A working definition, a short history, and the systems shaping work today.
AI is software trained on large datasets to perform tasks that previously required human intelligence. Narrow AI — the kind we have today — excels at specific tasks: translating languages, generating images, summarising documents, predicting outcomes. It does not understand the world the way a person does; it pattern-matches at scale.
General AI (AGI) — a system capable of any cognitive task a human can perform, and able to transfer skills flexibly between domains — does not yet exist. Key milestones on the road so far: 1950 Turing Test, 1997 Deep Blue defeats Kasparov at chess, 2012 ImageNet breakthrough using deep convolutional networks, 2017 Transformer architecture published by Google, 2022 ChatGPT's public release brings generative AI to the mainstream.
Training Data: the system is fed enormous quantities of examples — text, images, code, audio. The more varied and accurate the data, the better the model can later generalise.
Neural Network: billions of internal parameters are adjusted so the model learns statistical patterns in the data. This is the "learning" stage — slow and computationally expensive.
Predictions / Output: once trained, the model produces answers, images, or actions in response to new prompts — using the patterns it learned, not by reasoning the way a human does.
Multimodal language model; reads text, images, and audio
Reasoning-focused assistant; strong at analysis and writing
Integrated into Google products; very long context window
AI pair programmer; autocompletes and generates code
Image generation from text prompts
Video generation from text descriptions