Deployment as research treats real-world use as an experimental pressure test. Scale, engineering constraints, and user demand reveal improvements and insights that isolated novelty may miss. section 4.4.5
The practice appears through concrete engineering and organizational choices:
- Baidu’s Deep Speech 2 improved on Deep Speech 1 by scaling the same end-to-end CTC-trained RNN with deeper networks, sequence-wise batch normalization, SortaGrad, synchronous multi-GPU training, and low-latency serving techniques. section 4.4.5
- OpenAI’s GPT work illustrates product-oriented research: scaling the Transformer into deployable systems rather than treating architectural novelty as the primary measure of progress. section 4.4.5
From a good idea to a working system
Deep Speech 1 established a strong end-to-end speech-recognition approach; Deep Speech 2 demonstrated how data, infrastructure, production constraints, and disciplined scaling could make it substantially better. section 4.4.5
