A Blank PDF Form Is My New Favorite AI Benchmark
GPT-5.6 impressed me with a 40-hour C++ implementation, while Grok 4.5 surprised me by reconstructing a fillable PDF. Together, they suggest better ways to judge models than one-shot answers.
bhart.org
AI, tech, and personal writing from Bruce Hart.
Latest posts
My Favorite Codex Skill Is a Dumb Little File Clerk
A small personal Codex skill organizes my Google Drive inbox, renames scanned documents, files receipts, and turns HSA expenses into spreadsheet-ready rows.
The /goal Function Made Loop Engineering Click for Me
How Codex /goal, loop engineering, and a weird Starting Lineup Talking Baseball reverse engineering project changed how I think about working with AI agents.
A 4th of July With Baseball, Firmware, and Codex
A July 4 weekend experiment where Codex played an audio-only 1980s baseball game, used ASR as feedback, and helped recover details from its ROM and memory map.
Codex Helped Me Rescue Texts From an Old Android Phone
A personal story about using Codex, ADB, and a manual Android app install to recover a decade-old SMS archive from a phone that could no longer sign into Google services.
I Put Codex Inside My Bedtime Stories App
A bedtime-story emergency turned into an experiment with Cloudflare Workers, Fly.io Sprites, Codex CLI, and the bigger idea of giving apps a controlled place where agents can do real work.
Celebrity Death Bot Is a Memento Mori Machine
Celebrity Death Bot can look weird or morbid from the outside. To me it is closer to an internet-age memento mori: a reminder that life is finite, attention is precious, and every name in the feed points to a real life lived.
AI Code Is Not Slop. Bad Process Is.
AI-generated code gets called slop, but I think that misses the point. Bad code has always been a process failure, and the right response is stronger review, clearer standards, and better security checks.
The AI College Split Is Already Here
Some colleges are treating AI like contraband. Others are treating it like the new internet: allowed almost everywhere, as long as students show their work. I think the second approach probably wins, but the tradeoff is not simple.
I Turned the 2026 NFL Schedule Into JSON in Minutes
A tiny schedule-release experiment with the new Codex Chrome plugin turned an old two-hour HTML parsing chore into a 2-minute JSON extraction.
Codex With Chrome Turns Messy Websites Into Usable Data
The new Codex Chrome workflow makes logged-in, UI-heavy websites much easier to extract structured data from. It does not replace automation scripts everywhere, but it changes the cost of one-off data work dramatically.
GPT-5.5 Is Better at UI When You Give It a Picture
GPT-5.5 still struggles to invent polished UI from text alone, but a visual-first workflow works much better: use GPT Image 2 to create a mockup, then give that mockup to Codex as the implementation target.
GPT-5.5 Feels Like an Incremental Upgrade That Actually Matters
GPT-5.5 has been a solid step above GPT-5.4 in speed, token efficiency, and agent stamina, but that does not mean every API workflow should move. A Codex CLI experiment with Chrome remote debugging and Yahoo Fantasy Football data made the upgrade feel real.
I Built Some SMS Backup Tools So My Texts Could Escape XML
SMS Backup & Restore gives me a real backup, but not a very usable archive. I built sms-backup-tools to load SMS and MMS exports into SQLite, extract MMS images, and hash them so I can move old message media into Google Photos without duplicate chaos.
The AI Race Is Starting to Look Like Logistics
Anthropic's March 2026 changes now look like a full shift from off-peak incentives to explicit peak-hour cutbacks. I still think Opus slightly edges GPT-5.4 on pure feel, but OpenAI looks stronger on the things that increasingly decide who wins: reliability, efficiency, and focus.
I Added AVIF Export to My Browser Image Editor in Under 10 Minutes With Codex
I wanted AVIF export in my Image Editor tool, realized `canvas.toBlob(..., "image/avif")` was not something I could trust across browsers, and had Codex wire up `@jsquash/avif` behind a WASM web worker. From idea to live feature took less than 10 minutes.
NYT Connections Helper Is Live in the Chrome Web Store
I turned NYT Connections Helper from a userscript into a real Chrome extension, added customizable category settings, and got it approved in the Chrome Web Store.
Before Polymarket, I Had a Side Hustle Trading Super Bowl Ticket Futures
In the late 2000s, a startup called Yoonew turned Super Bowl tickets into a strange quasi-prediction market. I built a live pricing model around its stale quotes and learned an early lesson about markets: slow liquidity is an invitation, not a defense.
Most of Life Is Pattern Matching Until It Isn't
The more I learn, the less random the world looks. Stories, people, math, and LLMs all seem to rest on patterns, but the most interesting work happens right where those patterns stop being enough.
I Got Starting Lineup Talking Baseball Running in the Browser
A tiny weekend project turned into one of my favorite uses of Codex yet: take a rough command-line C emulator for a 1988 Intel 8051 baseball game, have Codex adapt it for Emscripten and WASM, debug two nasty lockups, and play a childhood favorite again in the browser for the first time in 30 years.
Partial Minds: Intelligence, God, and the Limits of Access
Working with LLMs has made me less likely to equate intelligence with godlike power. This essay explores AI, theology, and what partial access to reality might mean for human minds too.
GPT-5.4 Is My Daily Driver. Claude Opus 4.6 Is Still My Specialist.
GPT-5.4 is still the model I use most, but this week Claude Opus 4.6 beat it decisively on two tasks that matter: intuitive explanation and last-mile debugging in messy JavaScript environments. The gap between frontier models now feels less like general capability and more like which kinds of hard problems they handle best.
GPT-5.4 Feels Like the Start of the Stability Era
OpenAI's March 5, 2026 release of GPT-5.4 feels less like a one-off model bump and more like evidence that coding is getting close to solved for everyday work. The next frontier looks more like reliability, long-running agents, and systems that can keep learning without falling apart.
What If the Best Models End Up a Little Weird?
Human intelligence is messy, uneven, and sometimes inseparable from cognitive flaws. That raises a strange AI question: if we keep sanding every rough edge off LLMs, are we also sanding off paths to unusual capability?
Google Stitch Is the Most Overlooked Tool in the AI Builder Stack
Google Stitch deserves more attention: it lets you rapidly explore UI directions, then hand those concrete designs to coding agents as real implementation context, especially now that Stitch supports MCP workflows.
Gemini 3.1 Pro Review: Better, But Still Behind Where It Matters
Gemini 3.1 Pro is a clear improvement over Gemini 3 Pro, but in real workflows it still trails on tool calling, long coherence, and feedback incorporation even as it shines in SVG and vision tasks.
GitHub Workflows Turned My README into a Living Homepage
My first GitHub Workflow now auto-refreshes my profile README from my blog RSS feeds every six hours. The setup is simple, but it changed how I think about lightweight personal automation.
I Put Codex Spark in My .bashrc and It Changed My Command Line
I added a tiny Bash helper that uses Codex Spark to turn plain-English intent into shell commands. The speed makes it practical for everyday one-offs, with a few safety guardrails.
Codex Spark: Speed vs Depth
Codex Spark is fast enough to change what I bother automating. The trade is a little less thoroughness on harder tasks, so I now use a two-model workflow.
Seedance 2.0 and the Future of Infinite TV
Seedance 2.0 feels like a preview of a new entertainment era: cheaper, faster video generation that makes it plausible to resurrect formats we loved, with licensing and ethics becoming the real product layer.
Unlocking More with Codex Skills
Treat Codex skills like evolving capabilities, not static tools. When you refine them, script the busywork, and review logs for gaps, Codex stops being a helper and starts feeling like a compounding system.
OpenClaw and the Security Cliff for AI Agents
OpenClaw feels like a preview of where agent tooling is headed, and it also exposes the security cliff we are about to step toward. A few mental models help explain why the current wave is exciting and why it can fail fast without guardrails.
A Word of Caution: Tools Can Leak Secrets
A quick story about a near-miss where automation leaked API keys into GitHub comment history, plus a few mental models and guardrails to avoid the same trap.
Introducing harlite: query HAR files with SQL
Harlite turns HAR browser captures into SQLite so you can analyze web traffic with SQL, optional body storage, deduped blobs, and full text search.
The Hidden Roadmap in Sam Altman's Q&A
Most people heard a Q&A. I heard constraints: how OpenAI wants to price compute, ship agents safely at scale, hide specialization behind one mental model, and turn identity plus memory into the sticky layer. This is my attempt to read the roadmap between the lines.
A HAR-First Pattern for Web Scraping with Codex
Capturing a HAR file and letting Codex trace the requests is my new shortcut for scraping complex sites—faster than guessing and easier to automate.
The Optimization Game Inside AI Labs
AI labs are running a multi-axis optimization: capital, release timing, research direction, and hardware constraints all move together. The result feels less like pure research and more like a strategy sim where the board shifts under your feet.
When Coding Gets Cheap
LLMs are approaching top programmers; the bottleneck shifts from syntax to intent, Jevons Paradox likely drives more software, and software IP moats move beyond code.
Chrome, the Clipboard, and Two Buttons
A practical AI workflow: use Gemini to generate one-off console scripts, and use small TamperMonkey buttons to turn web pages into clean Google Sheets rows and predictable Google Drive receipt filenames (Walgreens + HSA receipts case study).
The Good-Enough Model Era
If open source catches up quickly and inference keeps getting cheaper, model IQ stops being the bottleneck. The next wins come from reliability, memory, context, and distribution - and that shift changes pricing too.
Web Tools: my tiny-tool factory on Cloudflare Workers
I built a single Cloudflare Worker that serves a suite of small web tools. Codex helps me spin up new ones fast, and when a tool needs login + persistence, I bolt on Google OAuth and D1 without turning the whole project into a full app.
Celebrity Death Bot: curation beats detection
A slightly morbid, very practical bot that watches Wikipedia’s recent deaths, uses an LLM to decide what’s actually notable, and sends my family the updates that matter.
LLM Coding Power Rankings (January 2026)
An unscientific, Bill Simmons–style power ranking of the LLMs I actually code with in January 2026.
Introducing Codex Transcripts
codex-transcripts converts Codex CLI session JSONL files into clean, mobile-friendly HTML transcripts (with pagination, search, and optional Gist sharing). Inspired by Simon Willison’s claude-code-transcripts.
The agent hit a content filter (and fixed it anyway)
Codex built a bedtime-story media API in under 30 minutes—then diagnosed an "Inappropriate Content" error and fixed the pipeline by regenerating the image.
Cloudflare Workers, Practically: Shipping More, Babysitting Less
I’m using Cloudflare Workers for this blog and my small personal projects because it’s cheap, global by default, and removes most of the annoying parts of deployment. When infrastructure stops demanding attention, shipping feels lightweight again and building becomes fun instead of fragile.
Hello, Internet (Again)
Starting a blog in 2026 as a time capsule, a public workshop, and a way to connect while AI rewrites the rules.