标签：#ai

It's important to be aware of disc rot, but there's no need to be afraid.

Deezer said more than 90,000 AI-generated tracks were uploaded daily on the platform in June.

I first slipped on Halliday's original smart glasses at CES 2025. I was not a fan. The glasses had a tiny, movable display window embedded into the frame that was incredibly finicky to see, and my 30-minute demo left me with achy eyeballs. It was an interesting concept with terrible execution, especially compared to the […]

Nintendo does a lot of weird things, but Splatoon might just be one of the strangest - or at least the boldest. The series started as an attempt to muscle into the ultra-competitive online shooter space, but in a way that feels a world away from the violence and bombast of Call of Duty or […]

Building AI Agents That Don't Hallucinate: Structured Workflows, Guardrails, and Per-Step Evaluation How we replaced fragile prompt chains with typed schemas, validation gates, and evaluation at every step — 94% task success vs 60% baseline The Prompt Chain Trap January 2024. We built a "research agent" — 12 prompts chained together: Decompose question → 2. Search planning → 3. Execute searches → 4. Extract facts → 5. Synthesize → 6. Fact-check → 7. Format → ... It worked 60% of the time. The other 40%: Step 3 returned malformed JSON → Step 4 crashed Step 5 hallucinated citations → Step 6 missed it Step 7 output wrong format → Downstream consumer failed No visibility into which step failed Debugging meant reading 12 LLM calls' worth of logs. Adding a step broke three others. The Shift: Agents as Typed Workflows We moved from prompt chains to structured workflows with: Pydantic schemas for every step input/output Guardrails that validate and auto-retry Explicit state machine (not implicit chaining) Evaluation harness per step (not just end-to-end) ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ Decompose │──▶│ Search │──▶│ Extract │──▶│ Synthesize │ │ Question │ │ Planning │ │ Facts │ │ Answer │ │ │ │ │ │ │ │ │ │ In: Query │ │ In: Plan │ │ In: Results │ │ In: Facts │ │ Out: SubQ[] │ │ Out: Steps │ │ Out: Fact[] │ │ Out: Answer │ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │ │ │ │ ▼ ▼ ▼ ▼ [Schema] [Schema] [Schema] [Schema] [Guardrail] [Guardrail] [Guardrail] [Guardrail] [Eval: 0.9] [Eval: 0.85] [Eval: 0.9] [Eval: 0.95] Core Abstractions # agent_eval/schemas.py from pydantic import BaseModel , Field from typing import Literal , Any class DecomposeInput ( BaseModel ): user_query : str context : dict = Field ( default_factory = dict ) class DecomposeOutput ( BaseModel ): sub_questions : list [ str ] = Field ( min_length = 1 , max_length = 5 ) requires_tools : bool reasoning : str class PlanInput ( BaseModel ): sub_questions : list [

AI agents no longer just answer questions — they book meetings, approve refunds, call APIs, update records, and chain together dozens of small decisions into a single outcome. Most of the time, this works quietly and well. But when something goes wrong, a troubling question surfaces: why did the agent do that? For a large share of deployed systems today, there's no good answer. The reasoning, the data consulted, the tools invoked, and the intermediate steps simply weren't recorded. This is the problem of undocumented agent decisions — and it's becoming one of the central risks of the agentic AI era. What "Undocumented" Actually Means An undocumented decision isn't necessarily a bad one. It's simply one that can't be reconstructed after the fact. In practice, this shows up in a few common ways: Output without reasoning. The system logs what the agent did, but not the chain of thought, tool calls, or data sources that led there. Ephemeral intermediate state. Multi-step agent chains often discard the "scratch work" between steps — the very material that would explain a decision — once the final output is produced. Reviewer blind spots. A human approves a final recommendation without ever seeing the reasoning that produced it, which looks like oversight but isn't meaningful oversight. No durable storage. Logs exist for a few days or weeks, then age out — so when someone asks for the record months later, it's gone. The common thread is a gap between acting and accounting for the action. This Is Becoming Urgent A few forces are converging to make this problem harder to ignore: Agents are doing more, autonomously. As agents move from single-turn assistants to systems that independently call APIs, touch databases, and trigger downstream workflows, the number of undocumented micro-decisions multiplies. A single customer request might now involve a dozen internal steps, each a potential decision point. Incidents are already happening. Surveys of enterprise AI deployments in 2

AI coding tools have gotten very good at one thing: generating code fast. What they haven't gotten good at is discipline. They don't know your architecture. They don't remember that you rejected a pattern last sprint. They don't know which parts of your context window are signal and which are noise. And they have no concept of a development workflow — no phases, no review gates, no verification steps. You describe what you want, they generate, and you hope the output fits. For small tasks this works fine. For anything that touches your real codebase at scale — refactors, new features with cross-cutting concerns, compliance-sensitive changes — the lack of workflow structure creates subtle, expensive problems that compound over time. We've been building Ortho to address two of these problems: workflow discipline (through ASES, a 6-phase AI development methodology) and context discipline (through a 9-component token optimization pipeline). This post explains both. The workflow problem with AI coding tools When a junior engineer joins your team, you don't just hand them a task and say "generate." There's a process: understand the codebase, plan the change, get the architecture reviewed, build it, test it, verify it, get it reviewed. The process exists because individual steps catch different categories of mistakes. AI coding tools collapse all of that into one step. You prompt. You get code. Done. The result isn't always bad code per file. The problem is architectural: the AI has no model of your layer boundaries, so it imports from layers it shouldn't touch. It has no memory of past decisions, so it re-proposes patterns you've already rejected. It has no verification step, so it confidently generates code that looks right but has subtle issues a reviewer would have caught in 30 seconds. The missing piece isn't a smarter model. It's a workflow. ASES: A 6-phase workflow for AI-assisted development ASES (v1.2) is the methodology built into Ortho's orchestration layer. It

It’s a lot : According to information obtained by The Tech , MIT is spending over $3 million on more than 500 AI surveillance cameras in academic buildings, residence halls, and outdoor areas along Memorial Drive. Installation of the new cameras, along with the wiring and infrastructure that will support them, began November 2025 and will likely continue until September 2026. Technical specifications for the cameras suggest that they will be capable of collecting real-time face and object classification data, including detection of motion, loitering, crowds, face masks, and camera tampering. Individuals can also be automatically classified on the basis of clothing color, gender, and age, up to a distance of 35 feet (11 meters) from the camera. According to a statement from MIT spokesperson Kimberly Allen, any collected data is “retained up to 30 days,” unless an exception is granted...

Bhavuk Jain discusses translating foundational AI into scalable mobile products. He shares the engineering challenges behind AI Wallpapers and Circle to Search, detailing how to implement robust runtime guardrails, fine-tuning, and seamless OS integration. For engineering leaders, he explains balancing UX constraints with model latency and infrastructure cost to deliver safe, reliable AI. By Bhavuk Jain

Gritt is coming out of stealth with $34 million and plan to automate the hardest tasks on construction sites.

Yelp has launched Training Orchestrator. This new internal framework replaces individual team Spark training scripts. Now, it uses a configuration-driven, DAG-based execution model. By Claudio Masolo

You have built an AI agent harness. It calls tools, routes requests, and returns results. Your team trusts its telemetry to tell you which tool was selected and why. That trust is a liability. An agent's own logs are self-reported. They tell you what the agent thinks it did, not what actually happened. A hallucinated tool name, a misrouted parameter, a silent fallback to a different function — none of these surface in the agent's own trace. You need an external witness. Here is how to build one. The Problem: Self-Reported Truth Is Not Truth Most teams validate agent behavior by reading the agent's own output. They check the tool_calls field in the response, match it against an expected schema, and call it done. This works until it doesn't. Consider a common failure mode: the agent decides to call search_knowledge_base but the LLM formats the tool name as searchKnowledgeBase . The routing layer silently normalizes it, the call succeeds, and the agent logs search_knowledge_base . Your test passes. The actual execution path was different from what you verified. Another pattern: the agent selects the correct tool but passes a parameter that the tool silently coerces. A date string gets parsed into a different timezone. A user ID gets truncated. The tool returns a result, the agent logs success, and your test never catches the drift. The root cause is the same. You are testing the agent's intent , not its execution . Intent is cheap to fake. Execution leaves fingerprints. The Solution: An External Observer You need a layer that sits between the agent and the tools it calls. This observer records every invocation — tool name, parameters, response, latency — without the agent knowing it is being watched. The observer does not trust the agent's logs. It trusts what it sees on the wire. Here is the architecture at a high level: Intercept every outbound call from the agent to a tool. Record the raw request before any normalization or routing. Compare the recorded call against

How I Built a Full-Stack Quality Skill for AI Coding Agents AI coding agents are getting very good at writing code. But I kept running into the same problem: They can move fast, but without strong project rules they can also create messy architecture, duplicate utilities, inconsistent APIs, weak security checks, and frontend components that slowly drift away from the design system. So I built Full-Stack Quality Skill . It is a reusable AI coding skill for full-stack audits, architecture guidance, long-term project memory, and CI quality gates. Repo: https://github.com/lablnet/full-stack-quality-skill Website: https://skills.lablnet.com Why I Built It When I use AI agents like Cursor, Codex, Claude Code, Antigravity, or similar tools, I do not only want them to "write code". I want them to think like a careful senior engineer: Is the database normalized correctly? Are backend layers clean? Is business logic leaking into controllers? Are frontend components consistent? Are Vue components using composables? Are React components using hooks correctly? Are HTTP methods and status codes right? Is GraphQL safe from N+1 problems? Are security and privacy risks checked? Are tests missing for critical paths? Is documentation still matching the code? That is a lot to remember every time. So instead of repeating the same instructions in prompts, I turned them into a reusable skill. What It Covers The skill includes audit areas for: Database Backend Frontend Mobile HTTP APIs GraphQL Security Privacy Accessibility i18n Analytics Background jobs Infrastructure Testing Performance Observability Delivery / CI Multi-tenancy Payments Notifications Data import/export API compatibility Developer experience AI/LLM safety It also includes examples for common stacks: Node.js / TypeScript Python Django Laravel Java / Spring C# / ASP.NET Core Go Ruby on Rails React Next.js Vue Angular SvelteKit Flutter React Native Kotlin / Android Swift / iOS SQL GraphQL Read-Only Audits by Default One impo

When people think about delays in Building Management System (BMS) projects, they usually blame installation issues, communication failures, or commissioning problems. In reality, many delays begin much earlier. They start during engineering. Before a single controller is installed, engineering teams spend significant time reviewing I/O lists, selecting controllers, designing panels, preparing wiring documentation, planning network architecture, and coordinating procurement. These activities are essential, but they are also repetitive, manual, and prone to errors. As modern buildings become larger and more connected, traditional engineering workflows are struggling to keep up. The Hidden Cost of Manual Engineering A typical BMS project may contain hundreds or even thousands of points: Temperature sensors Humidity sensors Pressure transmitters VFD controls Damper controls Pump status points AHU controls Chiller interfaces Each point must be reviewed, categorized, mapped, documented, and connected to the correct controller. While this process is necessary, it creates a bottleneck that often goes unnoticed. A small mistake in controller sizing or wiring documentation can trigger a chain of revisions, procurement changes, and commissioning delays. The result is a project schedule that slowly expands before installation even begins. Why Traditional Workflows Don't Scale The challenge isn't engineering knowledge. The challenge is repetition. Engineering teams repeatedly perform similar tasks across projects: Reviewing I/O schedules Selecting controllers Allocating points Generating documentation Creating wiring drawings Verifying network configurations As project complexity increases, the amount of repetitive work increases as well. This leads to: Longer engineering cycles Increased project costs More documentation reviews Greater risk of human error The Shift Toward Engineering Automation Many industries have already embraced automation in design and manufacturing. Build

Building AI Agents That Don't Hallucinate: Structured Workflows, Guardrails, and Per-Step Evaluation How we replaced fragile prompt chains with typed schemas, validation gates, and evaluation at every step — 94% task success vs 60% baseline The Prompt Chain Trap January 2024. We built a "research agent" — 12 prompts chained together: Decompose question → 2. Search planning → 3. Execute searches → 4. Extract facts → 5. Synthesize → 6. Fact-check → 7. Format → ... It worked 60% of the time. The other 40%: Step 3 returned malformed JSON → Step 4 crashed Step 5 hallucinated citations → Step 6 missed it Step 7 output wrong format → Downstream consumer failed No visibility into which step failed Debugging meant reading 12 LLM calls' worth of logs. Adding a step broke three others. The Shift: Agents as Typed Workflows We moved from prompt chains to structured workflows with: Pydantic schemas for every step input/output Guardrails that validate and auto-retry Explicit state machine (not implicit chaining) Evaluation harness per step (not just end-to-end) ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ Decompose │──▶│ Search │──▶│ Extract │──▶│ Synthesize │ │ Question │ │ Planning │ │ Facts │ │ Answer │ │ │ │ │ │ │ │ │ │ In: Query │ │ In: Plan │ │ In: Results │ │ In: Facts │ │ Out: SubQ[] │ │ Out: Steps │ │ Out: Fact[] │ │ Out: Answer │ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │ │ │ │ ▼ ▼ ▼ ▼ [Schema] [Schema] [Schema] [Schema] [Guardrail] [Guardrail] [Guardrail] [Guardrail] [Eval: 0.9] [Eval: 0.85] [Eval: 0.9] [Eval: 0.95] Core Abstractions # agent_eval/schemas.py from pydantic import BaseModel , Field from typing import Literal , Any class DecomposeInput ( BaseModel ): user_query : str context : dict = Field ( default_factory = dict ) class DecomposeOutput ( BaseModel ): sub_questions : list [ str ] = Field ( min_length = 1 , max_length = 5 ) requires_tools : bool reasoning : str class PlanInput ( BaseModel ): sub_questions : list [

Two posts ago I said I'd shipped code I never read. It's easy to write that as a general observation about the industry. It's less comfortable to go open your own repo and count. So I did. Here's what I found, and what I've actually done about it. The repo It's a desktop app I built for my own company. Roughly fourteen thousand lines of Python, a Tkinter UI, invoicing and documents and reports, the kind of internal tool nobody else will ever see. I had never written Python before I started it. I built it anyway, with a lot of AI help, learning as I went. That combination - no prior experience, heavy AI assistance, a real deadline because the business actually needed the thing - is basically a diff debt factory. I wasn't cutting corners on purpose. I just didn't have the knowledge to evaluate half of what I was merging, and it worked, so I moved on. What I actually found The clearest example took me months to notice. Five different parts of the app generate PDFs. Quotes, proformas, reports, petitions, heat treatment certificates. Every one of them was failing, in different ways, at different times, and I kept fixing them individually. Different error, different module, different patch. The actual cause was one thing: LibreOffice wasn't installed on the machine. Every one of those modules quietly depended on it to do the document conversion. Not one of them said so anywhere. I'd merged that dependency five separate times without ever registering that I'd taken it on. That's diff debt in its purest form. The code wasn't messy. It wasn't badly written. It just made an assumption I'd never read, and I paid interest on it five times over before I understood the principal. There were smaller ones too. A path handling bug that only showed up because my Windows username has a Turkish character in it - the code assumed ASCII and nobody, including me, had thought about it. Two Python versions installed side by side, quietly fighting. A stray quotation mark in my system PATH th

I spent a week using video editing Skills to build a video editing Agent. It feels amazing! It can automatically edit a 30-minute video in just 10 minutes. Video editing Agent demo: automatically editing a 30-minute video in 10 minutes. I often use CapCut to edit talking-head videos, but after using it for a long time, I found several problems. Problem 1: Smart talking-head editing does not understand meaning Because it cannot understand the meaning, it sometimes fails to identify repeated sections. If I speak continuously for 20 or 30 minutes, editing the video myself becomes exhausting. Problem 2: The subtitle quality is poor The automatically generated subtitles contain many incorrect words and typos. So I used the Skills feature in Claude Code to build a video editing Agent. The fundamental difference is simple: CapCut vs. Agent: a fixed tool vs. an adaptive assistant. The key difference is: CapCut = fixed tool + manual operation Agent = adaptive system + automatic learning I am not replacing CapCut with a better algorithm. I am replacing it with a system that can continuously improve itself. But that is not even the most impressive part. The most impressive part is this: the more I use it, the better it understands me, and the faster it becomes. Three Core Designs 1. Agent Logic It only takes four steps. Video editing Agent workflow: from the video file to the final video. 2. The Skills System At first, I put every function into one large Skill. I had to add instructions to distinguish between different tasks, which was very inconvenient. Now I have separated the five core video editing tasks into five independent Skills and placed them in the .claude/skills/ directory. This makes the structure clearer and the tasks easier to select. When I enter /v , Claude Code automatically lists the five available Skills. The list of five independent Skills. I select one, and the AI runs that Skill. Simple, right? A manual task that used to take 10 minutes now only requires

Remember Tamagotchis? Those little egg-shaped keychains from the 90s with a pixelated pet that got hungry, got sad, and, if you were a negligent eight-year-old like me, got dead during math class. I've been looking for an excuse to play with the Waveshare ESP32-S3 Touch AMOLED 1.8" , a $30 board with a gorgeous 368×448 AMOLED touchscreen, an accelerometer, a microphone, and a speaker. And I found one: I'm building a Tamagotchi. But with a twist that makes it worth a four-article series The pet's body lives on the device. Its brain lives in AWS. The device renders an adorable pet and captures your taps and shakes. But its hunger, mood, and energy are rows in DynamoDB. It gets hungry overnight because EventBridge Scheduler says so. And, my favorite part (and because it's 2026, we can't not have genAI in a project 🤪), every morning it fetches the AWS news, has Amazon Bedrock rewrite them in its own squeaky personality, and reads them to me out loud through Amazon Polly. A Tamagotchi with a job. In this series, I will walk you through the whole build: This article : the architecture, and connecting the ESP32-S3 to AWS IoT Core (the pet's nervous system). Part 2 : giving it a face: animations and touch with LVGL on the AMOLED. Part 3 : the serverless brain: Lambda, DynamoDB, and a pet that gets hungry while you sleep. Part 4 : Bedrock + Polly: my Tamagotchi reads me the AWS news. ⚠️ Reality check before you get too excited (sorry 😅): this is a hobby project, not a product. You'll need the specific Waveshare board (or the patience to adapt the code to yours), an AWS account, and a USB-C cable. The AWS bill for the whole series is well under a dollar a month, but it is not zero, and neither is the time you'll spend staring at idf.py monitor . Worth it, though. All the code lives in this repository . Each article has its own branch. For this one, git checkout article-1 . Why put a pet's brain in the cloud? Fair question. The original Tamagotchi ran on a chip with less power

Optimizing RAG at Scale: Chunking, Retrieval, and the Bayesian Search That Cut Latency 40% How we moved from "semantic search + hope" to a measured, tunable retrieval pipeline with 95% recall@10 The RAG Reality Check Everyone ships RAG the same way: chunk by 512 tokens, embed with text-embedding-3-small , top-k=5, stuff into context. It works for demos. Then you hit production: Legal contracts: 512 tokens splits clauses mid-sentence API docs: 1000-token chunks drown signal in noise Customer tickets: Conversational context needs overlap, not fixed windows Latency: 500ms embedding + 200ms vector search + 300ms LLM = 1s+ per query We rebuilt our retrieval layer from first principles. Here's what actually moves metrics. Chunking: One Size Fits None # rag/chunking.py from abc import ABC , abstractmethod from dataclasses import dataclass @dataclass class Chunk : text : str metadata : dict token_count : int chunk_id : str class ChunkingStrategy ( ABC ): @abstractmethod def chunk ( self , document : str , metadata : dict ) -> list [ Chunk ]: ... class FixedTokenChunker ( ChunkingStrategy ): """ Baseline. Good for homogeneous content. """ def __init__ ( self , chunk_size = 512 , overlap = 50 ): self . chunk_size = chunk_size self . overlap = overlap class RecursiveChunker ( ChunkingStrategy ): """ Respects structure: markdown headers, code blocks, paragraphs. """ def __init__ ( self , separators = [ " \n ## " , " \n ### " , " \n\n " , " \n " , " " ], chunk_size = 512 ): self . separators = separators self . chunk_size = chunk_size class SemanticChunker ( ChunkingStrategy ): """ Uses embedding similarity to find natural boundaries. """ def __init__ ( self , model = " text-embedding-3-small " , threshold = 0.7 ): self . model = model self . threshold = threshold class AgenticChunker ( ChunkingStrategy ): """ LLM decides boundaries. Expensive but highest quality for complex docs. """ def __init__ ( self , model = " gpt-4o-mini " ): self . model = model Our production config by

The short answer is: while OpenAI Sora offers unmatched visual quality and physics rendering, it remains restricted behind a paid subscription structure. For creators who want a completely free, unlimited AI video generator, the newly released open-source LTX Desktop app by Lightricks allows you to run the LTX-2.3 video model locally on your own computer with zero usage costs or filters. The AI Video Paywall Frustration If you have tried building AI video content for YouTube, TikTok, or social marketing, you know how expensive it is. Platforms like Runway Gen-3 and Luma Dream Machine charge by the second. A simple five-second clip can cost up to fifty cents in API credits, making creative experimentation almost impossible for solo developers. OpenAI Sora is a powerhouse, but its high computational overhead means it will likely remain a premium, paid tool for the foreseeable future. To bypass this, our dev team set up the new open-source LTX Desktop application on our local workbench to see if local video generation is actually viable for production. Here is our hands-on review. |---|---|---|---| | OpenAI Sora | Closed Cloud | None (Paid Plan) | Cloud-Only | Cinema-grade physics, long multi-action shots. | | Runway Gen-3 | Closed Cloud | Daily Free Credits | Cloud-Only | Cinematic camera pans, high texturing quality. | | Wan2.1 | Open Weights | Free Hugging Face Spaces | 16GB VRAM (Local) | Photorealistic human movement, natural lighting. | | LTX-2.3 | Open Weights | LTX Desktop (Free) | 8GB VRAM (Local) | Fast generation speeds, local desktop interface. | Running Video Models Locally: The LTX Desktop Solution LTX Desktop, developed by Lightricks, is a standalone, open-source desktop application that lets you run their LTX-2.3 video generation model on consumer-grade graphics cards. Why LTX Desktop is a Game-Changer Low VRAM Footprint: Unlike HunyuanVideo or Wan2.1 which require massive 16GB-24GB VRAM cards to compile locally, LTX-2.3 is highly optimized and runs com