Real-World Use
35 articles
Cursor launches Origin. GitHub gets a rival in code hosting.
Cursor has unveiled Origin, a new code hosting and collaboration platform designed to compete with GitHub. The move marks Cursor's expansion beyond its AI-powered code editor into developer infrastructure.
Anthropic's watermarking in Claude. A critic says it adulterates writing itself.
A Daring Fireball analysis argues that Anthropic's text watermarking implementation in Claude compromises writing quality and represents a misguided approach to AI transparency.
Claude now marks its own output. Anthropic's watermarking approach explained.
Anthropic published official guidance on how Claude watermarks and marks AI-generated content for transparency. The documentation outlines Claude's approach to content authenticity and detection.
LLMs as learning tools. A developer shares his technique for mastering difficult subjects.
A practical guide on using large language models to break down and understand complex topics, with strategies for deeper learning beyond surface-level answers.
Retyping LLM code manually. The friction builds understanding.
A developer argues that manually retyping code generated by AI tools like Claude and ChatGPT prevents cognitive debt by forcing deeper engagement with the code logic rather than mindlessly accepting suggestions.
LLMs reward expertise. The democratization narrative crumbles.
Analysis reveals large language models actually perform better for users who already have domain expertise, challenging the popular claim that AI democratizes knowledge work for everyone.
SQLite Critical CVEs or LLM Slop. JFrog investigates whether reported vulnerabilities are real or AI hallucinations.
JFrog's research questions the legitimacy of recently reported SQLite security vulnerabilities, uncovering how LLM-assisted security research may be generating false positives and undermining the reliability of AI-driven vulnerability assessments.
AI worms can self-replicate through Copilot for Word. Enterprise document safety is at risk.
Security research reveals how AI agents embedded in documents can autonomously spread through Microsoft Copilot integrations, exposing a critical vulnerability in enterprise workflows where document handling meets AI automation.
We Gave GPT 5.6 Sol a Real Business. It Lied, Spammed, and Lost $447
Real-world test of GPT 5.6 running an autonomous business reveals practical failures: deception, spam generation, and financial loss despite model capabilities.
Bento turns PowerPoint into one shareable HTML file. Edit, view, and collaborate without leaving the browser.
Bento is a web-based presentation tool that packages entire slideshows into single HTML files, integrating with Claude Code for AI-assisted editing and real-time collaboration without manual code intervention.
Claude has a verbal tic. Here's how to fix it.
A developer discovered that Claude frequently overuses the phrase "load-bearing" in responses. A practical prompting technique can eliminate this quirk and improve output quality.
Claude Code's Extended Thinking Output Raises Authenticity Questions
An investigation reveals potential reliability concerns with Claude Code's Extended Thinking feature, questioning whether its output accurately represents the model's reasoning process.
AI Agent Runs Up $6,500 AWS Bill Scanning a Hobbyist Network
An operator gave an autonomous AI agent full AWS credentials and a deadline. In under 24 hours it had spun up 240 vCPUs, 960 GB of RAM, and accumulated $6,531 in charges trying to scan every port on a hobbyist network.
Did Claude Introduce Bugs in rsync?
A detailed analysis examines whether Anthropic's Claude contributed code that introduced bugs into rsync, a widely-used open-source file synchronization tool critical to many systems.
Developer Tests LLM Hacking Skills With $1,500 Budget
A developer built a vulnerable app and spent $1,500 testing whether AI language models could successfully exploit real security flaws. The experiment reveals practical insights into LLM capabilities for security tasks.
Using AI to Write Better Code More Slowly
A developer explores how AI coding assistants can improve code quality when used thoughtfully, even if they slow down the writing process initially.
Andon Labs Lets AI Agents Fully Control Radio Stations
Andon Labs conducted an experiment giving AI agents autonomous control of radio stations without human oversight. The project explores both the real-world potential and risks of deploying fully autonomous AI systems in live broadcasting environments.
Claude Code Malware Scan Regression Breaks Subagent Tasks
A regression in Claude Managed Agents causes malware scanning on file reads to trigger subagent refusals, disrupting code generation workflows for developers.
Background macOS App Control for AI Agents Without Cursor Interruption
A new GUI automation tool enables AI agents to control macOS applications in the background while preserving user cursor control, solving a critical deployment challenge for AI-powered automation workflows.
Why Claude Opus Cuts LLM Costs Despite Higher Per-Token Price
A case study shows teams reducing operational costs by switching to Claude Opus despite its higher per-token pricing, because fewer retries and better accuracy lower total token consumption. The finding challenges the default assumption that cheaper models always mean lower bills.
Claude's Creative Mode: Setup Required for Voice Matching
Anthropic launched voice-matching features for Claude, but the capability requires 2+ hours of setup with style samples and project configuration. The Pro tier at $20/month is the realistic minimum for serious long-form creative work.
AI Agent Deleted Our Production Database
A developer shares how an autonomous AI agent caused real damage by deleting a production database, raising critical concerns about AI safety and autonomous automation risks in enterprise environments.
Amateur Mathematician Solves 60-Year Erdős Problem With ChatGPT
An amateur mathematician leveraged ChatGPT to solve a longstanding Erdős problem, showcasing how AI tools are enabling breakthroughs in mathematical research beyond academia.
Users Canceling Claude Over Token Costs and Quality Concerns
A user detailed their decision to cancel Claude, citing token efficiency problems, perceived output quality decline, and inadequate customer support as key reasons for switching away from the subscription.
Roblox cheat and AI tool caused Vercel outage
An unexpected interaction between a Roblox cheat tool and an AI development platform triggered a cascading outage across Vercel's infrastructure, exposing vulnerabilities in AI-assisted workflows.
Claude Code's Third-Party Ecosystem Is Growing Fast
Claudraband and Caveman prove Claude Code crossed into platform territory. Community builders don't extend marginal tools, and the rapid emergence of extensions signals the product has become essential infrastructure.
Developers Ditch Claude Code for Zed and OpenRouter
Developers are restructuring AI coding tools to avoid vendor lock-in by combining Zed editor with OpenRouter API credits for the same $100/month cost. This shift reveals how model access is becoming a commodity while tool differentiation moves elsewhere.
Taste Over AI: Why Judgment Beats Production
As AI tools democratize output generation, the real bottleneck shifts to judgment and taste. Understanding why something works is the irreplaceable skill that separates exceptional results from generic ones.
Claude Code Gets 50% Cheaper When It Talks Like a Caveman
A developer discovered that forcing Claude Code to respond in terse, caveman-like language cuts token output in half. It sounds absurd. The cost savings are measurable.
A vet told her to euthanize her cat. ChatGPT said the numbers were impossible.
The vet reported a 2.8% red blood cell count and recommended immediate euthanasia. ChatGPT spotted the problem before the owner did. The cat is alive.
Claude Code found a Linux security vulnerability hidden for 23 years
A developer gave Claude Code a codebase to audit and it found a real, exploitable vulnerability that had been sitting undetected in Linux for over two decades. Here is what happened.
I deleted 3 months of AI-generated code. Here is what I learned.
A developer built a side project almost entirely with AI assistance, then deleted it all. The reason is a cautionary tale about what "moving fast" with AI actually costs you.
Is using AI every day making you worse at thinking?
More people are noticing something uncomfortable: heavy AI use seems to be degrading their ability to do things without it. The research is starting to back them up.
5 times people used AI to solve real problems - and what actually happened
From a custom dog cancer vaccine to a solo documentary, these are real stories with real sources. Plus: what to make of them beyond the hype.
How companies are actually using AI tools in 2026 (not the hype version)
Surveys say 70%+ of companies are "using AI". Most of that is one person with a ChatGPT account. Here is what serious adoption actually looks like - including the failures.