What if you could step inside your dreams and share them with others? In this fascinating conversation, we sit down with *Shahbuland Matiana* (Co-founder & Head of Research) and *Andrew Lapp* (Member of Technical Staff) from Overworld Labs to explore their groundbreaking new technology: *Waypoint 1* — an open-source world simulation model that runs on consumer hardware.

Unlike Google's Genie, which requires massive cloud infrastructure, Waypoint 1 is designed to run on your gaming PC. We're talking 3070s, 4090s, even Apple Silicon. This 2 billion parameter model generates interactive worlds at *60 frames per second* — and they're releasing the weights for free.

*The Vision: Dreams You Can Record*

Shahbuland shares a vivid lucid dream that shaped his entire research direction — a house floating in space, a circling dragon, a katana parry that cracked the floorboards beneath his feet. "This is the kind of thing where dreams can give you these really amazing fully immersive experiences, but there's no way to record them. There's no way to share them." That's what Overworld is trying to change.

*How It Actually Works*

The technical architecture is genuinely novel — a hybrid between a causal language model and an image diffusion model. Instead of predicting the next token like ChatGPT, it *denoises the next 256 tokens* representing each frame. Every sixteenth of a second, the model generates a new frame conditioned on all previous frames, your text prompt, and your controller inputs in real-time.

*Why Privacy Matters Here*

These simulations are extensions of our minds. When we imagine future scenarios, run mental simulations, or explore creative spaces — that's deeply private. The team agrees that running locally gives users ownership over their experiences in a way that cloud streaming never could.

*We're Still Early*

Perhaps most refreshing is their honesty about where the technology stands. Unlike LLMs, which have crossed from research into engineering, world models are still in active research territory. "Every other week, it feels like someone comes out with a paper that finds a way to make it 100 times faster."

This conversation covers the brain as a simulator, chaos theory in diffusion models, why fewer sampling steps reduce diversity, and the future of interactive entertainment. Whether you're a researcher, a game developer, or just someone curious about where AI is heading — this one's worth your time.

---
TIMESTAMPS:
00:00:00 Introduction & Overworld Demo
00:02:49 Core Technical Capabilities
00:05:38 Image Prompting & Experience Sharing Vision
00:08:59 Lucid Dreaming Vision & Shared Experience
00:11:34 Open Source Origins & Platform Philosophy
00:14:30 Technical Architecture Deep Dive
00:17:56 Optimization, Distillation & Future Outlook

---
REFERENCES:
Company:
[00:00:00] Overworld Labs
https://over.world/
[00:12:01] Stability AI
https://stability.ai/
AI Model:
[00:00:18] Google Genie
https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/
[00:00:45] Waypoint 1
https://huggingface.co/Overworld
[00:11:51] Sora (OpenAI)
https://openai.com/sora
[00:17:15] HunyuanVideo (Tencent)
https://github.com/Tencent/HunyuanVideo
AI Tool:
[00:06:13] Claude Code (Anthropic)
https://www.anthropic.com/claude-code
Paper/Technique:
[00:18:04] Rectified Flow Models
https://arxiv.org/abs/2209.03003
[00:24:00] DMD (Distribution Matching Distillation)
https://arxiv.org/abs/2311.18828
Code Repository:
[00:26:25] Overworld GitHub
https://github.com/Overworldai
Community:
[00:26:30] Overworld Discord
https://discord.gg/overworld

---
LINKS:
PDF Transcript: https://app.rescript.info/api/sessions/9987e73cc06390de/pdf/download
RESCRIPT:
https://app.rescript.info/public/share/mDC42R65XfkXOsRkbqN7qaBjHwUvIPTMUrkoBAeN6g4

Machine Learning Street Talk

Overworld Labs introduces Waypoint One, a 2-billion parameter real-time world model that runs on consumer GPUs (RTX 3090/4090/5090) at 60 fps. Unlike cloud-based solutions like Google's Genie, this open-source model enables interactive text-to-experience generation locally, allowing users to explore AI-generated worlds with real-time control inputs. The system uses a hybrid architecture combining autoregressive transformers with diffusion models, compressing 360p video into 32×32 latent representations and generating frames sequentially using 4-step rectified flow. Technical innovations include avoiding temporal compression VAEs for low-latency input handling, distillation techniques that maintain quality while reducing inference steps, and compute-bound optimization strategies. The team prioritizes privacy, local execution, and community experimentation over cloud streaming, with plans to extend context windows, add image conditioning, and explore mixture-of-experts architectures.

"We Made a Dream Machine That Runs on Your Gaming PC"