I spent a bit of time recently pulling together 50 of the most important AI and ML papers from the major labs and academic groups. Wanted to see how they actually linked up over the years instead of treating them as random standalone breakthroughs.
I fired up some agents for about an hour to go through the data and they put together this chart, sorted by year, field and longer term impact.
Go through it in sequence and the whole path is there. AI came together in distinct layers.
First phase, roughly 2014 to 2017, mostly locking down the foundations. Computer vision took some major steps. Reinforcement learning began in a serious way. Representation learning improved. Then the Transformer paper landed and reset the bar for everyone.
After that everythings toward scaling.
Labs realized they could forecast performance gains fairly reliably just by scaling up compute, data volume & parameters. It was right around then that LLMs started merging all those separate research threads together. GPT architecture, BERT pretraining, instruction tuning, alignment, multimodal inputs, open weights, low-rank adaptation. etc etc.