AGI Frontier / Research Lab Manual
From token prediction to systems that improve themselves.
A public, reproducible, continuously updated tutorial series tracing the shared path from LLMs and multimodal models to generative models, world models, VLAs, and self-improving agents.
A map that keeps growing¶
01 · REPRESENT Represent the world How text, images, video, states, and actions enter a model. 02 · PREDICT Predict the future The shared structure behind token, latent, trajectory, and action prediction. 03 · VERIFY Get feedback From labels and preferences to verifiers, rewards, and environment feedback. 04 · EVOLVE Keep improving Automatic improvement of tasks, data, policies, code, and architectures.
Current state¶
How to use this site¶
If you are new here, start with the tutorials. Use the Technology Path when you want the historical order. The series does not try to cover every paper. Each page selects a mechanism, explains it, records where it fails, and proposes the next test. A page can ship before its experiment exists.
Current judgments live in the Frontier Radar. Experiments are optional. GitHub Markdown is the source of truth; revisions and corrections remain public.
Updates: Start with GitHub Watch. RSS/Atom and email subscriptions will come after the publishing cadence is stable, so the learning loop does not become an operations project.
One-line promise¶
Do not maintain a list of what I have read. Maintain how my AGI judgments changed, and how to test the next one.