There’s something magical about Recurrent Neural Networks
Andrej Karpathy, 21 May 2015 (Andrej Karpathy blog). Karpathy, then a PhD student at Stanford, opened his blog post on character-level text generation with recurrent networks in May 2015. With the post he released code on Github for training character-level language models.
The Unreasonable Effectiveness of Recurrent Neural Networks
More from Andrej Karpathy · AI and science · AI and hype · Before ChatGPT
Who else?
What it refers to: Recurrent Neural Networks, which in his first image-captioning model generated nearly sensible image descriptions within a few dozen minutes of training.
Other voices
- “Whatever you see there, I want more of it” Alexander Mordvintsev et al.,
- “We don't understand how neural networks work” Leo Gao,
- “Machine learning has become alchemy” Ali Rahimi,
- “I challenge the claim that next-token prediction cannot surpass human performance” Ilya Sutskever, Builder,
- “Pointillistic masses of blurry memory” Gary Marcus, Researcher,










