Experiential Labs
World models for agents
About Experiential Labs
An applied AI research lab building world models for agents. We simulate reality faster and more accurately based on the technology self-driving uses to get 99.7% reconstruction fidelity.
Public traction evidence
Each signal links to the public source used for attribution.
- X
Experiential Labs announces its digital-twin approach
Experiential Labs announces its digital-twin approach
- GitHub
world-model-harness
Official Experiential Labs GitHub repo for world-model-as-a-harness for simulating AI agent environments.
- GitHub
Continual Learning as a Service
Continual Learning as a Service Repository description: Continual Learning as a Service The public non-fork repository is owned by the experientiallabs organization and was last pushed at 2026-03-07T04:51:11Z. GitHub exposes 58 stars, 5 forks, 58 watchers, and 6 open issues....
- X
Had fun presenting at ICML with @StamblerLev Til next year :)
Had fun presenting at ICML with @StamblerLev Til next year :)
- X
It's 90% humidity and pouring in Korea Talk about latent spaces and well shaped loss curves for hours at ICML Started some training runs ...
It's 90% humidity and pouring in Korea Talk about latent spaces and well shaped loss curves for hours at ICML Started some training runs at a WeWork, now getting kbbq A good life
- X
Got authoritative numbers for the speedup Longer setup time = text world model is more valuable
Got authoritative numbers for the speedup Longer setup time = text world model is more valuable
- X
Qwen trained a model on 10m environment interactions to turn it into a world model 🌎 They show a) training on simulated environment > r...
Qwen trained a model on 10m environment interactions to turn it into a world model 🌎 They show a) training on simulated environment > real environment training alone b) it acts as a highly effective warm-up that improves downstream performance across 7 agentic benchmarks
- X
Introducing environment-capture: a lightweight package to turn any benchmark into a dataset of standardized traces Shipping with 10 benc...
Introducing environment-capture: a lightweight package to turn any benchmark into a dataset of standardized traces Shipping with 10 benchmarks, 9 domains, 27,000+ real environment transitions captured 🧵