AI Got Smarter When Engineers Stopped Trying to Be Clever

The best way to make artificial intelligence smarter turned out to be simpler and less clever than experts expected. Instead of carefully teaching a system the rules of a field, researchers got better results by feeding it much more data and much more computing power, and letting it find the patterns itself.

Eric Landau, co-founder of Encord, which builds the data-handling tools behind robots and self-driving cars, says he lived through that shift. He started in particle physics, the study of the tiniest bits of matter, where he and colleagues filtered their measurements using what they already knew about how particles move and how much push they carry. He then spent 10 years as a quant, a trader who uses math to bet on markets, where the work was to figure out which economic factors moved prices and then build those ideas into a trading model.

Both jobs relied on what AI researchers call feature engineering, meaning experts decide in advance which details matter and hand those clues to the machine. What researchers learned, in an observation known as the bitter lesson, is that this careful handcrafting usually loses. A larger, simpler system, given a giant pile of raw data and enough computing power to chew through it, generally learns better on its own.

That lesson shaped where Landau started his company. Encord began with vision, meaning photos and video, because those were the toughest and most data-packed inputs to handle first. The bet was that AI systems would naturally become multimodal, a technical term for taking in many senses at once, because, in his view, people are multimodal and the best decisions come from accepting as many sensory inputs as possible.

What happened next, in his telling, is that the market came in his direction. The biggest application for that multisense AI turned out to be physical AI, machines that act in the real world. Encord now works heavily in robotics, autonomous vehicles, logistics, and manufacturing.