for SEO, alt text, and a downstream tagging pipeline. The downstream pipeline, btw, is the part that actually makes us money. Garbage captions in, garbage tags out. We were calling gpt-4o for everything. Every image. No caching. No batching. No thought. Each call cost us $2.50 per million input tokens and $10.00 per million output tokens. You don't have to be a math PhD to know that scales badly. I was not a math PhD. I am still not a math PhD. But I can do division. When I started pulling the numbers, the situation was grim. We were processing roughly 8 million images a month, and each one was generating more tokens than it needed to. I found one image in the logs that had produced a 4,000-token caption. The image was a screenshot of an error message. The caption was longer than the error. The Wake-Up Call: Actually Reading the Catalog One Saturday morning, coffee in hand, I decided to actually look at what was on offer. I'd been ignoring the multi-model world b

Last week I couldn't shake an idea: what if I had an AI that knew everything I know ? Not ChatGPT — something on my hardware, holding my knowledge, answering to no one's API bill. Yesterday I built it. Here's the honest breakdown. What it does NEXUS runs on a regular Windows laptop — aging i7, 16GB RAM, no GPU. It: Remembers everything. Drop any file in a folder; 60 seconds later it's searchable memory. Answers from MY knowledge. "Which of my projects were formally closed and why?" — it answers from my actual records. Watches the live web. Every 2 hours it pulls Hacker News and news feeds, learns what's trending, pings my Telegram. Reports to my phone. 7 AM daily briefing: what it learned, what's running, what needs me. The stack — all free, all open source Ollama runs the models (Llama 3.2, Mistral 7B). Open WebUI is my private ChatGPT. Qdrant stores memory. n8n automates. SearXNG searches privately. PostgreSQL, Redis, and MinIO handle data. Commercial equivalent: $300–500/month . My cost: electricity. The memory trick nobody explains simply Parse — extract text from any file Chunk — split into ~300-word pieces Embed — each chunk becomes 768 numbers representing its meaning Store — a database that searches by similarity Your question becomes 768 numbers too, and the database finds memories with similar meaning — not matching keywords. I asked "how do I get clients cheaper" and it found my notes on "reducing customer acquisition cost." Different words. Same meaning. That's the magic. What surprised me A 2GB model is genuinely useful. Llama 3.2 3B answers from my knowledge in seconds, on CPU. The automation matters more than the AI. The watched folder + Telegram bot turned a cool demo into a system I actually use. Windows is fine. Docker Desktop + WSL2 ran all nine services without drama. The bill, honestly Hardware: $0 (laptop I own) Software: $0 (open source) APIs: $0 (all local) Time: one focused day The only future cost is a cloud GPU server (~$65/mo) when I outg

Introduction Zhipu AI (THUDM) has officially released GLM 5.2 , the latest iteration of its flagship open-weights model family. Announced today by Jie Tang on Twitter, the release is already making waves on Hacker News — racking up 269 points and 146 comments within hours. For developers who have been watching the open-weight LLM race, this is a significant moment. What's New in GLM 5.2 GLM 5.2 builds on the GLM-4 series that put Zhipu on the global map. The release focuses on three areas that matter most to production teams: Stronger reasoning and coding : Improved performance on multi-step reasoning benchmarks and competitive code generation against closed-source models like GPT-5 and Claude 4.5. Better multilingual behavior : GLM has always been strong in Chinese; 5.2 pushes English-quality code reasoning and longer-context retrieval closer to frontier levels. Longer context window : Reports point to a 200K+ token context with reduced degradation on long-document tasks — useful for codebase-level analysis. Weights, inference code, and a technical report have landed on Hugging Face under the THUDM organization, with an OpenAI-compatible API endpoint exposed by Zhipu's own platform. Why It Matters The open-weights race has consolidated around a handful of serious contenders — Llama, Qwen, DeepSeek, Mistral, and now GLM. Zhipu's positioning is unique: a Chinese lab that consistently weights-and-releases frontier-class models while still maintaining a hosted commercial API. For developers, that translates to real options: You can self-host on a single H200 or a pair of RTX 5090s and skip per-token API costs entirely. You can route between self-hosted GLM 5.2 and a hosted Anthropic/OpenAI endpoint depending on cost, latency, and capability. You get an OpenAI-compatible endpoint, so dropping GLM into an existing stack is a config change, not a rewrite. The Bigger Picture GLM 5.2 lands on the same week that U.S. regulators have reportedly cracked down on Anthropic model

Token economics is not only about token price. It is about designing the rules, incentives, and long-term logic of a Web3 ecosystem. When people start building a Web3 project, they usually focus on the visible parts first. They think about the smart contract, the frontend, the wallet connection, the token launch, the whitepaper, and maybe the community. All of those are important. But there is one part that can decide whether the project survives or fails: Token economics. A project can have clean smart contracts, a nice UI, and strong marketing, but if the token economy is weak, the project can slowly collapse. Users may come only for rewards, early investors may dump, inflation may destroy value, and the token may lose its reason to exist. That is why token economics should not be treated as just a “crypto finance” topic. For developers and Web3 builders, token economics is closer to system design . It defines how value moves inside the ecosystem, how users are rewarded, how supply is controlled, how governance works, and how the project can grow without depending only on hype. What Is Token Economics? Token economics, often called tokenomics , means the design of how a token works inside a project. It answers questions like: Why does this token exist? Who receives the token? How is the token used? How many tokens will exist? How are rewards distributed? When can team and investor tokens unlock? How does the project treasury work? What creates real demand for the token? In simple words, token economics is the rule system behind a token. A token is not only something people buy and sell. In a real Web3 product, a token can be used for payments, staking, governance, access, rewards, collateral, or network fees. If the token has no clear role, it becomes only a speculative asset. That is dangerous because speculation can bring attention, but it cannot support a project forever. Why Developers Should Care Some developers think token economics is only for founders, eco

Meta starts dismantling its $2 billion Manus acquisition after Beijing ordered the deal reversed.

GitHub's Agentic Workflows preview has the kind of headline that makes people reach for the wrong conclusion. Natural language Markdown can turn into GitHub Actions workflows. That sounds like "the YAML is going away." I do not think that is the interesting story. The interesting story is that the agent is not escaping the workflow engine. It is being pulled into it. That matters because a lot of agent demos still pretend the future is a smart process floating above the boring machinery: the agent understands the request, edits the repo, runs some commands, and hands back a neat result. Nice demo. Very clean. Production engineering is not clean like that. Production engineering has permissions, logs, runner groups, approval rules, secrets, firewalls, budgets, weird old repositories, compliance questions, and someone who has to explain what happened when the helpful automation did something surprising. So the shape of Agentic Workflows is useful precisely because it is less magical than the demo version. GitHub is putting agents inside the same CI/CD world that already carries a lot of organizational trust. That is the right direction. markdown is not the control plane The cute part is that a developer can describe a workflow in Markdown and have GitHub turn that into standard Actions YAML. That is useful. YAML is not a personality test, and most teams have better things to do than memorize every Actions syntax edge case. But Markdown is only the input surface. The control plane is still Actions. That distinction matters. If the generated workflow is a normal Actions workflow, then all the existing machinery can still matter: repository permissions, runner selection, logs, environments, approvals, branch protection, organization policy, and whatever security controls the company already built around CI. This is where I get more optimistic about agentic tooling. The bad version of agents asks every organization to trust a new, parallel execution model because the mode

The hype cycle for AI adoption in businesses often follows a familiar, and often frustrating, trajectory. It begins with the undeniable allure of a powerful new tool—seeing ChatGPT effortlessly summarize documents or Copilot intelligently autocomplete code is undeniably impressive. The immediate reaction is almost universally: "We need to get AI." So, tools are procured, workshops are run, and initial enthusiasm soars. Yet, more often than not, six months down the line, that excitement has waned, and the needle on actual business transformation hasn't moved. The fundamental operational rhythm of the organization remains unchanged. As a Senior IT Consultant and Digital Solutions Architect with over a decade of experience, I've observed this pattern repeatedly across various client engagements. This isn't a failure of the technology itself, but rather a profound strategy failure . It’s the single most common and costly mistake I encounter in AI adoption today. The Critical Distinction: Tool Acquisition vs. Strategic Integration Think of buying an AI tool like purchasing a state-of-the-art machine for a workshop. Its inherent value is immense, but if it sits in a corner, unused or without a defined process around it, that value remains entirely theoretical. It's an expense, not an asset generating return. Real AI integration , on the other hand, is a disciplined, multi-faceted endeavor. It's about deeply understanding your existing business processes, meticulously identifying precisely where intelligent automation can create genuine, measurable leverage. It involves designing a holistic system that delivers this leverage, and critically, establishing robust mechanisms to measure the actual outcomes against predefined business objectives. This holistic approach is strategy. The tools—be it an LLM, a specific AI platform, or an automation suite—are merely components that strategy carefully selects and orchestrates to achieve a greater aim. Where AI Integration Truly Crea

According to the Wall Street Journal, the export control directive that led to Anthropic cutting off access to Fable 5 and Mythos 5 was triggered in part by cybersecurity research from Amazon and conversations between CEO Andy Jassy and the White House. According to the report, the paper from Amazon claims that, through a series […]

AI Agents Level Up Workflows: Terraform MCP, WebMCP, Pinecone Integrations Today's Highlights This week showcases significant advancements in AI agent orchestration and workflow automation, with new tools enabling AI to manage infrastructure, interact with the web, and leverage enterprise data. These developments highlight the growing maturity of applied AI frameworks for real-world production use cases. Terraform MCP Server Enables AI Assistants to Interact with Terraform Infrastructure (InfoQ) Source: https://www.infoq.com/news/2026/06/terraform-mcp-server-ga/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global The Terraform MCP (Machine Code Platform) Server has recently achieved general availability, marking a significant step forward in the integration of AI assistants with infrastructure-as-code paradigms. This new server allows AI agents to directly interpret and execute operations on Terraform-provisioned infrastructure, providing a robust and standardized interface for AI-driven automation. Instead of relying on complex scripting or indirect API calls, AI assistants can now receive natural language instructions, translate them into appropriate Terraform commands, and manage resources like virtual machines, networks, and databases across various cloud providers. This capability introduces unprecedented potential for advanced workflow automation within DevOps environments. Teams can leverage AI for tasks ranging from autonomous resource provisioning based on demand surges to intelligent incident response that dynamically scales or reconfigures infrastructure. The MCP Server acts as a crucial middleware, ensuring secure and controlled interaction between intelligent agents and critical infrastructure, thereby reducing manual operational burdens and enhancing system resilience through automated, intelligent responses. This direct interaction paves the way for a new era of self-managing, AI-orchestrated cloud environments. Comment: This i

Once again, AI proves to be an unreliable source of information about AI.

In 2026, Marvel and Mattel are both releasing projects designed to capitalize on people's love for iconic animated heroes from their childhoods. Masters of the Universe has put a live-action He-Man on the big screen, and the second season of X-Men '97 is about to fling some of Charles Xavier's mutants into an apocalyptic future. […]

Amazon CEO Andy Jassy may have been the source of security concerns that led Anthropic to cut off worldwide access to two models on Friday.

Dulu, nyari query yang bikin database spike itu bisa makan berhari-hari. Yang nyari capek, yang nge-fix juga capek. Sekarang? Hitungan jam — dan bonusnya, sambil belajar hal baru juga. Ceritanya begini. Kalau kamu pernah kerja di aplikasi yang pakai ORM (Object-Relational Mapping — semacam "penerjemah otomatis" antara code dan database), pasti familiar sama situasi ini: database tiba-tiba lambat, kamu dapet raw query yang jadi biang kerok, tapi di codebase kamu nulis pakai syntax ORM yang bentuknya beda jauh dari SQL mentah itu. Buat yang belum pernah deal sama ORM, bayangin gini: kamu nulis pesan dalam bahasa Indonesia, lalu ada "penerjemah otomatis" yang convert jadi bahasa Jepang sebelum dikirim ke penerima. Suatu hari ada masalah di pesan yang terkirim — tapi kamu cuma bisa lihat versi bahasa Jepang-nya. Nyari bagian mana dari tulisan Indonesia kamu yang bikin terjemahan-nya bermasalah? Itu effort-nya yang bikin pengen balik tidur aja. Sekarang dengan bantuan Kiro, cukup kasih raw query + akses ke codebase, dia otomatis nyari bagian mana di code yang nge-generate query bermasalah itu. Yang dulu butuh berhari-hari, sekarang bisa selesai dalam hitungan jam — dan itu baru tahap investigasi, belum termasuk fixing-nya. Ceritanya Kenapa Bisa Pakai Kiro Akhir-akhir ini lagi aktif pakai Kiro di tempat kerja. Awal tahun lalu kantor dapat credits melalui program Kiro for Startup , jadi ya sekalian dimaksimalkan. Selain buat debug dan explore query di MS SQL Server, kadang pakai Kiro juga buat analisa log AWS CloudWatch — sambil kasih context aplikasi yang running biar analisa-nya lebih akurat dan gak generic. Di tulisan kali ini, saya mau sharing gimana pakai Kiro sebagai partner beberapa minggu terakhir buat improve query performance di aplikasi .NET Core. Kenapa "partner"? Karena Kiro-nya gak boleh langsung akses ke database — jadi wajib melalui perantara saya. Kita discuss, kolaborasi, dan nge-solve bareng. Bukan AI yang dikasih tombol terus disuruh jalan sendiri. Wakt

If you use Claude daily, you've probably hit a wall mid-task. The reason it's so easy: Claude doesn't have one limit. It has several. A rolling 5-hour session window A weekly pool shared by all models Separate weekly caps per model (Opus, Sonnet, Claude Design) And the weekly pool is sneaky: Fable has no cap of its own and draws that shared pool down ~2× faster than Opus . So you can burn the week without realizing it. I wanted that state visible at a glance instead of digging through settings, so I built Claude Usage Monitor — a small browser extension for Chrome and Firefox. What it does It puts a color-coded % badge in your toolbar , and one click shows the full breakdown: Current 5-hour session + live countdown to reset Weekly all-models pool (with the Fable draw-down flagged on the card) Per-model weekly caps: Opus, Sonnet, Claude Design Daily Claude Code routine runs Extra-usage credits + prepaid balance for pay-as-you-go Auto-detects your plan (Pro / Max 5x / Max 20x / Team), so every number matches your subscription Six themes The privacy part (the part I cared about most) It's open source (MIT) and runs 100% locally — no backend, no analytics, no servers. A design decision I'm happy with: the host permissions are scoped to exactly four claude.ai endpoints (org list, usage stats, routine-run budget, prepaid balance). It physically can't read your chats, projects or files. Narrow permissions = less to trust, and less to exploit. One honest caveat: it reads claude.ai's own usage endpoints, so it can break if Anthropic changes them. Not affiliated with Anthropic. Try it It's free, no account needed (desktop Chrome & Firefox): 🔗 claude-monitor.com Chrome Web Store Firefox Add-ons Happy to hear what else you'd want tracked — drop a comment.

I take a lot of screenshots. The article I'll read later. The recipe I'll cook on Sunday. The movie name from someone's Instagram story. A job post, a product, a tender. Most of them die in my camera roll. So I built Chista — an iOS app that auto-imports every screenshot, classifies it with AI (Article, Product, Event, Reference, Media), and surfaces a one-tap action: Buy on Amazon , Add to Calendar , Reserve on OpenTable , etc. It shipped on the App Store thirty days after I started, for about $300 in total cost. The interesting part wasn't the app. It was what the build revealed. What I built Chista is a native iOS app + Python backend. iOS reads new screenshots in the background via PHPhotoLibraryChangeObserver , scoped to PHAssetMediaSubtype.photoScreenshot (so it literally can't see your other photos). Each new screenshot gets OCR'd on-device with Apple Vision , then the image + OCR text get POSTed to the backend. Backend sends the pair to OpenAI GPT-4o with a structured prompt that returns a CategorizationResult JSON: category, subtype, title, suggested action, extracted data (price, deadline, URL, etc.). Result gets persisted and pushed back to the inbox via Supabase real-time. That's the whole thing. The "magic moment" is just: you screenshot something, switch to Chista a few seconds later, it's already sorted with a contextual action button. Stack Layer Tool Why iOS app Swift 5.10, SwiftUI, StoreKit 2 iOS 17+, modern surface Backend FastAPI on Railway One-file ergonomics, fast cold starts Database + Auth Supabase Postgres + JWT auth out of the box AI OpenAI GPT-4o (Pro), gpt-4o-mini (Free) Tier-routed at categorization time Push APNs via aioapns Direct, no Firebase middleman Subscriptions StoreKit 2 + app-store-server-library Server-side JWS verification Affiliate routing Custom matrix in Supabase tables Amazon Associates wired, more pending Hosting (web) Cloudflare Pages Free, fast, never goes down No frameworks I wouldn't reach for again. What it cost Lin

Six months ago, I could tell you which model to use for almost any job, and I would have said it with confidence. Today I hedge, and so does almost everyone I talk to who builds with these tools. The reason is simple. The ground keeps moving under us. Models get smarter on a schedule no one can forecast, and they get cheaper to run on a second schedule that is just as hard to predict. Both curves are bending at once, and they point in directions that change how I build and how I think you should build too. I spend my days crafting development, content and productivity workflows that lean on these models. I wire up agents, route tasks, and watch the bills. So this is not a far-off observation for me. It is the thing I am living with week to week, and it has forced me to rethink habits I held for years. This is not the usual story about a single breakthrough. It is four shifts happening together. Frontier performance is climbing past what most of us guessed was possible this year. Small models are getting good enough to run on a phone or a thirty-five dollar computer. The smart move has stopped being "pick the best model" and started being "build a system that picks for you." And a coding startup with a rocket company behind it is showing what happens when product, data, and compute sit under one roof. Let me take these one at a time. Then I want to show you what they add up to, because the sum is bigger than the parts. Performance Is Outrunning the Forecasts Start at the top of the market, where the most capable models live. Anthropic now ships a tier above its Opus line. The Fable and Mythos family is a class of model built for problems that smaller systems still fumble: long chains of reasoning, deep code work, research that needs to hold many threads at once. Claude Fable 5 carries extra safety work so it can go out to the public. A more powerful sibling, used inside a small set of trusted partners, stays behind tighter controls. The names are not the point. The p

The thing I'm building, App Store Analyzer, is a website that does one thing: it reads an iOS niche and writes a deep market analysis for indie devs. For a long time that analysis had one brain — and it spoke German. That made sense at the start. German is my home market and my own language, so I built the analysis logic in German first. I could actually feel whether the output was good or garbage, section by section, because I was reading it in the language I think in. It got deep. Reliable. I trusted it. Then it started to hurt. Every time I wanted the analysis in another language, I was basically running the whole expensive thinking step again from scratch. German code, German slugs, German routes, German everything — and a goal of serving 14 languages. The whole thing fought itself. So I rebuilt the brain in English. Not "translated the code" — rebuilt the canonical brain so English is the one source of truth. Now the engine thinks once in native English, and that single analysis gets translated and cached into 13 other languages . Generate once, translate many. It was not a clean ride. The lows. A refactor left a pile of undefined names and quietly 500'd my detail pages — live, in production, while I thought everything was fine. I misread a normal cache warm-up window as a dead backend more than once and "fixed" things that were never broken. I spent an embarrassing stretch hammering an endpoint with a wrong key, watching 403 scroll by, before realizing my terminal had eaten the line that set the key. Small things. Hours each. The highs. Two of them I didn't expect: It got cheaper , not just cleaner. I'm not paying for a full deep analysis per language anymore — one real generation, then lightweight translations. For a solo dev watching every API cent, that's the whole game. And the English brain was actually sharper . I ran the old German output against the new English one side by side, fully expecting English to be the weaker copy. It wasn't. In a few section