Software 2.0

Software whose behavior is determined by the learned weightsparameter, n.: One of the adjustable numbers inside an AI model that are tuned during training. Large language models have billions of them, and together they store what the model has learned. of a neural networkneural network, n.: A kind of computer program that learns a task from examples by tuning thousands to billions of connections between simple number-crunching units, an arrangement loosely inspired by nerve cells in the brain. rather than by hand-written instructions. In this framing, developers stop writing the program directly; they curate datasetsdataset, n.: An organized collection of examples gathered for a purpose — such as photos with their captions, or questions with answers — used to teach or test an AI system., choose an architecture and objective, and let optimizationgradient descent, n.: A method for training models by repeatedly nudging every setting a small step in whichever direction most reduces the error, like a hiker who cannot see the valley but keeps stepping wherever the ground tilts down. search the space of possible programs, so that the resulting 'code' is millions of numbers no one reads. Andrej Karpathy contrasted it with Software 1.0 — explicit code in languages such as Python or C++ — and predicted it would progressively absorb tasks such as vision, speech recognition and translation. He listed its advantages (uniform computation, fixed run time and memory, easy porting to specialized chips) and its costs (opacity, silent failures, inherited biases).

Our lane detector went Software 2.0 years ago; the hand-tuned rules are gone.