Historical overview
The field was named at the 1956 Dartmouth workshop. It progressed through symbolic systems (1960s–80s), expert systems (1980s), statistical machine learning (1990s–2010s), deep learning (since 2012), and the transformer era beginning with the 2017 'Attention Is All You Need' paper.
Scientific basis
Modern AI is dominated by deep neural networks trained on large datasets via gradient descent. Transformers use self-attention to model long-range dependencies, enabling large language models with hundreds of billions of parameters and frontier multimodal systems.
Strengths
- Massive parallel processing of structured and unstructured data
- Perfect recall of training distribution and instant duplication
- Superhuman performance on narrow benchmarks like protein folding and Go
- 24/7 availability at marginal per-query cost
Limitations
- Brittle out-of-distribution generalisation
- Hallucinations and confidently wrong outputs
- Energy and data demands that scale faster than capabilities
- No grounded embodied experience or intrinsic goals
Relationship to other intelligence systems
Human Intelligence
Models human language and reasoning patterns from text.
AGI
Considered the substrate from which AGI is expected to emerge.
Machine Intelligence
AI is the largest commercially deployed form of machine intelligence.
Future implications
Convergence toward agentic systems, world models, and reasoning models that combine search with learned heuristics is the dominant research trajectory through the late 2020s.