Software 2.0: when learned models replace hand-written rules
Andrej Karpathy describes Software 2.0 and Tesla's data engine, George Hotz explains comma's learned driving policy, and the DeepSeek analysts describe reinforcement learning's aha moment. Each shows systems improving when engineers stop writing rules and let data and training do the work.
What is Software 2.0? Show answer
Andrej Karpathy's name for code written in a neural network's weights, where engineers supply data and objectives and training does the rest.
Why did networks take over Tesla's Autopilot code, according to Karpathy? Show answer
Humans are not good at writing software for problems like perception, where the difficulty is a long tail of rare cases.
What is the data engine? Show answer
A loop: deploy the model, find the rare cases where it struggles, reconstruct the truth, add them to training and retrain.
What three qualities does Karpathy say training data needs? Show answer
It must be very large, accurate and diverse.
Why does Karpathy call extra sensors liabilities? Show answer
Each brings supply chains, firmware, calibration and fusion code, and different versions add noise to the data.
What reward does George Hotz use for driving? Show answer
Whether a human would take over, a disengagement, instead of how human-like the driving is.
What was the aha moment in DeepSeek's reasoning model? Show answer
Reinforcement learning on verifiable questions, with no human preference data, made behaviors like checking its work emerge on their own.
Why did math and code improve first with trial and error? Show answer
Their answers can be checked automatically, so a model can try, fail and improve without a person in the loop.
Why should you not start reinforcement learning from scratch? Show answer
Karpathy's early project failed because a randomly initialized agent never stumbled onto a correct result. A pretrained model gives a starting point.
Where did the engineering go in Software 2.0? Show answer
To data quality, reward design, verification and the infrastructure that makes training cheaper and bigger.
Why do multi-step agent tasks fail, in Dylan Patel's comparison? Show answer
Like manufacturing yields, high reliability at each step still compounds into failure over many steps.
What does Karpathy say direct use of a system gives you? Show answer
An intuition for its failures that numbers and plots hide.