bhart.org

AI progress in mathematics raises a question I cannot shake: if machines eventually reason beyond our best minds, will their discoveries still make sense to us?

A few days with Claude Fable changed how I think about AI compute: capable local models could handle everyday work while large clusters make sustained investigations affordable.

Codex and GPT-5.6 helped me investigate a device I own and build a local browser interface after the official web option disappeared.

WSL, herdr, and a custom tmux usage meter help me manage several coding agents without losing track of the work.

GPT-5.6 impressed me with a 40-hour C++ implementation, while Grok 4.5 surprised me by reconstructing a fillable PDF.

A small personal Codex skill organizes my Google Drive inbox, renames scanned documents, files receipts, and turns HSA expenses into spreadsheet-ready rows.

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 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.

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.

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 can look weird or morbid from the outside.

AI-generated code gets called slop, but I think that misses the point.

Some colleges are treating AI like contraband.

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.

The new Codex Chrome workflow makes logged-in, UI-heavy websites much easier to extract structured data from.

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 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.

SMS Backup & Restore gives me a real backup, but not a very usable archive.

Anthropic's March 2026 changes now look like a full shift from off-peak incentives to explicit peak-hour cutbacks.

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.

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.

In the late 2000s, a startup called Yoonew turned Super Bowl tickets into a strange quasi-prediction market.

The more I learn, the less random the world looks.

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…

Working with LLMs has made me less likely to equate intelligence with godlike power.

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.

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.

Human intelligence is messy, uneven, and sometimes inseparable from cognitive flaws.

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 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.

My first GitHub Workflow now auto-refreshes my profile README from my blog RSS feeds every six hours.

I added a tiny Bash helper that uses Codex Spark to turn plain-English intent into shell commands.

Codex Spark is fast enough to change what I bother automating.

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.

Treat Codex skills like evolving capabilities, not static tools.

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 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.

Harlite turns HAR browser captures into SQLite so you can analyze web traffic with SQL, optional body storage, deduped blobs, and full text search.

Most people heard a Q&A.

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.

AI labs are running a multi-axis optimization: capital, release timing, research direction, and hardware constraints all move together.

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.

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).

If open source catches up quickly and inference keeps getting cheaper, model IQ stops being the bottleneck.

I built a single Cloudflare Worker that serves a suite of small web tools.

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.

An unscientific, Bill Simmons–style power ranking of the LLMs I actually code with in January 2026.

codex-transcripts converts Codex CLI session JSONL files into clean, mobile-friendly HTML transcripts (with pagination, search, and optional Gist sharing).

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.

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.