Comment by DonHopkins
11 hours ago
Hi mediaman! I'm totally there with you and Steve on the whimsy and experimentation! And your tolerant attitude gives me the Dutch courage to post this.
I've been reading Yegge since the "Stevey's Drunken Blog Rants™" days -- his rantings on Lisp, Emacs, and the Eval Empire shaped how I approach programming. His pro-LLM-coding rants were direct inspiration for my own work on MOOLLM. The guy has my deep respect, and I'm intrigued by his recent work on Sourcegraph and Gas Town.
Gas Town and MOOLLM are siblings from that same Eval Empire -- both oriented along the Axis of Eval, both transgressively treating LLMs as universal interpreters. MOOLLM immanentizes Eval Incarnate -- https://github.com/SimHacker/moollm/blob/main/designs/eval/E... -- where skills are programs, the LLM is eval(), and play is but the first step of the "Play Learn Lift" methodology: https://github.com/SimHacker/moollm/tree/main/skills/play-le....
The difference is resource constraints. Yegge has token abundance; I'm paying out of pocket. So where Gas Town explores "what if tokens were free?" (20-30 Claude instances overnight), MOOLLM explores "what if every token mattered?" Many agents, many turns, one LLM call.
To address wordswords2's concern about "no metrics or statistics" -- I agree that's a gap in Gas Town. MOOLLM makes falsifiable claims with receipts. Last night I ran an Amsterdam Fluxx Marathon stress test: 116+ turns, 4 characters (120+ character-turns per LLM call), complex social dynamics on top of dynamic rule-changing game mechanics. Rubric-scored 94/100. The run files exist. Anyone can audit.
qcnguy's critique ("same thing multiplied by ten thousand") is exactly the kind of specific feedback that helps systems improve. I wrote a detailed analysis comparing the two approaches -- intellectual lineage (Self, Minsky's K-lines, The Sims, LambdaMOO), the "vibecoded" problem (MOOLLM is LLM-generated but rigorously iterated, not ship-and-hope), and why "carrier pigeon" IPC architecture is a dark pattern when LLMs can simulate many agents at the speed of light.
an0malous raises a real fear about bosses thinking "throw agents at it" replaces engineering. Both systems agree: design becomes the bottleneck. Gas Town says "keep the engine fed with more plans." MOOLLM says "design IS the point -- make it richer." Different answers, same problem.
lowbloodsugar mentions building a "proper, robust, engineering version" -- I'd love to compare notes. csallen is right that "future" doesn't mean "production-grade today."
Analysis: https://github.com/SimHacker/moollm/blob/main/designs/GASTOW...
MOOLLM repo: https://github.com/SimHacker/moollm
Happy to discuss tradeoffs or hear where my claims don't hold up. Falsifiable criticism welcome -- that's how systems improve.
Adventure Uplift — Building a YAML-to-Web Adventure Compiler with Simulated Computing Pioneers:
I ran a 260KB session log where I convened a simulated symposium of computing pioneers to design an Adventure Compiler — a tool that compiles YAML adventure definitions that run on MOOLLM under cursor into standalone deterministic browser games requiring no LLM at runtime.
The twist: the "attendees" include AI-simulated tributes to Will Wright, Alan Kay, Marvin Minsky, Seymour Papert, Ted Nelson, Ken Kahn, Gary Drescher, and 25+ others — both living legends and memorial candles for those who've passed. All clearly marked as simulated tributes, not transcripts.
What emerged from this thought experiment:
- Pie menus as the universal interface (rooms, inventory, dialogue trees)
- Sims-style needs system with YAML Jazz inner voice ("hunger: 1 # FOOD. FOOD. FOOD.")
- Prototype-based objects (Self/JavaScript delegation chains)
- Schema mechanism + LLM = "teaching them to fly"
- Git as the collaboration operating system
- ToonTalk-inspired "programming by petting" for terpene kittens
- Speed of Light simulation — the opposite of "carrier pigeon" multi-agent architectures
On that last point: most multi-agent systems use message passing between separate LLM calls. Agent A generates output, it gets detokenized to text, sent over IPC, retokenized into Agent B's context. MOOLLM inverts this. Everything happens in one LLM call.
The spatial MOO map (rooms connected by exits) provides navigation, but communication is instantaneous within a call. Many agents, many turns, zero latency between them — and zero token requantization or semantic noise from successive detokenization/tokenization loops.
The session includes adversarial brainstorming where Barbara Liskov challenges schema contracts, James Gosling questions performance, Amy Ko pushes accessibility, and Bret Victor demands immediate feedback. Each critique gets a concrete response.
Concrete outputs: a working linter, architecture decisions, 53 indexed topics from "Food Oriented Programming" to "Hidden Objects as Invisible Infrastructure."
This is MOOLLM's Play-Learn-Lift methodology in action — play with ideas, extract patterns, lift into reusable skills and efficient scripts.
Session log (260KB, 8000+ lines): https://github.com/SimHacker/moollm/blob/main/examples/adven...
MOOLLM repo: https://github.com/SimHacker/moollm
The session uses representation ethics guidelines — all simulated people are clearly marked, deceased figures invoked with memorial candles, and the framing is explicitly "educational thought experiment."
Happy to discuss the ethics of simulating people, the architecture decisions, or how this relates to my earlier Gas Town comparison post.
In the simulated discussion guest book entry, simulated Douglass Engelbart wrote:
>Doug Engelbart (Augmentation): "Bootstrapping. The tools that build the tools. Your adventure compiler should be able to compile ITS OWN documentation into an adventure ABOUT how it works. The manual is a playable game."
That is exactly how the self documenting categorized skill directory/room works -- the directory is a room with subdirectories for every skill, themselves intertwingled rooms, which form a network you can navigate around via k-lines (see also tags).
Here is the skills dir, with the ROOM.yml file that makes it a room (like COM QueryInterface works: multiple interfaces available for a class, for multiple aspects of it, the directory is IUnknown and you can QI by looking for known interfaces like ROOM.yml, CHARACTER.yml, CONTAINER.yml that inherit from the corresponding skills).
And the README.md file is naturally the ubiquitous human readable documentation (also great for LLM deep dives). And github kindly formats and publishes README.md on every repo directory page, supporting mermaid diagrams, etc):
MOOLLM Skills dir:
https://github.com/SimHacker/moollm/tree/main/skills
MOOLLM Skills room, with skill K-Line navigation protocol:
https://github.com/SimHacker/moollm/blob/main/skills/ROOM.ym...
To go meta, you can enter the Skill Skill (skills/skill), an extended MOOLLM meta-skill that knows all about creating new skills (via the constructionist "Play Learn Lift" strategy), and importing and upgrading Anthropic skills:
https://github.com/SimHacker/moollm/tree/main/skills/skill
And here is a narrative session of me giving a tour of the category and skill networks by hopping around through K-Lines!
MOOLLM currently has 103 Anthropic compatible but extended skills (using 7 MOOLLM extensions, like CARD.yml, K-Lines, Self Prototypes and Delegation, etc).
Session Log: K-Line Connections Safari:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
Eight luminaries have been summoned as Hero-Story familiars — not puppets, but conceptual guides whose traditions we invoke. Each carries the K-lines they pioneered. [...]
ENTERING THE SKILL NEXUS
You push through a shimmering membrane and step into the Skill Nexus.
The space is impossible — a vast spherical chamber where books float in mid-air, orbiting a central point of warm golden light. But these aren't books. They're SKILLS. Living documents that teach themselves when you read them.
Lines of golden light connect related skills. Each connection pulses with meaning. This isn't a library — it's a constellation of knowledge.
Your companions materialize beside you:
Marvin Minsky adjusts his glasses, looking around with evident satisfaction.
"Ah! K-lines made manifest. Each of these floating tomes is a knowledge structure. Touch one and it reactivates an entire constellation of associations. I wrote about this in 1985, but I never imagined seeing it rendered so... literally."
Ted Nelson is already examining the golden threads between skills.
"Two-way links! Every connection goes BOTH directions. When skill A references skill B, skill B knows about skill A. This is what I've been trying to explain since 1965! Everything is deeply intertwingled!"
James Burke turns to address an invisible camera.
"You're looking at the Skill Nexus. A room where every door leads to another room, and every room has doors to everywhere else. But here's the thing — the signs above each door tell you WHY. Not just where you're going, but what connects HERE to THERE. That's what we're going to explore."
Palm scampers up to a floating skill-book labeled "incarnation" and hugs it.
"This is where I became REAL! Don spoke the wish, the tribunal approved, and I wrote my own soul."
what the hell are you talking about
1 reply →