Essay · June 2023
My team built a fully-playable D&D-style RPG in a single day, winning the a16z AI Tech Week 2023 Virtual Worlds Hackathon. The approach we used—semantic programming—prefigured the agentic engineering workflows that are now becoming mainstream.
What We Built
Tales of Mythos is a roleplaying game where you roll up a character by selecting Tarot cards that determine your destiny and nemesis, then embark on an open-ended adventure. The story can lead anywhere, yet maintains consistency and a startling ability to recall earlier episodes—thanks to Anthropic's Claude operating with a 100K context window.
We combined Claude with Blockade Labs for dynamic 3D skybox generation (creating scenes wherever the story takes you, without pre-designed environments), Scenario for generative 2D character portraits, and Beamable for the persistent backend infrastructure.
Semantic Programming: Language as Code
Andrej Karpathy predicted a "Software 2.0" model where traditionally-coded software would increasingly be replaced by neural network components. We took this further: we used a language model as the functional equivalent of subroutines you'd normally write code for. The character creation system, combat resolution, narrative progression, and NPC behavior were all managed by structured interactions with the LLM.
The key innovation was what I call NLP-XML subsystems: natural language replaces traditional coding for the logic, while XML provides consistent, parseable data exchange between components. This lets you chain between different AI modalities—text generation, image generation, 3D generation—through a unified architecture where the LLM acts as both the logic layer and the integration layer.
Why This Matters
Semantic programming is excellent for things that need to work with natural language (like storytelling in a game), things you want to rapidly prototype, and systems where the behavior is complex but the specification can be expressed in plain language. The chat interface effectively becomes an IDE—you can isolate functionality, test it conversationally, and iterate at the speed of thought.
Tim Berners-Lee's long-hoped-for "semantic web"—where data and applications interoperate via well-defined schemas—never caught on through traditional approaches. But AI may supply the translation layer that finally makes semantic applications connect. More pragmatically, this translation layer can interconnect software modules within teams, vastly accelerating rapid prototyping.
Looking back from 2026, this hackathon experiment was an early glimpse of agentic engineering. The moment when describing what you want starts to replace specifying how to build it. What was a weekend experiment in 2023 is now a default workflow for building software.
Read the full essay: Semantic Programming and Software 2.0 on Metavert Meditations.
Related: Software's Creator Era · The Direct from Imagination Era · Projects