AGI Capability Map¶
This map does not treat AGI as a single score. It decomposes the problem into capabilities that can be observed, reproduced, and challenged.
flowchart LR
A[Represent\ntext · image · video · state] --> B[Predict\ntoken · latent · trajectory · action]
B --> C[Get feedback\nlabels · preferences · verifiers · environment]
C --> D[Close the loop\ntasks · data · policies · systems]
D --> A
B --> E[Agent / VLA]
E --> C Four connecting questions¶
Place every new model under the same four questions:
- Representation: what does it compress into a computable internal state?
- Prediction: does it predict the next symbol, a full distribution, or a future trajectory?
- Feedback: how does it know that it is wrong?
- Closed loop: can feedback produce better tasks, data, or policies for the next round?
This vocabulary makes seemingly different work comparable: LLMs predict in token space, diffusion models denoise in latent space, VLAs predict in state–action space, and self-improving systems search over experiments and update rules.
Reading path¶
- To understand the core mechanism: start with Transformer and language modeling.
- To recover historical causes: read the Technology Timeline.
- To assess current progress: open the Frontier Radar.
- To test ideas yourself: open Experiments.