AI today
Where AI stands now: what it does well, its limits and open questions.
What AI does well
- Recognising patterns in images, speech and text.
- Drafting, summarising and translating language.
- Writing and explaining code.
- Helping scientists, for example by predicting protein structures.
Where it falls short
- Language models can state false things confidently, often called hallucination.
- Systems can reflect biases present in their training data.
- Training large models takes a lot of data, computing power and energy.
- Performance can be uneven: a system can excel at one task and fail at a similar one.
Open questions
- How can we check that AI systems are reliable and safe?
- How should AI be governed, and who is responsible when it errs?
- How will it change work and education?
- How far can current methods go, and what else is needed?
For how we got here, see the timeline and the breakthroughs.