Comparing PDF readers for my finance app
I compared eight readers on 78 financial PDFs. The results show what I still need to check before trusting my capital gains estimates.
AI
Writing on AI, newest first.
43articles in this collection2025–2026
1 min
Let's talk about the $200 ChatGPT Pro plan
Promotions and resets made 20× feel like what $200 bought. The allowance can shrink on schedule while the expectation outlasts the offer.
Dot is helpful, but not clever enough yet
Dot made airline miles easier to understand. It also made me ask what it takes to trust a personal agent with work beyond a chat.
Building a San Francisco map feels a bit like cheating
I’ve always wanted a small, animated SVG map of just San Francisco. The city’s open data was already there. Codex made the rest feel easy.
The AI Slowdown Protects the Labs Already Ahead
Anthropic says frontier AI is dangerous but keeps racing. Regulation makes its competitors pay for the slowdown it will not take alone.
What’s There to Learn from WordPress?
The WP Engine fight, the Valkey fork, and agent-written code raise a harder question: what keeps an open-source project honest?
Choosing When Hermes Agent Automations Need an LLM
I use Hermes Agent to interpret and configure automations, then let Home Assistant or a script execute repeatable actions with visible failure states.
AI Pricing Is Becoming UX
GPT-5.6 turns capability into a price ladder, while Codex Spark turns speed into a separate button. AI pricing is becoming part of the interface.
Claude’s Guardrails End at Claude
Closed-model guardrails remain tied to the provider, even as comparable open weights spread. Public safety needs controls that survive that reality.
AI-Generated Code Review: What Reviewers Should Still Block
Models make awkward code cheaper to replace. Review should block future risk and leave private ideas of neat to automated conventions.
The Best Coding Agent Is the One You Can Keep Using
Grok Build produces useful patches, but its weekly limit changes how I use it. The harness and quota now matter more than the model.
Learning Rust with Claude Code Through Local Hardware Projects
Claude Code lowers the cost of being a Rust beginner. It kept two local-hardware projects moving while I learned the language inside them.
GLM Is Cheap. Steering It Isn't.
A lower token price pushed me toward GLM, but slow responses and manual mode switches made the work harder to steer.
I Can't Avoid the Price Gap Anymore
My company merged separate AI tool budgets into one shared pool. That's when Opus versus GLM stopped being an abstract question.
Tripsy Now Has a CLI (and an MCP)
Tripsy released an official CLI and MCP server in May 2026. The CLI is the better tool. It replaces the unsupported SQLite writes I used before.
How to Connect Hermes Agent to Home Assistant
Two environment variables and a gateway restart. I said 'turn off all my lights' and they went off two seconds later.
Document Your Quirks: Why CLAUDE.md Changes Everything
I spent 45 minutes teaching Claude a fix I already knew. Putting that project quirk in CLAUDE.md stopped the same rediscovery next time.
Origin Made Me Review Work Copilot Already Handles
I tried Origin as an alternative to Copilot Money. More financial products did not make up for the transaction review it handed back to me.
October's Layoff Numbers Did Not Prove AI Did the Work
October's report counted 31,039 cuts attributed to AI. It recorded what employers said—not whether deployed systems actually replaced the work.
One Claude Code Install Is Enough
My machine has one Claude Code executable and one updater. Here is how I check the path, remove conflicts, and keep it that way.
Meta is Cutting 600 AI Roles to Speed Up Decisions
Meta cut roughly 600 AI jobs while still hiring for TBD Lab. I read the reorganization as a shift in priorities, not an AI spending retreat.
ChatGPT Atlas Automates the Easy Part of Browsing
After using Comet for coupons, search, and summaries, I am not convinced putting an agent in the browser solves the hard part: judgment.
Claude Code Comes to the Web and iOS
Claude Code on the web changes coding from an editor interaction into delegated work. That matters more than adding another interface.
Sora Isn't the Problem: It's the Mirror
Using Sora in a TikTok-style feed made social media's incentive problem hard to ignore: fabricated video is now cheaper than recording reality.
Samsung's Tiny Model Made Scale Look Wasteful
Samsung's 7-million-parameter TRM beat much larger models on specific reasoning benchmarks by revising its answer up to 16 times.
Claude Haiku 4.5: Cheap and Fast
I use Claude Haiku for small refactors, review, and debugging. Its lower cost and faster responses matter more there than deeper reasoning.
What Is a Website For When AI Gives the Answer?
AI search is turning websites from destinations into raw material. The web may survive that transition, but human voices will be harder to find.
Anthropic's Fair-Use Win Did Not Excuse Its Pirated Library
The court treated AI training as transformative while leaving Anthropic exposed for keeping a library of pirated books.
We Rebuilt the UI Instead of Refactoring It
A new data-model entity touched several old pages. Rebuilding those UI components was faster than carrying their Flow types and old assumptions forward.
AI Is the New Backdoor Layoff
Companies can shrink teams without announcing layoffs. They call it AI efficiency, lower hiring needs, and let attrition do the rest.
Windsurf Simplified Pricing While OpenAI Circled
Windsurf removed flow action credits and cut team pricing while OpenAI acquisition talks were being reported. I cared more about the simpler bill.
GPT-4.1 Needed Fewer Retry Loops in My Editors
GPT-4.1's benchmark jump showed up in my Windsurf and Cursor use as cleaner diffs and fewer back-and-forth retries.
The Small Models Were the Interesting Part of OpenAI's Leak
A leaked OpenAI lineup showed five models, but the mini and nano pricing mattered more to me than another flagship.
ChatGPT Came Within 30 Million Visits of X
Similarweb March 2025: ChatGPT at 4.55B visits, X at 4.58B. ChatGPT grew faster that month. Infrastructure, not a toy.
GitHub Copilot Adds Agent Mode and Claude 3.5 Sonnet
Copilot shipped agent mode and more models. Microsoft also tightened extension licenses so some tools refuse to run in Cursor.
AI Will Make the Digital Divide More Expensive
AI access is becoming another way to buy time, better service, and economic leverage. My son will grow up inside that divide.
Claude Code's Strengths and Weaknesses in March 2025
Claude Code for hobby refactors, Cursor at work — two tools, two jobs, and where each one still breaks.
How Claude Code Built This Blog - Part 2
Claude Code helped build the parser, generator, templates, and Cloudflare deployment. Syntax highlighting and timezones took more work.
How Claude Code Built This Blog - Part 1
I paid about $10 to see whether Claude Code could scaffold the TypeScript static-site generator behind this blog. Most of its choices held up.
Google Keeps Replacing Products Before Trust Can Compound
Replacing Assistant with Gemini fits Google's long pattern of abandoning products instead of earning trust through steady maintenance.
Public Code Gives AI More Context to Work With
Public libraries give coding agents examples, documentation, and discussions they can retrieve. Private APIs must supply that context themselves.
A Reported $20k AI Agent Is Not Cheap Labor
A 2025 report put specialized OpenAI agents at $2k–$20k a month. Even as a reported figure, that belongs in a labor budget—not a footnote.
Perplexity replaces Google Search and Apple News for me
I use Perplexity for most search and for a daily news skim. Google for the stubborn queries. Apple News lost the paywall fight.