Debates about AGI can drift far from reality. Here's a grounded snapshot of typical strengths and weaknesses of current frontier systems.
Strengths
- Fluent writing, translation and summarisation.
- Programming across many languages, including substantial multi-step tasks.
- Strong performance on many exam-style questions in science, law and medicine.
- Understanding images, charts and documents.
- Using tools and completing multi-step tasks as agents.
Weaknesses
- Confident errors and invented facts.
- Inconsistent reliability: success on hard problems alongside failures on simple ones.
- Limited ability to learn continuously from experience after training.
- Difficulty with very long, open-ended tasks without supervision.
- Weaker physical-world understanding and robotics.
Uneven Profiles
Capabilities are "jagged": superhuman in some areas, surprisingly weak in others. This makes simple comparisons with human intelligence misleading.
Changing Quickly
These lists shift with each model generation. Evaluate systems on your own tasks rather than relying on general claims.