My Insatiable Appetite for Meaty Projects

I’m on a mission to deeply understand all of AI. Yep, now you know why I tend to be a bit frazzled.

I built a system for knowledge compilation that helps, an implementation of Karpathy’s LLM Knowledge Base pattern that’s become a substantial, constantly growing knowledge base on the AI industry — 1,120 pages as of this morning, compiled and cross-linked on the Mac Mini in my basement from the tweets, articles, podcasts, and meetings I feed it. For me this is essential, and extremely useful.

It’s also not enough. Even with the knowledge refinery up and operating, “reading about” does not equate to “deeply understanding.” I need to actually do real things with AI, to exercise the capabilities I’m reading about.

So I’m constantly on the prowl for meaty projects: work that enables me to exercise the newest developments, from new models to new agent harnesses to new tooling to new knowledge-work capabilities. My appetite for such projects is insatiable, and the moment a candidate occurs to me — bang, I drop it into my Obsidian task list.

The Current Project Mix

Right now the mix focuses on Herdr, Oh My Pi, and a raft of major model releases including Opus 5, Fable 5, three flavors of GPT-5.6, and tons of open-weight models.

First up, Herdr: a model-provider-agnostic orchestration cockpit for AI agents. Not just coding agents, either — lately mine has been herding writing, presentation, and video-creation sessions too. Herdr has been my command center since late June: one window, each project a named workspace, a sidebar that shows at a glance which agents are working, which are done, and which are blocked waiting on me. It also exposes its full control surface as an API, so one agent — one harness-plus-model combination — can orchestrate a whole roster of others (the “Fable to plan, gpt-5.6 to execute” pattern everybody is sketching right now). And with Moshi, an iOS terminal app with native Herdr support and push-when-blocked notifications, I’ve been herding from my phone since the middle of last week.

Alongside Herdr, I’ve added a model-provider-agnostic agent harness: omp, short for Oh My Pi, a batteries-included fork of Mario Zechner’s widely-adopted pi harness. It runs inside Herdr as a first-class citizen and can drive models from 40+ providers, with role-based routing (planning on one model, commit messages on a cheaper one) and mid-session model switching.

Finally, there’s a continuously-refreshing basket of great models, large and small, closed and open, general and special purpose. Prominent from the major labs are Opus 5, out this past week, Fable 5 in Anthropic’s new top tier, and GPT-5.6 in three flavors — Sol the flagship, Terra the balanced middle, Luna the fast cheap one. High-visibility open-weight (and often local-inference-capable) models include GLM-5.2 (MIT-licensed and currently the strongest all-round open model), DeepSeek V4 in Pro and Flash tiers, Moonshot’s Kimi K3 (the K3 weights dropped just yesterday), and Tencent’s Hy3, a 295B mixture-of-experts that until minutes ago was the model running my Hermes Agent environment. With Herdr and omp both highly model-agnostic, I get to drive all of these in real work instead of just reading the model cards.

Why “Model-Provider-Agnostic”?

First: among the major model labs, there are no good guys anymore. Not so long ago I sorted the players into good actors and bad-faith actors, committed to the first group, and avoided the second. The major, massively-funded labs have since convinced me, through their actions over time, that no one belongs in the good-actor group anymore. We can only trust the big players to act in their own interests. A move that looks altruistic? Wait a month, and you’ll learn why it wasn’t. And if there are no good guys, why put yourself at the mercy of an untrustworthy party? With agnostic tooling, no lab can hold my workflow hostage.

Second: as I propose here: if you’ve only experienced one instance of a class, you can’t really know that instance, let alone the class. Until recently the Anthropic stack dominated my experience space — one family of models, one harness, one orchestrator. My current project mix gives me a second instance at every layer: Herdr to compare with Claude Code’s Agent View, omp versus the Claude Code harness, the GPT-5.6 family and a fleet of increasingly viable open-weight models versus Anthropic’s.

Workflow and the Test of Longevity

There’s a great side effect to all this: my project work has compounded into an incredible personal workflow, rigged and automated to the max. The knowledge refinery in the basement, the herding cockpit, the writing pipeline, the analytics on my own hardware — none of these were the point of the projects that produced them, but together they’ve become The Workflow.

Building things that actually get used over time also exposes the quality of the AI’s work. I see people like Ethan Mollick testing new models by having them write games, which I have to assume get minimal gameplay before landing in the trash bin. Build-a-game testing measures some facets of quality — “how creative was the model, how good were its UI-crafting skills, how large a project could it successfully tackle one-shot” — but not quality of technical execution, that is, how many messes are hidden in the code the agent produced. That’s not likely to show up in a few minutes or even hours of playing a game.

Most of my workflow projects, on the other hand, are still in regular use months after their creation, so I get a steady diet of both AI’s successes and its messy failures. Long-term usage exposes work quality in a way that a quick test cannot.