Thoughts

Making a snazzy video with Opus 5.5. Wow, if everyone can make things like this, the value of everything other than production quality goes up.

I’m winding down a very broad and deep tour of building lots of different things with AI, for fun.

It involved everything from creating a distributed AI job/harness scheduling system to execute repeatable job chains with unique strategies in vendor harnesses, to an in-loop AI plan editor, to a 3D web card game, to a personal blog/site revamp, to a portfolio rebalancing tool, to migrating a complex 15-year-old iOS app with numerous no-longer-supported dependencies to current SDKs and debugging performance issues with AI, to building out some websites on domain names I had sitting around accumulating dust, to building AI skills and workflows, generating music and 3D models with AI plus an intermediate programmatic language, using paid and free models and running all manner of experiments to assess the edges of those. It has been a blast!

Probably both the most visible and the thing I have heard the most positive feedback on is my personal site revamp, so I wrote up (with some help, of course!) a recap on all the little customizations and niceties I pulled together there

Cheers!

Can Stanley cup personalization be taken too far?

Doing a little experiment with having a synth/cylon/droid/bot research agent that checks LLM generated prose phrasing of claims against sources.

Got a 3D printer recently (P2S)… Learned a bare minimum amount of CAD and then discovered gpt-5.6 can do quite a bit without it!

and…revise…

TIL: ASD-STE100 ftw!

Also in 2026: new tools enable new possibilities.

Prototype of adding HTML plan file support via in-loop browser launch of rep tool. It won’t work cleanly over ssh/tmux/etc., but great for local HTML plan revisions. Uses same keyboard shortcuts as the TUI, for quick plan navigation too.

2026 collateral damage: I dropped fish as my preferred shell this year because generating zsh scripts with LLMs is just so good that for me, the benefits of using fish no longer justify the weight of the additional dependency in my default setup.

related: also attempting to not use Homebrew at all on personal Mac, in favor of pinned project-level dependencies, and mise installing binaries from GitHub releases via their provider.

I just kicked off Fable doing a bakeoff, training LoRA weights with a few image diffusion (DiT) models.

This is an area I know very little about.

I fed it images, briefs, and judging criteria for guidance, and now it is renting GPUs on @runpod, downloading models, and configuring+executing training runs, reviewing results, and adjusting.

It’s so crazy it can even attempt this … it’s a little nerve-racking waiting to see how it turns out. 🙃

result:

I am dazzled… Trained up LoRA weights with open source model, remotely, renting Nvidia chips, and generated an automated re-ranker based on human reviews, and can now generate on-style tiny icons both on remote CUDA, and local Mac w/MLX.

Thousands of them🤯

TIL: If you can keep a lot of the experiment organized in context and recorded thoughtfully in logs on disk, a good [llm] model can do an amazing job of summarizing it and explaining it from multiple perspectives. One area I discovered was weak, though, was follow-on and productionizing for repeatability. In service of meeting a /goal, model tended to ‘work around’ things and sometimes not clean up well.

Fable and gpt-5.6-sol are like rocket fuel for prototyping fun things to put on old domain names I have sitting around otherwise collecting dust. Part 2:

Here’s a tier ranking web app for sharing ranked preferences.

It uses Cloudflare (Pages, D1, R2, Turnstile, worker analytics engine, worker cron), Auth.js w/Google OAuth and Resend integration for passwordless sign in, an open source image diffusion model with LoRA weights for icon generation, and Playwright for E2E browser tests.

Feel free to poke at it if interested. Try it here

LOL. You keep using that word. I do not think it means what you think it means.

Fable and gpt-5.6-sol are like rocket fuel for prototyping fun things to put on old domain names I have sitting around otherwise collecting dust. Part 1:

Here’s Klondike Solitaire implemented in Babylon.js with custom shaders, 9 shuffle variations, 3D camera pans, responsive layout for desktop/tablet/mobile, solvability vetting and some fun easter eggs like being able to toss the cards in the fire when you are stuck lol.

There’s still no substitute for human taste judging how fun something is and fine-tuning.

Thus, a pre-LLM proven game concept is a huge leg up, not only because the LLM has seen similar code before to smooth implementing/iterating, but because the design of the game and its ruleset is proven to at least be interesting to people.

Try it here

Added dynamic music which moves with game progress using https://strudel.cc, because, why not?

New unlock found while stacking up AI capabilities:

Have the LLM generate a custom rich (HTML w/screenshots, videos, keyboard shortcuts, etc) comparison/review UI for providing feedback and decisions, w/embedded code and UI that can submit follow-up jobs back to orchestrator. 🤯

Finally got all the knobs turned right to record a demo of how this experimental tool I made for revising Markdown plans works in the loop of a Claude Code session.

“One of the challenges we will face in the next 2 years is the economics of tokens. The marginal cost of productivity improvement has to match the marginal cost of the token. That’s a management discipline. […] The business has to benefit from it.” — SN

LLMs enable a new brainstorm, review, revise, and select mechanism for prototyping and iterating toward a goal. It requires tracking intention, rules vs. guidance, optionality in forward paths, and optimizing review UIs within a directed graph.

Whoever can make that process durable, distributed, multiplayer, and multi-model is going to accelerate beyond all reason.

A lot of the current established industry players are trying to build ‘software factories’ on top of existing primitives.

IMHO, if there is one thing I have learned from prototyping my own local distributed job runner for AI to implement various strategies of how to form the work, it is that we could really use some new primitives, like yesterday.

Ohhhhh…

For fun, I’ve been building a little distributed agent scheduling system to explore different AI-related dev practices and workflows … with an eye towards stacking up primitives and finding compounding effects …

I just added Option ‘A’ here, which works sooooo much better than anticipated, I am dazed. There’s no substitute for a pro [human] designer, but with some taste (reject most ideas and adopt/try a select few), this raises the floor so much it feels crazy.

Here’s a little peek inside the engine for that first run and the models used, that I am still digesting…

Would have cost me ~$12 in tokens (if not for subscriptions).

It feels like something akin to the current momentum around people vibe-coding prototypes that they then try to hand to developers and ask them to support, enhance, and maintain in production forever.

I made a fun little experimental TUI (Rust and Ratatui) to quickly revise generated Markdown plan files. If you wrap an agent in tmux, a skill can present this as a modal (from Claude Code or Codex) and feed the results (strict annotations and loose intent feedback) directly to an agent loop.

Yikes. I need to rethink this:

The generated chain has 61 model workers plus the collector, still under the 64 dynamic jobs allowed from this source job. I’m submitting it now and will poll until the collector completes.

OK, it’s Friday. Musssssst… put… thisssss… down… Ahhh. Happy weekend!

Tool tip: I am very satisfied after switching to keep.md to have email newsletters, GitHub stars, Twitter bookmarks, RSS feeds, and read-later bookmarks from mobile and browser all in one inbox for follow-up or wiring into automation (MCP, API, CLI).

Using AI models to surface potential improvements for your particular code and product along the dimensions that are important to you is a fun way to find things you might otherwise miss.

Below, a rough visual ranking from an experiment I ran, asking six different models to look through a codebase and brainstorm along dimensions I wanted to focus on.

Aside: even Opus 4.8 agreed with me that GPT-5.5 did the best job. LOL. 🤣

What do you see when you look at these latest model rankings from Artificial Analysis?

WWDC 2026 keynote: Hey developers, we’re not going to say [wink] why we’re so focused on your mobile apps being resizable [wink], but check out this origami (literally folding paper) example app [wink]. A folding phone is coming soon? Waaaaaaaaat? Where would you get that idea?

Playing around with small LLMs that run entirely locally in a user’s web browser, seeded with context from my site.

Come on… please!

A fun thing you can leverage AI for is learning about and tuning your own harness and habits with long-running experiments using data, or derivable data, that only you have.

Your results can vary widely from others based on your codebase, your models, and your context/tools tuning.

For example, I wanted to get a sense of whether I was defaulting to the right reasoning level when using GPT-5.5 on a particular codebase.

It’s easy to default to GPT-5.5 high/xhigh, but higher spending, wall-clock time, and model overthinking are often creeping results that compound over time.

I had Claude Code with Opus 4.8 orchestrate an experiment analyzing 100 past good commits (no bugs/follow-ups) to assess what reasoning level Codex with GPT-5.5 needed to reproduce those changes effectively.

Each experiment run was isolated, with no knowledge of how the changes were actually implemented in the real codebase, a prompt to express intent, and a starting point (the previous commit).

Across 400 runs (4 reasoning levels x 100 commits), a few patterns emerged: high/xhigh was rarely needed to solve correctly, and the highest thinking level often took ~2.5x the wall-clock time and 10x the tokens/cost.

I’ll definitely be adjusting my default when working on this particular codebase.

Of course, what works for you will depend on your codebase, your models, and your context/tools tuning, so YOU SHOULD RUN YOUR OWN experiments like this!

Playing around with some ‘personal software’… a Rust-based TUI built on top of the keep.md API for reading and sorting through everything I have it collecting …

Thanks for providing the API, @iannuttall!

Inside cmux, a little experiment to have this control a browser in a split is working well:

Open-sourcing this under the MIT license here: github.com/mattorb/sift

Good judgment and diligence have always been important to building scalable systems, and that has not changed. It’s easy to lose sight of that when there is a daily onslaught of new AI skills, models, and techniques competing for your attention.

Judging what to carry in your AI toolkit

Succinctly and accurately describing work to be done is a superpower with this new crop of AI tools, so the cruelest irony is that using them carelessly can erode those very skills over time.

If you know, you know.

Doing some fun Friday experiments: