TECHNOLOGY

Machines That Learn

~2012 CE

History · Contemporary · Holocene · Meghalayan

In the early 2010s, an old idea started working: neural networks with many layers, trained on oceans of data, began to see, hear and translate better than the hand-built programs before them. The approach then crossed from perception into science, predicting protein structures to atomic accuracy in 2021, more than fifty years after the problem was first posed, and into everyone's pocket as systems that converse, write and code. This timeline's last invention is a machine that improves with experience.

Why it matters

Every earlier tool on this timeline extended muscle, senses or memory; this one reaches towards the faculty that made all the others. Its trajectory is undecided, which is precisely the point: like fire, writing and the atom before it, learning machines enter the story as pure capability, and what the species does with them becomes the next entry. The timeline does not end; it hands over the pen.

Dating & uncertainty

Neural networks are old; the anchor marks the early 2010s, when deep learning, many-layered networks trained by backpropagation on big data and fast chips, began winning at vision and speech, on the field's own account. A decade later the same approach brought atomic accuracy to protein structure prediction, an open problem for more than fifty years, and conversational systems reached the general public.

Sources

1. LeCun, Bengio & Hinton (2015), Nature · Deep learningdoi:10.1038/nature14539
2. Jumper et al. (2021), Nature · Highly accurate protein structure prediction with AlphaFolddoi:10.1038/s41586-021-03819-2
3. The Nobel Prize in Chemistry 2024, awarded in part to Demis Hassabis and John M. Jumper for protein structure prediction. Institutional source: NobelPrize.org press release, accessed 2026-08-30.

Our species’ history as one day

23:59:563.8 s before midnight

if the 315,000 years of Homo sapiens were compressed into a single day