Skip to content

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:

  1. Representation: what does it compress into a computable internal state?
  2. Prediction: does it predict the next symbol, a full distribution, or a future trajectory?
  3. Feedback: how does it know that it is wrong?
  4. 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