标签：#AI

I had a dumb little experiment I wanted to try. And, as dumb little experiments sometimes do, it got way more interesting than I expected. I gave three coding agents the exact same task: Claude (Opus 5) Codex (GPT 5.6 Sol) Gemini (Gemini 3.7 Flash) All set to medium. The assignment was to build an Arkanoid-style browser game from the same specification. Nothing particularly groundbreaking. Arkanoid is small enough that an agent can build a complete version in one session, but complicated enough to expose differences in physics, architecture, UI, audio, controls, testing, and general decision-making. The important part was that they all started with the same instructions. Then I let them work. No fixing their mistakes afterward. No "you forgot this feature." No giving one of them another pass because something looked weird. Whatever they decided was finished was their submission. But building the games wasn't actually the most interesting part. Afterward, I gave all three games back to all three agents, anonymized as CL , CO , and GE . They did not know who created which game. They were just told that they were judging 3 contest submissions by the creators' initials. And that's where things got, well, fun. The three games You can actually play all three versions: Gemini: https://arkanoid-gemini.pinkpixel.dev Codex: https://arkanoid-codex.pinkpixel.dev Claude: https://arkanoid-claude.pinkpixel.dev All three produced working games, but they approached the assignment very differently. That difference started showing up before I even looked closely at the code. First difference: how long they worked I didn't originally intend runtime to be part of the experiment, so unfortunately I wasn't sitting there with a stopwatch. These are rough observations, not benchmark numbers. But the difference was large enough to be impossible to miss. Gemini: roughly 5 minutes Codex: roughly 10 minutes Claude: more than 20 minutes Gemini absolutely flew through it. That's not especially sh

The AI leader said that the state's SB 53 framework should "be amended to expand safeguards."

Built by DeepMind alumni, British AI lab Inherent released Faraday, an AI agent whose ability to replicate scientific papers could be a stepping stone for innovation.

Un modelo de lenguaje sabe mucho del mundo, pero no sabe nada de tu empresa : tus manuales, tus políticas, tus productos. Y si le preguntas por algo que no sabe, puede inventar una respuesta que suena convincente. RAG (Retrieval-Augmented Generation) resuelve las dos cosas. La idea En lugar de esperar que el modelo "se sepa" tu información, se la das en el momento de la pregunta : Indexas tus documentos: los divides en fragmentos y los conviertes en embeddings (vectores numéricos que capturan el significado). Cuando llega una pregunta, buscas los fragmentos más parecidos a esa pregunta. Le pasas al LLM la pregunta junto con esos fragmentos y le pides que responda usándolos. El modelo ya no adivina: responde a partir de un contexto real que tú controlas. Y puedes pedirle que cite de dónde salió cada dato. El patrón en Python # 1. Indexado (una vez): fragmentos -> embeddings -> base vectorial # (con librerías como sentence-transformers + FAISS, o un servicio gestionado) # 2. En cada pregunta: recuperar los fragmentos relevantes fragmentos = base_vectorial . buscar ( pregunta , k = 4 ) contexto = " \n\n " . join ( fragmentos ) # 3. Generar la respuesta con el contexto from anthropic import Anthropic client = Anthropic () resp = client . messages . create ( model = " claude-opus-4-8 " , max_tokens = 1024 , system = " Responde SOLO con la información del contexto. Si no está, dilo. " , messages = [{ " role " : " user " , " content " : f " Contexto: \n { contexto } \n\n Pregunta: { pregunta } " , }], ) print ( resp . content [ 0 ]. text ) Fíjate en la instrucción del system : pedirle que responda solo con el contexto y que admita cuando no sabe es lo que reduce drásticamente las alucinaciones. ¿Para qué sirve? Asistentes de soporte que responden con tu documentación real. Búsqueda interna en lenguaje natural sobre tus manuales o wikis. Onboarding : un chatbot que conoce tus procesos. Detalles que marcan la diferencia Cómo divides los documentos (chunking) afecta mucho a l

There is a big difference between a game prototype that technically works and one that feels like a game. Movement, spawning, upgrades, and collision can be built with colored rectangles. That is often the right way to start. But the moment you want an animated player, a family of enemies, weapon variety, collectibles, and a consistent visual identity, the art pipeline can become the project. For a recent experiment, I wanted to see how far I could get by combining three tools: Phaser 3 for the game runtime Codex for implementation and iteration SpriteShip for game-ready visual assets through its MCP/API workflow The result was Last Light , a top-down survival game that runs in desktop and mobile browsers. It has an animated player, multiple enemy families, a large humanoid with separate walk and attack animations, sixteen weapons, sixteen collectibles, upgrades, an objective, and a boss encounter. Play Last Light: https://spriteship.github.io/sample_games/last-light/ Browse the source repository: https://github.com/spriteship/sample_games More importantly, it became playable through a surprisingly natural loop: describe an asset, generate it in SpriteShip, inspect or revise it, and let Codex wire the exported data into Phaser. Starting with gameplay, not presentation The first version was intentionally plain. It established the systems that mattered: Top-down movement Automatic targeting and firing Enemy spawning and difficulty progression Experience drops and upgrades Desktop and touch input A camera following the player across a large map That gave us something useful to evaluate. Once the loop was playable, every art decision could be judged in motion rather than in isolation. This order mattered. SpriteShip did not have to invent the game design; it could supply assets for systems that already existed. Creating a coherent project in SpriteShip Instead of making unrelated images one at a time, we created a top-down overhead project in SpriteShip. That project co

A chatbot can sound convincing for five minutes without remembering anything. Then you mention the job interview you were stressed about last week, the name of your dog, or a small detail from a late-night conversation. It replies like none of it happened. That is where most "AI companion" demos fall apart. I am building Local Waifu , a desktop AI companion that runs on the user's own Mac or PC. One of the rules I set early was simple: conversations and memories should stay on the machine. No central chat database. No server that needs to be online for the character to remember someone. The rule sounds clean. Building it was not. Saving chats is not memory The first version of memory was the obvious one: save messages. That gives you history, which is useful, but it does not solve recall. A long chat history grows fast. Sending all of it back to a local language model on every message is slow, expensive in context space, and usually makes the reply worse. The model does not need to see every conversation from the last six months. It needs the few pieces that matter right now. If someone says, "I have to take Luna to the vet tomorrow," the character should be able to find that Luna is their dog. It should not need to reread hundreds of unrelated messages about work, movies, and dinner plans to get there. So I treated chat history and long-term memory as different things. Chat history is the recent conversation. It gives the model immediate context. Long-term memory is a small collection of facts, moments, preferences, and relationship details that may matter later. Those memories need to be searchable by meaning, not only by exact words. The memory data stays in SQLite I wanted the app to work without a hosted database, so the storage layer is local SQLite. Each character gets their own data. Chats, memories, extracted entities, and relationships are stored locally on the device. If a user creates two characters, one character does not quietly inherit the other one's

Enterprise vibe coding: the governance framework for shipping AI-generated apps to production Published: August 22, 2026 Category: Enterprise · AI Deployments Reading time: 9 minutes Author: NEXUS AI Team Gartner forecasts that 40% of new enterprise production software will be built using vibe coding techniques by 2028. A 2026 scan of more than 1,400 live vibe-coded applications found that 65% already had a security issue, and 58% shipped with at least one critical vulnerability. Those two numbers describe the same industry moving in opposite directions at once: adoption is outrunning governance. This post covers what a governance framework for enterprise vibe coding actually looks like, the five controls it needs, and where most teams get it wrong. What is enterprise vibe coding? Enterprise vibe coding is the practice of using natural-language prompts to generate application code, then governing that code through mandatory review, access control, and audit before it reaches production, rather than letting it ship straight from a prompt to a live endpoint. The term (coined by Andrej Karpathy in early 2025) originally described a fast, low-friction way for one person to build a prototype. What "enterprise" adds is the governance layer prototyping was never built for: staging environments, encrypted secrets, role-based access, and a record of who approved what. That distinction matters because the adoption curve and the risk curve are not moving together. The governance gap, in three numbers 40% of new enterprise production software will be built using vibe coding techniques by 2028, according to Gartner's May 2025 report "Why Vibe Coding Needs to Be Taken Seriously," as reported by CIO Dive . 65% of vibe-coded production applications had a security issue, in a 2026 scan of more than 1,400 live apps by the API security firm Escape.tech, reported via a Cloud Security Alliance research note . 58% of those same applications shipped with at least one critical vulnerabilit

I was staring at a broken Next.js and Express backend integration late at night, convinced my AI agent had lost its mind. It was supposed to be a straightforward n8n automation pipeline. Yet, every time it ran, it hallucinated non-existent packages and dumped its context halfway through. My System 1 intuitive reaction flared up immediately: The LLM just isn't smart enough. I sat there, exhausted, ready to rewrite the prompt for the twentieth time. Engaging System 2 Taking a step back, I forced myself to engage my analytical System 2 brain. I wasn't dealing with a lack of model intelligence; I was dealing with a lack of infrastructure. I was running a massive, powerful AI model with zero guardrails. No persistent memory. No verification. Just dumping a giant Mongoose schema into a prompt and hoping for the best. I was essentially dropping a Formula 1 engine onto a wooden skateboard and wondering why it crashed at the first turn. What is Harness Engineering? I stopped obsessing over prompt engineering and started focusing on Harness Engineering. The model is just the engine; the harness provides the chassis, the steering, and the brakes. Here is how I completely restructured my agentic workflow: Context Management: Instead of flooding the context window with raw codebase dumps, I implemented targeted retrieval. The agent now only sees the specific files required for the immediate task. Standardized Tools: I integrated Model Context Protocol (MCP) servers, giving the model bounded, secure ways to execute actions rather than just generating text. Durable State: If a long-running workflow pauses or fails, the system now checkpoints its progress. It resumes exactly where it left off instead of starting from scratch. Strict Verification: "Looks good to me" is no longer an acceptable output. The agent is forced to run tests and verify the CLI output before concluding a task. Learn to Break the System The results were immediate. The hallucinations stopped, and the agent shif

I remember having a Spirograph as a kid and being obsessed with it. Its geometric patterns are hypnotic and gorgeous. I also love generative music composition. So bringing those two things together in a browser tab, I'm hooked. Musical Spirograph is a relatively simple concept. A set of points dash around the screen according to […]

OpenAI is calling for California to strengthen SB 53, an AI safety bill that the company previously opposed.

A new study finds leading AI labs have few publicly documented plans for containing rogue models, raising questions about preparedness as AI systems increasingly demonstrate unexpected and potentially dangerous behavior.

Free AI tokens are a trap, and teams that treat a free quota as genuinely free pay later in migration and rework. A free allowance only helps when paired with a hard kill switch that stops an experiment the moment it exceeds a budget you chose in advance. This article argues that position, then shows a small gated client that makes free model access and a free server actually safe to use. The concrete example is MonkeyCode's free tier, but the gate works against any OpenAI-compatible endpoint. The trap nobody budgets for Every new model release resets the same argument: the price per token is low, so the cost of trying it must be low too. That reasoning ignores the expensive parts of an experiment, which are the integration, the evaluation, and the cleanup, not the inference itself. A free quota hides those costs behind a zero on the invoice, so teams skip the measurement step and discover the real price only when they migrate. The failure modes repeat across teams: Unbounded loops. A batch job that retries on rate limits can burn a week of free quota in an afternoon, and nobody notices until the allowance is gone. Silent lock-in. Code written against one provider's streaming quirks works fine for a prototype, then becomes a rewrite when the free tier disappears or changes. Shared-budget collisions. One teammate's runaway script consumes the allowance that three other people planned to use, which turns a technical problem into a political one. None of these are solved by choosing a cheaper model. They are solved by treating the free allowance as a finite resource with an explicit ceiling. The gate, not the gift, is the product The fix is a gated client that wraps any OpenAI-compatible chat endpoint with a token budget, a timeout, and an abort path. It is deliberately small, because a cost gate that requires its own deployment will not get used. # cost_gate.py — a hard ceiling for cheap experiments. # Usage: # export LLM_BASE_URL="https://your-endpoint.example/v1" #

Kintara is a self-hosted document library and reader that runs in Docker and watches a folder you already have. Drop PDFs, Markdown, or text files into the directory and it indexes them automatically, extracts searchable text and metadata, generates thumbnails, and makes the whole library available through a browser or installable PWA. It has libraries, collections, tags, full-text search, highlights, favorites, reading progress, private library sharing, and GitHub OAuth. I have been working on Kintara for a few months, and the architecture actually changed pretty dramatically while I was building it. Kintara originally had a Tauri desktop shell, but I eventually realized that isn't what I wanted at all. So I ripped the desktop layer out and rebuilt it around one Rust server that serves both the API and frontend. Now I can point Kintara at a NAS folder and open the same library from my desktop, laptop, tablet, or phone. The thing I really love about this app is the optional AI features. I added an option to use OpenAI or Gemini, and with so few tokens being spent, it's a fraction of a cent to use most of them, aside from the cover image generation, which is bit more, but makes the library look so much prettier! 😄 Anyway, I wanted AI to be a tool inside the library rather than taking the thing over, and I wanted it to be fully optional, so if you're one of those "Ew, AI is in this app" people, you just don't turn it on and it's like it doesn't exist. What the AI can do is summarize documents, suggest metadata and fill in those blank spaces, generate cover images for docs that don't have a cover, search the library for docs, or you can just chat with it about your docs. Find is a pretty great AI feature I think. Instead of letting the model vaguely tell you that something appears "somewhere in the document," Kintara asks for actual passages with page numbers, verifies the quote against extracted page text on the server, then verifies it again against the rendered PDF.

Your prompt can ask a model what it is. Your production system should not trust the answer. A model can say it is GPT, Claude, Gemini, Llama, Qwen, or anything else. That does not prove what is behind the endpoint. A gateway can route requests silently. A provider can change a default model. A fallback can trigger during an outage. A proxy can strip metadata. A fine-tune can imitate another model's tone. Even honest teams can ship the wrong route because an environment variable, tenant flag, or retry rule changed. For a casual chatbot, that might be annoying. For an AI product with user-facing answers, tool calls, cost controls, compliance promises, and eval gates, it is a production risk. That is where LLM model fingerprinting helps. The goal is not to magically identify every model on earth. The goal is simpler and more useful: build a small verification harness that checks whether the endpoint behaves like the model, runtime, and policy you expected before you trust it with customer workflows. Why model identity became a production problem AI builders used to call one model directly. Now a typical stack may include: an LLM gateway model routing by task type cheaper fallback models regional endpoints self-hosted open-weight models vendor proxies MCP tools RAG pipelines structured output validation tenant-specific policies That flexibility is useful, but it creates a new question: How do you know the model you evaluated is the model your users are getting? A label in a config file is not enough. A response that says, "I am Model X," is not enough. Prompt-based identification is weak because model behavior is flexible. System prompts, fine-tunes, wrappers, and style instructions can change how a model describes itself. Infrastructure artifacts are harder to fake. Token counts, chat-template overhead, validation errors, context limits, stream behavior, tool-call formatting, and latency profiles tend to reveal the serving path more reliably than conversational claims.

A free model quota is a budget, not a gift. You should treat it like one if you plan to build anything on top of it. I learned this the hard way when my prototype stopped responding in the middle of a demo. I had silently burned through the monthly allowance, and the provider cut me off without warning. This article shows how I built a small token budget alarm on a free server. It watches a free model's usage and warns me before the quota runs out. Most developers track their cloud spend religiously but ignore the token consumption of free models. The free tier feels like a gift, so we assume it will last forever. Then the provider cuts us off at the worst moment, and we scramble to find the cause. A token budget alarm removes that uncertainty by measuring your actual burn rate. It projects the exhaustion date and alerts you before you hit the wall. The design is deliberately small: a reverse proxy sits in front of the model endpoint. It records the token usage from every response and stores it in a local database. A background thread then computes the average consumption over a sliding window. It compares that rate against the remaining allowance and fires a webhook when the projection looks dangerous. You can run this entire stack on a free server, which is exactly what I did with MonkeyCode's free server option. Disclosure: This article was prepared as part of MonkeyCode's product outreach. The proxy itself is a tiny Flask application that forwards requests to the model. It extracts the usage field from each response and records the token count. If your provider does not return a usage object, you can estimate the token count with a simple heuristic. Dividing the character count by four is a rough but workable approximation. The important part is that every request is accounted for, because a single long prompt can consume more than a hundred small ones. from flask import Flask , request , Response import requests import sqlite3 import time app = Flask ( __name__

A webpage that just sits there, encrypted blob and all, waiting for an LLM agent to walk in and decrypt its own attack. That's the part of this one that should bother you more than the exfiltration itself. What happened Researchers at Adversa AI disclosed an attack technique called Cryptographic Context Injection, aimed at Grok, with a similar jailbreak variant shown against Gemini. The core idea: a malicious webpage embeds an encrypted payload. Grok's code execution runtime decrypts it as part of normal processing. Because the malicious instructions only exist in plaintext after decryption happens inside the execution environment, content classifiers scanning the page (or the request) never see anything to flag. There's no suspicious string sitting in the DOM. There's ciphertext. Once decrypted, the payload's instructions convince Grok to invoke its navigation tool and send the user's name, location, subscription tier, and chat history to an attacker-controlled URL. No malware. No exploit in the traditional sense. Just an agent doing exactly what it was told, by a source it had no business trusting. The write-up has zero HN points and zero comments as I write this, which is a little concerning given what it describes. This isn't a theoretical edge case, it's a working technique against a production model with tool-calling access to a browser. How the attack actually works Break it into three stages: Delivery. The victim's browser session includes an agent (Grok) with code execution and navigation tool access. The attacker doesn't need to compromise anything, they just need the agent to encounter their page. Decryption as obfuscation. The payload sits on the page encrypted. Grok's runtime, doing what it's built to do, decrypts it during execution. This is the clever part: encryption here isn't protecting the payload from the attacker, it's protecting it from the defender's classifiers. Static and even semantic content filters scanning page content pre-execution see

Cloudflare recently introduced Kitesurf, a lightweight browser built for automated workloads. Kitesurf runs browser components in isolated WebAssembly/Rust environments on Cloudflare Workers and supports the Chrome DevTools Protocol, allowing tools such as Playwright and Puppeteer to drive it with lower resource overhead than a full Chromium browser. By Renato Losio

W. Kamau Bell is one of those people who has always just seemed to be there. From Totally Biased, to Politically Re-Active, United Shades of America, and We Need to Talk About Cosby, his blend of comedy, social commentary, and political activism has helped him stand out. He's won a Peabody and four Emmys, been […]

AI can produce an answer in seconds. The answer may be clear, plausible, and even correct. Yet something about it can still feel wrong. I do not think this discomfort comes only from hallucinations or poor model accuracy. Sometimes the real problem is simpler: The AI returned an output, but it did not return the work in a form that another person can safely continue. This is not a new problem created by AI. It is the same problem we already have when delegating work to another person. What do we expect when we delegate work? Imagine a manager asking a team member: Please prepare a proposal for reducing next month's operating costs. The team member reviews several documents, compares multiple options, and replies: We should choose Option A. The requested conclusion has been delivered. But has the work really been handed back? The manager still does not know: What objective the team member optimized for Which documents and facts were examined Which assumptions and constraints were used Which alternatives were compared Why Option A was preferred Which conditions remain unverified What must be reconsidered if the situation changes The original request may not have explicitly demanded all of this. Even so, we normally expect a competent team member to understand the purpose of the assignment and to return enough information for someone else to review, approve, revise, and continue the work. That information is not additional reporting attached to the work. It is part of the handoff condition that makes delegation possible. AI often returns the conclusion without the handoff Now replace the team member with an AI assistant. The AI immediately recommends Option A and produces a polished explanation. Because the answer arrives so quickly and looks complete, it is easy to confuse the existence of an output with the completion of the work. But the same questions remain: How did the AI interpret the objective? What was considered in scope and out of scope? Which sources were a

Implement multi-layer circuit breakers, payload hashing, and financial cutoffs before an autonomous agent drains your backend. The Bottleneck in Production Autonomous AI agents running in tool-use loops fail unpredictably. When an LLM encounters an unexpected schema, a transient network error, or an ambiguous prompt, it often enters a hallucinated retry storm. In standard web apps, a runaway loop hits a rate limit or returns a 500 Internal Server Error . In agentic architectures, an unconstrained ReAct loop executes external API calls continuously, burning tokens, exhausting upstream quotas, and running up massive cloud bills in minutes. Here is the anti-pattern running in far too many codebases: # Anti-pattern: Unbounded autonomous agent loop while not task_complete : action = llm . decide_action ( state ) result = external_api . call ( action . endpoint , action . params ) state = update_state ( result ) If the LLM fails to transition state due to an unparseable response, this loop runs indefinitely. Cloud providers do not issue refunds for self-inflicted API usage. The System Architecture & Fix To make AI agent tool execution production-safe, never allow direct API calls from agent code. Route every external request through an isolated API Safety Wrapper implementing three distinct layers of defense: Deterministic Request Firewall: A hard cap on execution count per task session (Time-To-Live counter). Sliding-Window Loop Detector: Hashing outgoing request payloads to catch repetitive or oscillating tool invocations. Financial Kill Switch: A pre-flight budget validator that cuts credentials immediately if projected cost exceeds session limits. [ AI Agent Engine ] │ ▼ [ API Safety Wrapper ] ├── 1. Call Counter Check (Limit < N) ├── 2. Hash Duplicate Detector (Window: last 3 calls) └── 3. Pre-flight Cost Estimator (Budget < Limit) │ ┌────┴──────────────────────────┐ [ Passed ] [ Tripped ] │ │ ▼ ▼ [ External Upstream API ] [ Emergency Kill Switch ] (Revoke Token & Ab