Newsletter - Public beta, AI playtesting, and why hallucinations are not bugs
Two updates and three new reads.
GameGrammar Is Now in Public Beta
No invitation required. You can sign up, generate your first game design, and start iterating immediately.
New accounts get a 7-day Pro trial with full access to everything, including Nova (the conversational AI design partner) at 20 messages per day.
After the trial, free accounts keep 5 generations per day. Pro unlocks the premium Nova features that make iteration fast: multi-turn design conversations, balance critique with before-and-after comparisons, board topology generation, and AI playtesting with population tournaments.
What Is New Since Last Time
A few capabilities that shipped since the video tutorial newsletter:
- AI Playtesting. Simulate your game against different player archetypes (greedy, exploratory, risk-averse, momentum-seeking) and run every possible matchup. The payoff matrix shows you which strategy dominates and where the counter-dynamics are healthy. Balance issues that would take weeks of game nights to notice surface in two minutes.
- Enhanced Board Topology. For games with spatial elements, Nova generates structured board layouts with nodes, edges, and regions. The topology is not just visual. It feeds into the simulation engine so the AI reasons about movement, adjacency, and area control.
- Card-Aware Simulation. Cards are no longer interchangeable tokens in playtesting. The simulator now distinguishes between card types, executes compiled card effects, and lets MCTS agents learn which cards are worth playing. A "Draw 2 cards" card is genuinely more powerful than a "Gain 1 gold" card in the balance test.
The Research Behind It
Three recent articles if you want to go deeper on the ideas that make GameGrammar work.
Generative Ontology: From Game Knowledge to Game Creation How game knowledge gets structured into a formal ontology, how six AI agents collaborate to turn that structure into playable designs, and what Neural Race (a board game about the race to AGI) taught us about the whole process. Read the article
AI Playtesting: When Your Board Game Tests Itself The technical deep dive on population tournaments, MCTS agents, and why AI players that cannot outperform random play still produce the most useful balance data. Covers the DeepMind-inspired "burnt pancake" discovery and what the payoff matrix actually tells you about your game. Read the article
Hallucinations Are Not Bugs: The Kantian Architecture of AI Consciousness A philosophical exploration of why large language models hallucinate, reframed through Immanuel Kant's 18th-century theory of mind. The argument: hallucination is not a failure of the technology. It is a structural feature of how any system, biological or artificial, constructs a coherent world from incomplete data. Understanding why it happens changes how you build with AI. Read the article
Full Video Tutorial Series
If you missed the last newsletter, the complete 8-video tutorial series is still the fastest way to see what GameGrammar does. From first generation to AI playtesting to Neural Race, the full game showcase.
Try It
GameGrammar is live at gamegrammar.dynamindresearch.com. Free to start. 7-day Pro trial for new accounts. No credit card required.
If you have been thinking about a game idea, this is the fastest path from concept to playable, testable design that I know of.