标签：#p

Microsoft's Surface Laptop Ultra is basically a MacBook Pro clone (that's powered by NVIDIA's RTX Spark).

NVIDIA claims the new RTX Spark chip for PCs will offer 1 petaflop of AI computing power.

Once upon a time, Microsoft had to write off $900 million betting an Arm-based Nvidia chip could power its first flagship Windows portable, the original Microsoft Surface. But today, it's trying again. Microsoft and Nvidia have just announced the Surface Laptop Ultra, a computer with a new Arm-based Nvidia chip at its core. There's a […]

This fall, Nvidia will officially become a consumer PC chipmaker like Intel, AMD, Apple, and Qualcomm, putting a complete computing chip - not just graphics - into the very heart of laptops and mini-PCs. After many months of leaks, it's finally announcing the RTX Spark, the first in a family of chips that will meet […]

The row looked perfect. rating: 7 . Valid JSON, right type, no nulls, no missing keys. My schema check waved it through. The page had returned HTTP 200. The selectors hadn't moved. Everything green. A rating of 7 on a 5-star site is impossible. The model invented it, formatted it correctly, and handed it to me with total confidence. That's the failure I want to talk about. Not the scraper that breaks loudly. The one that hands you a clean-looking row that is quietly, plausibly false — and sails past every check you have, because your checks are all looking at the shape of the data, and the lie is in the value . TL;DR HTTP 200, intact selectors, and valid JSON tell you the form is fine. They say nothing about whether the value is true. When an LLM extracts from messy free-text, structured-output mode guarantees you get valid JSON. It does not guarantee the content is real. The model fills uncertain fields rather than leaving them empty — because the schema demands a complete row. A ~60-line value-level sanity gate (ranges, dates, cross-field, reference, language) catches the obvious lies before they hit your database. Real code and real output below. The honest catch: this gate catches rule violations , not plausible lies inside the allowed range . A rating: 4 where the truth is 2 slides right through. I'll be specific about where the gate stops. Two different ways a scraper lies to you I wrote about source drift last week — the case where the page changes underneath you and a 30-line schema check catches the structure shifting. That's an input problem. The source mutated; your agreement with the page broke; you detect it by watching the shape. This is the other end of the pipe. The source is fine. The page is intact, the selectors are correct, the structure is exactly what you expected. The thing that lied to you is the model , on the extraction step, when you asked it to pull structured fields out of a paragraph of human prose. Those two failures feel similar and t

I spent way too long staring at this error. If you're here, you probably are too. Your requirements could not be resolved to an installable set of packages. Problem 1 - Root composer.json requires php ^8.3 but your php version (8.2.30) does not satisfy that requirement. My Laravel 13 app needed PHP 8.3. My Hostinger server was running 8.2. composer install refused to budge. Here's exactly what happened and the one-liner that fixed it. The Setup I was deploying a Laravel 13 + Inertia + React app to Hostinger shared hosting. Laravel 13 requires PHP 8.3 minimum — and so do its locked Symfony 8.x and PHPUnit 12.x dependencies. My composer.lock had been generated on a local machine with PHP 8.3, but Hostinger's CLI was defaulting to 8.2. The hPanel showed PHP 8.3 selected under PHP Configuration . The website itself was running fine on 8.3. But SSH? Still on 8.2. $ php -v PHP 8.2.30 ( cli ) That disconnect — hPanel vs. CLI — is the trap. What I Tried First composer update My first instinct was to just let Composer resolve newer compatible versions: composer update No luck. The root composer.json itself declared "php": "^8.3" , so Composer refused before even touching the lock file. The PHP constraint wasn't just in dependencies — it was in my own project requirements. composer install --ignore-platform-reqs This flag skips platform checks and forces the install anyway. It works , but it's a lie — you end up with packages that may behave incorrectly or fail at runtime because they genuinely require PHP 8.3 features. Not a real fix. Changing PHP in hPanel Hostinger's control panel has a PHP version switcher under Hosting → Manage → PHP Configuration . I had already set this to 8.3. This controls the web server / FPM version — what runs your .php files in the browser. It does not change what php points to in your SSH terminal. That's the key distinction most tutorials miss. What Actually Fixed It Hostinger installs multiple PHP versions in parallel. They live in /opt/alt/ph

Based on real system architecture decisions. About a $660K AI platform, three AI agents that kept the dashboard green, and a P0 incident that cost $3.15M over one weekend. Act 1 · The All-Hands Meeting Wang Lei, VP of Product, stood in front of the big screen, a smile on his face. Behind him, a dashboard rolled data from the "Axon AI Client Engineering Platform — Q1 Performance Report." Numbers cascaded across the wall: Metric Axon Platform Human Team (Last Q1) Improvement Avg daily tickets processed 847 312 +171% Avg first response time 12s 4h 17m ↓ 99.92% Customer satisfaction 4.8/5 4.1/5 +17% Monthly operating cost $52K $133K −61% Twelve department heads sat in the room. Dead silence. Wang Lei planted both hands on the table and scanned the room. His eyes landed on me. "Alex. Your team processed 312 tickets last Q1. Axon processed more than that in a single day last month." He smiled. Not a friendly smile. A sentencing smile. "And Axon costs less than a third of your team's operating expense." "We invested $660K in the whole platform. At current operating costs, it pays for itself in eighteen months." "After management review — the Client Engineering technical liaison function is being fully transitioned to the Axon platform." He clicked to the next slide. "Employees in replaced roles will complete exit interviews within the week." Someone inhaled sharply. I didn't. I opened my notebook to page 37. "Wang, what dimensions are these numbers from?" "What do you mean, 'what dimensions'?" His smile tightened. "Of those 847 daily tickets — how many are auto-tagging and routing, and how many are actual technical resolutions?" The room went quiet for about five seconds. Wang Lei looked at me. "Axon's ticket closure rate is ninety-three percent." "What's the reopen rate?" He paused. "What?" "After Axon replies — how many customers reopen the same ticket within twenty-four hours?" "We're still collecting that —" "Let me save you the trouble." I turned my notebook toward th

[Paper Notes] JEPA: Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture 🔗 TL;DR: JEPA learns a a generalized semantic representation with less data pairs by predicting missing information in the embedding space , which helps it disregard unnecessary noisy from input(pixel)-level details and learns at a higher abstraction level with good semantic generalization. 1. Innovation & Significance The Bottleneck: Image-text data pair labels are hard to find Pixel level pre-training paired & data augmentation are strongly biased towards trained data distribution, hard to determine proper generalization and level of abstraction. JEA's (Joint Embedding Architecture) collapse probelm: encoder & decoder attempts to cheat by always landing on trivial constant when predicting itself (reconstruction) and gets away with an easy Error=0. The Solution: > Chain-of-thought ⭕ Mask pre-training to reduce data & generalize↓❌ Bad/lower semantic representation without semantic target, could be learning noisy local pixel correlation↓⭕ Learn at the embedding level to omit pixel input and generalize⭕ Adds context encoder & positional encoding to inject context and force model to pick up image inherent structure from reconstructing multiple masked patches with one target.↓❌ JEAs wants to cheat: if I always map all pixels to a constant for both the predictor and end target encoder then the reconstruction error is always collapsed to zero! Hehe~ ↓ ⭕ EMA (Exponential moving avg.): Update target encoder parameters from the EMA of context encoders. This 'delays' the target encoder to prevent collapsing (a trick from the BYOL paper[2020], proven essential to training JEAs with ViT). 2. Model & High-Level Intuitions 2.1 Model Architecture Input: randomly samples block masks from original image within certain aspect ratio changes, and apply mask for context image 2.1.2 Context Context Encoder: ViT encodes context image to embedding SxS_x S x ​ Mask Token : an [1,D] random

Last year I needed to ask for a raise. I knew my number, I'd read the guides, I had bullet points in my notes app. Then my manager said "let's chat about your goals for next quarter" and I said "sounds great, looking forward to it" and hung up. Never brought up money. Same thing kept happening elsewhere. Coworker taking credit for my work, I said nothing. Relationship that should've ended months earlier, I kept postponing. I always knew what to say. I just couldn't say it with someone actually looking at me. So I started building a thing to practice on. That thing became cosskill . What it actually is You pick a persona, tell it the situation in a sentence, and start talking. The persona doesn't help you. It holds position and pushes back. You practice not folding. Think of it as a flight simulator for hard conversations. You rehearse until your opener comes out steady, then go do the real thing. 20 personas across five categories: Operators (Musk, Jobs): first-principles thinking, harsh product feedback Strategists (Trump, Buffett): treat everything as a deal or a bet Relationship (Ex, Coworker): breakups, workplace friction, family money Philosophy (Socrates, Aurelius, Confucius, Sun Tzu, four more): each tradition frames problems differently Psychology (Rogers, Rosenberg, Ellis, Frankl, Kahneman, Jung): therapeutic frameworks on real situations These aren't celebrity impressions. The Buffett persona won't hype your startup idea. It'll ask "what's the downside?" and keep asking until you have something concrete. Tech stack Next.js 16 on Cloudflare Workers. DeepSeek for inference. Cloudflare D1 (SQLite at edge) for the bits that need to persist. No user accounts, chat history lives in localStorage. Monthly cost stays low enough that the free tier (10 messages/day) doesn't worry me. Why I made these choices DeepSeek instead of GPT-4/Claude. Each conversation is 10-30 messages. At GPT-4 pricing a free product bleeds money. DeepSeek gives maybe 90% of the quality for

Frontend request code often involves repetitive state management. This article compares Axios and alova through a paginated list example, analyzing how request strategization reduces boilerplate and when it's a good fit. The Pattern: Paginated List in Two Ways A common requirement: fetch a user list with pagination. Approach 1: Axios const [ data , setData ] = useState ([]); const [ page , setPage ] = useState ( 1 ); const [ total , setTotal ] = useState ( 0 ); const [ loading , setLoading ] = useState ( false ); const [ error , setError ] = useState ( null ); const fetchUsers = async ( currentPage ) => { setLoading ( true ); setError ( null ); try { const res = await axios . get ( ' /api/users ' , { params : { page : currentPage , pageSize : 10 }, }); setData ( res . data . list ); setTotal ( res . data . total ); } catch ( e ) { setError ( e . message ); } finally { setLoading ( false ); } }; useEffect (() => { fetchUsers ( page ); }, [ page ]); This pattern appears in nearly every data-fetching component. The actual business logic — GET /api/users — occupies a single line. The rest is infrastructure: state declarations, loading toggles, error handling, and effect management. Approach 2: alova with usePagination const { data , total , loading , error , page , pageSize , nextPage , prevPage , } = usePagination ( ( page , pageSize ) => alovaInstance . Get ( ' /api/users ' , { params : { page , pageSize }, }), { page : 1 , pageSize : 10 } ); Both implementations are functionally identical. The key difference is where the state management logic lives: in the component (Axios) vs. inside the hook (alova). What Changed Component of Axios version Handled by alova loading state + toggling Managed internally by usePagination error state + try/catch Managed internally by usePagination data state + assignment Returned as reactive value page state + change handler Built-in nextPage / prevPage total state extraction Extracted from response automatically useEffect dependency tr

This isn't a tutorial. It's real experience. Most articles about making money with Python are vague: "Learn Python to make money" (then what?) "Do data analysis freelancing" (how to get clients?) "Write web scrapers" (legal gray area) I'll share my actual path: building an Excel template generator with Python, listing it for sale, and earning my first dollar. Why This Direction My background: Know Python, but not expert Made some automation scripts No product design experience Want products (scalable) not services (time-for-money) The opportunity: Huge Excel template market (many 10k+ sales on Gumroad) Templates are static, hard to customize I can make a "template generator" for customization Technical feasibility: Python's openpyxl generates Excel programmatically JSON config is user-friendly ~300 lines of code The Product Not an Excel file. A Python script that generates Excel files . Users get: generator.py - generator code config.json - configuration README.md - documentation Workflow: Edit config.json → Run python generator.py → Get customized Excel Technical Implementation Core code is simple: from openpyxl import Workbook from openpyxl.styles import Font , PatternFill wb = Workbook () ws = wb . active # Header style header_fill = PatternFill ( start_color = ' 6366F1 ' , fill_type = ' solid ' ) header_font = Font ( bold = True , color = ' FFFFFF ' ) # Write header ws [ ' A1 ' ] = ' Project Name ' ws [ ' A1 ' ]. fill = header_fill ws [ ' A1 ' ]. font = header_font # Add dropdown from openpyxl.worksheet.datavalidation import DataValidation dv = DataValidation ( type = ' list ' , formula1 = '" In Progress,Completed,Paused "' ) ws . add_data_validation ( dv ) dv . add ( ' B2:B100 ' ) wb . save ( ' output.xlsx ' ) Loop to create sheets, set styles, add validation. Productization Process Step 1: MVP One module only (knowledge base) Test generation Use myself for a week Step 2: Expand Add 6 modules Add JavaScript version (using exceljs ) Improve docs Step 3: Package

Introduction Understanding how to find prime numbers is one of the best ways to develop logical thinking in programming. It looks simple on the surface, but it teaches you how to break a problem into smaller steps, build a solution gradually, and then improve it into a clean and reusable structure. In this blog, we will not jump directly into code. Instead, we will start from basic thinking, slowly convert that thinking into logic, and finally refine it into a proper Python program using functions and loops. The goal is not just to find prime numbers, but to understand how programming logic is actually built in real development. 1. Understanding the Problem First Before writing anything in Python, we need to understand what a prime number actually means. A prime number is a number that: is greater than 1 has exactly two divisors: 1 and itself So the real question becomes: How do we check whether a number has any divisors other than 1 and itself? That is the core problem we are trying to solve. 2. Thinking Like a Human Before Coding Let’s take a number, for example 13. To check if 13 is prime, we naturally try dividing it by smaller numbers: 2 → does not divide 13 3 → does not divide 13 4 → does not divide 13 5 → does not divide 13 and so on If none of these numbers divide 13 completely, then 13 is prime. So the logic is simple: Try dividing the number by possible candidates and see if any divide it perfectly. 3. Turning Thinking into a Basic Algorithm From the above idea, we can form a basic structure: We need: a number to test a variable that moves through possible divisors a way to detect whether a divisor exists We start checking from 2 because every number is divisible by 1 anyway. We also do not need to check beyond half of the number, because a number cannot have a divisor greater than half (except itself). So the idea becomes: Start divisor from 2 Go up to number // 2 If any number divides it evenly, it is not prime 4. First Working Logic (Direct Implementati

Six months into running my SaaS, my "feedback system" was three browser tabs, a starred Gmail folder, and a sticky note on my monitor that said "check Discord." That was the whole system. It held together until the day I found a three-paragraph email from a paying user — a genuinely detailed feature request with a real use case — sitting unread for 24 days. His last line was: "Happy to pay more if you can support this." I replied the same afternoon I found it. His reply: "Switched last week, thanks anyway." That was the moment I stopped treating feedback management as a nice-to-have. Why the usual fixes didn't fix anything I tried the obvious things first. I want to document them because I see a lot of people cycling through the same failed solutions. Notion database 🪦 Built a beautiful one. Color-coded tags, priority columns, status tracking. It lasted 11 days before nobody — including me — was maintaining it. The friction of "open Notion, find the right database, fill in six fields" is invisible when you're designing the system and fatal when you're in the middle of a support conversation. Airtable form 🪦 Better entry point, still disconnected from where users actually were when they had feedback. Nobody bookmarks your Airtable form. They DM you on Discord and you think "I'll add that later" and you don't. Canny — this one actually worked, for a while I genuinely liked Canny. Clean interface, users could upvote requests, I could see what was popular. It felt like a real system. Then our user count grew and the pricing tier jumped. I was looking at $99/month for a feedback board for a product still finding its footing. That's not a moral judgment on Canny — it's a fair product — but for a bootstrapped indie dev, it started feeling like a tax on momentum. The deeper problem with all three solutions was the same: they were inboxes, not loops. User submits → enters the void → user never knows if anyone saw it → user assumes nobody did → trust erodes → churn. I had bui

Canvas first AI agent harness. MCP native. Local first. Discussion | Link

``This is a submission for the GitHub Finish-Up-A-Thon Challenge What I...

A stack de autenticação em .NET fica sólida quando separamos duas responsabilidades: ✅ Argon2id para guardar senhas (hash irreversível, lento, memória-intensivo) ✅ JWT Bearer para provar identidade depois do login ✅ Validação de iss , aud , exp e assinatura em cada request ✅ Segredos fora do repositório (ambiente / Key Vault) Se o ecossistema .NET já oferece hosting, APIs e pacotes maduros, combinar Argon2 (referência da Password Hashing Competition , testável em argon2.online ) com JWT é o caminho natural para microsserviços e Web APIs. Neste artigo, mostro o fluxo registo → login → token → rotas protegidas com foco no que implementar no dia a dia. ⚠️ Observação importante JWT não substitui Argon2. Nunca coloque senha ou hash no payload do token. Argon2 protege a credencial na base de dados; JWT é sessão assinada com expiração. 🧠 Visão Geral Aspecto Argon2 (senha) JWT (sessão) Foco Resistir a offline cracking Autorizar requests após login Onde vive Coluna password_hash na BD Header Authorization: Bearer Algoritmo Argon2id (OWASP) HMAC-SHA256 ou RSA (config) Ferramenta de estudo argon2.online docs Microsoft JWT Bearer Runtime Biblioteca .NET (ex.: Konscious Argon2) Microsoft.AspNetCore.Authentication.JwtBearer Erro clássico MD5/SHA rápido na senha Token sem validar aud / iss 🧩 O que o Argon2 resolve (camada 1) O Argon2 é o vencedor da Password Hashing Competition — hoje a referência para novas passwords . 1️⃣ Hash irreversível com Argon2id var hash = hasher . Hash ( password ); await store . CreateAsync ( email , hash ); ✅ Salt único por utilizador ✅ Parâmetros m , t , p documentados no próprio hash ✅ Verificação com tempo constante ( FixedTimeEquals ) 2️⃣ Calibrar custo com consciência Em argon2.online podes experimentar memory cost e iterations — útil em laboratório. 📌 Em produção usa biblioteca auditada (.NET), não hashes de utilizadores reais em sites públicos. 3️⃣ O que não fazer na senha ✅ Não “criptografar” senha com AES reversível ✅ Não MD5 / SHA-1 / SHA-256

This is a submission for the GitHub Finish-Up-A-Thon Challenge What I Built VKara is a browser-based karaoke room app for singing at home with friends or family. It is not trying to replace YouTube. YouTube is already great at playing videos. It already has almost every karaoke song we need. But YouTube is not really designed to manage a karaoke night where many people want to choose songs together. That is the gap VKara tries to fill. You open VKara on a TV or laptop as the main playback screen. Everyone else joins the same room from their phone using a 4-digit room code or QR code. Then anyone can search for songs, add them to the queue, pause, resume, or skip. The TV only needs to play the video. Everyone's phone becomes their own remote. That is the whole idea. Simple enough to explain in one sentence. Not simple enough to build in one weekend. I learned that part the hard way. Demo Links: Live demo: https://vkara.vercel.app/en GitHub repo: https://github.com/lehuygiang28/vkara Before branch: https://github.com/lehuygiang28/vkara/tree/before Old backend repo: https://github.com/lehuygiang28/vkara-api Small warning: the demo is running on limited resources, so if it is slow, please give it a moment. My wallet is still a student wallet. lol. The flow is: Open VKara on a TV or laptop. Join the room from a phone by code or QR. Search for a karaoke video. Add it to the shared queue. Control playback together. Before: the idea worked, but the product still felt like a video app squeezed into a karaoke use case. After: the mobile flow is now focused on joining, searching, choosing an action, and controlling playback. The Comeback Story I started VKara around early 2025. At that time, my goal was very personal. I wanted a better way to sing karaoke at home with friends. The normal setup was: open YouTube on a TV, search for karaoke videos, and pass control around. It worked, but it was awkward. One person was searching. Another person accidentally played a video immedia

There is a specific failure mode in AI-assisted QA work that most tooling discussions skip entirely, and it shows up earliest when you are working solo on a real engagement. Every new chat session is stateless. You paste the ticket, describe the feature, explain your severity logic, set up the context, and by the time the AI is actually useful, you have rebuilt your methodology from scratch for the third time that week. That is not a workflow problem you fix with better prompts. It is an architecture problem, and the fix is a skill file. QAJourney has a full breakdown of this system at qajourney.net/ai-qa-workflow-for-real-projects, including the actual skill files as free downloads. The short version: a skill file is a context document you load as a system prompt. It carries your test surface tiers, your three-path testing framework, your bug report format, your severity and priority logic, your Playwright conventions, and an explicit definition of what the AI does and does not get to call. Load it once per session. The AI operates inside your methodology from the first message instead of a blank slate. The local LLM layer solves a different problem. On a freelance or retainer engagement, tickets contain real product logic and real client data. Sending that to a cloud API on every session is a data exposure question whether or not it rises to a compliance issue. Running Ollama locally with the same skill file as system context keeps the engagement data on the machine. For the output quality required on QA tasks, current 7B to 14B models are sufficient. The cost at zero marginal per token makes it infrastructure rather than a service you pay by the session. The three-role setup in the workflow: engineer as judgment layer, cloud AI loaded with the skill file for complex reasoning and active session output, local LLM for lightweight tasks and client data work. The skill file is the constant across all three. The part that took time to internalize: AI dev teams already

RecursiveMAS (arXiv 2604.25917) showed that agents sharing internal reasoning state outperform agents that share only final outputs. The average accuracy gain across benchmarks was 8.3 points. The mechanism: each agent passes not just its answer but the latent embeddings from its own reasoning process, and the next agent conditions on both. The paper is a good result. The catch is access. RecursiveMAS requires open-weight models with hidden states exposed at inference time. That rules out Claude, GPT-4o, and Gemini. I built a Claude-native version using the Anthropic extended thinking API. The core idea transfers: instead of passing latent vectors, pass the full thinking text. The paper calls it internal state sharing; the Claude version calls it thinking-block relay. The architecture problem Claude's extended thinking blocks carry an encrypted signature tied to the originating conversation. You cannot pass a signed thinking block into a different agent's messages array. The API rejects it. The workaround: extract the text from the thinking block and inject it as a regular user message. # Extract thinking text from Agent 1 thinking_text = next ( ( b . thinking for b in response . content if b . type == " thinking " ), "" ) # Inject into Agent 2 as regular context, not as a thinking block context = f " Prior agent reasoning: \n { thinking_text } " The signature does not transfer. The reasoning does. relay-structured: what I built first The first architecture was a Planner > Critic > Solver loop where each agent emits a compact mental model JSON instead of raw thinking text. Raw thinking at a 1024-token budget is often compressed and fragmented. The hypothesis was that 150 tokens of structured signal carries more information per token than 1024 tokens of compressed prose. The schema each agent emits: { "interpretation" : "how the agent read the problem" , "key_steps" : [ "step 1" , "step 2" ], "rejected_approaches" : [ "approach tried and discarded" ], "confidence" :