Andrej Karpathy’s char-rnn made recurrent text generation tangible: train on essays, then generate new prose one character at a time. His Paul Graham generator used about 1 MB of concatenated startup essays—roughly one million characters. section 4.1
The model was a two-layer stacked LSTM with 512 hidden units per layer, about 3.5 million parameters, and dropout set to 0.5. At generation time, it sampled each next character from a probability distribution. Lower sampling temperatures made the output more predictable and repetitive; higher ones made it more creative but less coherent. section 4.1
Startup advice, almost
The generated prose mentions investors, fundraising, and acquirers, yet its sentences fail to form coherent advice. Familiar vocabulary is not the same as a sound argument. section 4.1
