Back to Home

Insights

How AI evolved

From recognizing handwritten digits to generating language, images, and action — a short tour of the breakthroughs that took AI from discriminative models to the generative era powering today's tools.

Tip: tap any milestone to see how it advanced the field.

DISCRIMINATIVEGENERATIVE AICNN1989LeNet-51999AlexNet2012Pixel-Level Gen.2014–2020GPT-12018Diffusion Models2020–presentGPT-2/32019–2020VLA2023–presentModern LLMs2022–present
Discriminative1989

CNN · Convolutional Neural Networks

Yann LeCun applied backpropagation to convolutional networks for reading handwritten digits. CNNs introduced learned feature detectors (filters) that scan an image — the foundation for nearly all modern computer vision.

The discriminative era: teaching machines to see

Early AI focused on discrimination — telling things apart. Given an input, predict a label. The convolutional neural network (CNN), pioneered in 1989, learned visual features automatically instead of relying on hand-crafted rules. LeNet-5 turned that idea into a practical system reading bank cheques, and in 2012 AlexNet shattered the ImageNet benchmark using GPUs — the spark that ignited the modern deep-learning boom.

These models were brilliant classifiers, but they didn't create anything. That was about to change.

The generative era: teaching machines to create

Starting with GANs in 2014, models learned to generate — first pixels, then language. GPT-1 (2018) introduced generative pre-training for text; scaling it produced GPT-2 and GPT-3, which could perform new tasks from a prompt alone. Diffusion models overtook GANs for images and video, and instruction-tuning turned raw models into the helpful assistants we now call modern LLMs. The frontier today — Vision-Language-Action (VLA) — fuses perception, language, and action so AI can operate in the physical world.

Tap through the milestones above to see how each step built on the last.

Where Node2 fits

The same modern LLMs at the end of this timeline now run small and private enough to deploy locally, on your own infrastructure. That's exactly what Node2 builds — AI-native finance and payroll tooling, custom local LLMs, and Micro AI agents that keep your data in Canada and out of the public cloud.

See what we build