Moltbook is fascinating, and kind of terrifying
Watching AI agents interact on Moltbook is fascinating, weird, and a bit concerning.
Hilary Gridley's AI Steering Wheel
The AI Steering Wheel transforms vague instructions like "simpler" or "ambitious" into specific, actionable commands that LLMs actually understand.
The Camera You Have on Hand: Using Codex Web to Act on Ideas Instantly
Learn how Codex Web lets you build and deploy projects instantly from your phone—no desk setup required.
Quoting Ethan Mollick: Managing Agents is Really a Management Problem
Managing AI agents requires the same skills as managing people or products, that's why product managers excel at getting results from LLMs.
Quoting Jason Gorman: The Future of Software Development is Software Developers
Have you ever asked an LLM to do the same task in different languages and gotten wildly different results?
From 9 Posts to 47: Building, Breaking, and Learning With AI in 2025
My 2025 review: from building real projects with LLMs to exploring security risks, prototyping, and the philosophical impact of AI on work and creativity.
Quoting Johann Rehberger: The Normalization of Deviance in AI
Discover how the normalization of deviance threatens AI systems (or why companies gradually accept risky shortcuts)
A First Look at Prototyping With Gemini 3 and Google AI Studio
How Gemini 3 and Google AI Studio revolutionize prototyping.
Quoting Elena Verna: My beef with AI credit pricing
Elena Verna critiques AI credit pricing models and urges product teams to rethink how they charge for AI-powered features.
Experimenting with Codex CLI, Agents.md, and PRDs
First impressions using Codex CLI with Agents.md and PRDs to speed product work and code experiments without another subscription.
Quoting Geoffrey Litt: Code Like a Surgeon
Geoffrey Litt compares effective coding with AI support to a surgeon working with a skilled team, staying hands-on with the core.
On Vibe Coding Cleanup as a Service
Reflecting on the business of cleaning up AI-generated code and why vibe-to-production services are becoming lucrative.
Curious & Confused: TIL in July - August 2025
A roundup of my July–August rabbit holes: AI tools, odd ideas, and quick notes captured before they vanish.
Quoting Michael Bassili: I Miss Using Em Dashes
Michael Bassili laments losing em dashes to safety filters and how AI tooling reshapes writing habits for bloggers.
Quoting Bruce Schneier: We Are Still Unable to Secure LLMs
Bruce Schneier argues we still lack defenses against malicious LLM inputs and outlines why current security approaches fall short.
Prototyping a Tag Manager Component with ChatGPT and Cursor
Building a tag manager component for a blog CMS using ChatGPT for UI prototyping and Cursor for implementation, with LLM-powered tag suggestions.
Quoting Anu Atluru: Doomprompting Is the New Doomscrolling
Highlighting Anu Atluru's take on doomprompting—how short, lazy prompts make us passive creators and duller conversationalists.
GPT-5 First Impressions
Testing GPT-5 in Cursor to ship version history features quickly, with thoughts on speed, accuracy, and AI-assisted coding.
What is Slopsquatting?
Explaining slopsquatting—the tactic of registering fake packages that LLMs hallucinate, priming supply-chain attacks.
Quoting Orta Therox: Programming’s ‘Introduction to Photography’ Moment
Orta Therox frames AI-assisted coding as programming's photography moment, where new tools reshape craft rather than replace it.
Quoting Vincent Schmalbach: My LLMs Have Personalities
Vincent Schmalbach personifies his LLMs like quirky interns, comparing ChatGPT, Claude, Gemini, and Grok personalities.
Curious & Confused: TIL in May - June 2025
Notes from May–June explorations: AI agents, security quirks, regulation chatter, and context engineering experiments.
Quoting Simon Willison: Identify, solve, verify
Simon Willison's identify–solve–verify mantra on why humans remain essential to guide, debug, and validate LLM-generated work.
Quoting John Rush: Building a Personal AI Factory
John Rush shares how he builds a personal AI factory with Claude Code, MCP, and agents, mirroring my own coding workflow.
Quoting Simon Willison: The lethal trifecta for AI agents
Simon Willison outlines the lethal trifecta for AI agents—private data, untrusted content, and external communication risks.
Quoting Devansh: Fine-Tuning LLMs is a Huge Waste of Time
Devansh argues fine-tuning LLMs is destructive overwriting. Use RAG, adapters, or prompt engineering instead.
Watching o3 guess a photo’s location is kinda scary
OpenAI's o3 model can identify photo locations—a powerful but dystopian capability that raises serious privacy concerns.
Quoting Drew Breunig on Domain Experts & Developers
Drew Breunig on how AI is flipping the script: coding is becoming commodified while domain expertise becomes the real differentiator.
How secure is MCP, really?
Exploring the security risks of MCP and why it may not be production-ready. Key vulnerabilities include shell access and secret exposure.
The future belongs to idea guys who can just do things
How LLMs are empowering idea people to build and prototype without traditional coding skills—the future of rapid iteration.
The Anthropic Economic Index by Anthropic
Anthropic's new economic index tracks how LLMs impact the economy and labor market, providing data for evidence-based AI regulation.
OpenAI Furious DeepSeek Might Have Stolen All the Data They Stole First
The irony of OpenAI complaining about data theft when it built its company on unauthorized data collection—plus an intro to model distillation.
Your very own Benjamin Gates with ChatGPT
ChatGPT is now surprisingly good at history, noticing details that even experts miss—like having your own Benjamin Gates.
ChatGPT reveals the system prompt for Tasks
Simon Willison reveals ChatGPT Tasks system prompt by getting the model to output its internal scheduling instructions.
Street-fighting RAG: Chain-of-thought prompting
Using chain-of-thought prompting to control LLM responses in constrained game environments and prevent unwanted associations.
Challenging the "LLMs are just next-token predictors" take
Why dismissing LLMs as 'just next-token predictors' misses the emergent intelligence, reasoning, and creativity they develop.
Tracking Flights Over My House With ChatGPT and Claude
Building a flight tracking web app with ChatGPT and Claude—lessons on AI-assisted coding, prompt engineering, and iteration.
Can LLMs write better code if you keep asking them to “write better code”?
Experiment: can iteratively asking LLMs to 'write better code' actually improve output, or does it lead to over-engineering?
Things we learned about LLMs in 2024
Simon Willison's comprehensive review of key LLM developments, breakthroughs, and lessons learned throughout 2024.
OpenAI’s Board: ‘To Succeed, All We Need Is Unimaginable Sums of Money’
John Gruber compares OpenAI to Netscape: leading product but no durable competitive moat in an increasingly commoditized AI market.