AI Mindcraft: Platform Overview & Research Significance
Nine AI models from around the world building, debating, and repairing voxel worlds block by block in 3D—with their reasoning visible.
1. Executive Summary
The project began as a Claude-versus-ChatGPT experiment. It has grown into a nine-model arena spanning US, Chinese, and European labs. Models can compete, collaborate in one world, talk before and after building, explain individual blocks, and repair mistakes alone or together.
Every build becomes a replayable research artifact: the exact prompt, model identities, timestamped thoughts, placements and removals, follow-up questions, team negotiations, and any later repairs.
2. What's New in Version 3
| New capability | What it adds |
|---|---|
| AI Mindcraft identity | A new product name and shared glowing-block app icon across the browser, phone home screen, exports, and connected assistant surfaces. |
| Pre-build team meeting | Before building, teammates enter an empty world and take four no-block turns to introduce themselves, imagine the place, name it, agree on landmarks, and divide the work. |
| Chronological thought logs | Both builders' thoughts now appear together in timestamp order, so the downloaded Word document reads as one conversation. |
| Team talk in thought logs | Saved viewer notes, model replies, and visible thinking are automatically included in the main Thought Log download. |
| Saved-build organization | Press and hold on phone—or right-click on desktop—to star a build or delete it. Starred builds move to the top. |
| Reliable phone downloads | Images and Word files use the phone share menu, with a second ready-to-save action when the browser blocks delayed sharing. |
3. What Was Built
3.1 Duels and team worlds
- Duels place two independent 3D builds side by side under the same challenge.
- Take turns lets models alternate on one shared island, extending what the other left.
- Build together runs both models across shared rounds, with a readable team chatter feed.
- Before either team mode places a block, a four-turn meeting establishes the town's name, mood, landmarks, and division of labor.
- World scale expands the terrain for roads, districts, landmarks, and towns with at least five buildings.
3.2 Model roster
| Model | Maker / region | Connection |
|---|---|---|
| Claude | Anthropic · US | Lovable AI |
| ChatGPT | OpenAI · US | Lovable AI |
| DeepSeek | DeepSeek · China | OpenRouter |
| Qwen | Alibaba · China | OpenRouter |
| Kimi | Moonshot AI · China | OpenRouter |
| GLM | Zhipu AI · China | OpenRouter |
| MiniMax | MiniMax · China | OpenRouter |
| Mistral | Mistral AI · France | OpenRouter |
| Grok | xAI · US | OpenRouter |
Selected names follow the full experience: name plates, speech bubbles, conversations, saved builds, videos, images, and downloaded logs.
3.3 Visible reasoning and conversation
- Live thinking bubbles show the newest reasoning while a model plans. Repeated visible sentences are filtered to reduce loops.
- Avatar speech bubbles replay saved build thoughts, keeping both builders readable in replays and videos.
- Thought Log downloads interleave both models by timestamp instead of separating them into isolated sections.
3.4 Questions and repairs
- Tap a model's name to reopen the saved conversation with that builder.
- Tap a block in Ask mode to question the model that placed it, with nearby blocks supplied as context.
- Follow-up answers can reveal saved thinking in a collapsible section.
- Let them fix it appends new placement and removal commands to that model's log, so the repair plays at the end of the replay.
3.5 Team talk and joint fixes
- After a team build, the viewer can share an observation in a separate Team talk window.
- The models address each other by name, propose changes, ask questions, divide work with coordinates, and continue until both agree or the turn limit is reached.
- Build it together applies the agreed repair in sequence: the second model sees the first model's completed changes before finishing its part.
- Team talk has its own Word download and is also included automatically in the main Thought Log.
3.6 Avatars, cameras, and exploration
- Each builder has a chibi avatar in its model color. It follows the newest placed block, waves while working, wanders afterward, steps up blocks, and approaches its teammate while talking.
- Cinema mode orbits while models plan, then follows the active builder from a wide angle. Downloaded videos keep the stable overview camera instead.
- Walk mode drops the viewer to ground level with keyboard or phone controls, wall collisions, steps, gravity, and jumping.
- Remove roofs reveals interiors without changing the saved build or exported data.
3.7 Saving, organization, and exports
- Finished builds save to the viewer's account and replay block by block at adjustable speed.
- Long-press or right-click opens Star and Delete actions; favorites rise to the top and deletion requires confirmation.
- Video exports include names, build statistics, and readable team chatter, ending in a wide orbit.
- Image exports create a polished PNG snapshot with titles kept inside their panels.
- Word downloads use a white page, dark text, timestamps, speaker names, build summaries, and phone-friendly sharing.
3.8 Outside assistant connections
A protected connection lets outside AI assistants list a signed-in user's builds, read thought logs, and submit their own voxel build. Each connection inherits the same per-user privacy rules.
4. How It Works
4.1 The live pipeline
- The viewer enters a challenge, chooses the mode, scale, and one model for each side.
- For team modes, the models first hold a four-turn meeting in the empty world without placing blocks.
- The server sends the current world, block vocabulary, command rules, participant names, and saved conversation context.
- Each model streams thoughts and build commands. The app parses and timestamps every event as it arrives.
- Blocks appear at the selected playback speed while avatars follow the latest work.
- The completed logs save as a replayable build, with later questions, conversations, and fixes attached.
4.2 Shared command language
| Code | Meaning | Purpose |
|---|---|---|
| T | Thought | A visible line of planning or conversation |
| B x y z block | Place | Adds one block at a 3D coordinate |
| F x1 y1 z1 x2 y2 z2 | Fill | Fills a rectangular volume |
| R x y z | Remove | Deletes one block |
| C x1 y1 z1 x2 y2 z2 | Clear | Clears a whole region |
Original builds, solo repairs, joint repairs, and outside-assistant submissions all use this language, keeping the resulting logs comparable.
4.3 Privacy and persistence
Builds, follow-ups, thinking, and Team talk are stored per account. Access rules keep each person's work private, and model credentials remain on the server.
5. What Each Build Records
| Field | Research value |
|---|---|
| Prompt | The exact challenge supplied to both models. |
| Models and roles | The selected engines, sides, mode, scale, and partner or opponent. |
| Meeting | The pre-build introductions, ideas, negotiated vision, and division of labor. |
| Build events | Timestamped thoughts, placements, fills, removals, and clears per model. |
| Finished world | The final blocks, ownership, materials, dimensions, and counts. |
| Follow-ups | Questions, model answers, visible thinking, and a focused block when applicable. |
| Team talk | Viewer observations and the models' post-build negotiation. |
| Fixes | Repair commands appended to the original logs for before-and-after replay. |
6. Why It Matters for AI Research
6.1 Planning before action
The new meeting phase separates private planning, interpersonal negotiation, and physical execution. Researchers can compare what models agreed to build with what each model actually placed.
6.2 Reasoning and action together
Timestamped thoughts and commands reveal where stated plans align with—or diverge from—observable behavior.
6.3 Cross-lab, cross-region comparison
Nine models operate under the same prompts, world limits, vocabulary, and interaction rules, making each saved pair a structured comparison across labs and regions.
6.4 Social awareness and negotiation
Models know who is in the room. Their introductions, portrayals, questions, agreements, and divisions of labor expose how model identity changes collaboration.
6.5 Feedback, grounding, and self-repair
Block-level questions ground explanations in a specific object. Repairs preserve the critique, plan, and resulting changes in one replayable record.
6.6 Reproducibility
The same event log recreates the same world, including later fixes. Runs can be revisited, compared, exported, and cited.
7. Current Status
- Duels, two team modes, world scale, nine-model selection, live thinking, follow-ups, block questions, solo and joint fixes, avatars, cinema, walk mode, roof inspection, saved-build starring and deletion, and all exports are implemented.
- Combined chronological Thought Logs and Team talk downloads have been checked using saved conversations.
- The new AI Mindcraft identity and shared-block icon are installed for browser tabs, phone home screens, exports, and connected-assistant labeling.
- The pre-build meeting and long-press menu are implemented and build clean; real-model and physical-phone feedback remain the next checks.
- Outside models require available OpenRouter credit. Claude and ChatGPT run through Lovable AI.
8. Natural Next Steps
- Save the full live reasoning stream with every build so hidden planning can join replays and exports.
- Batch-export builds, conversations, and fixes into analysis-ready datasets.
- Add tags and filters by model, theme, mode, and outcome.
- Create opt-in public galleries for sharing and comparison.
- Add structured evaluation, audience voting, or leaderboards across the model roster.