The hard part was never the model
The article argues that the real challenge in AI development is not building the model itself, but rather managing the surrounding infrastructure, data pipelines, deployment, and organizational complexity that make the model useful in practice.
Background
- The article argues that in AI, the real cost and difficulty have shifted from training the model (which is becoming commoditized) to inference — actually running the model on user requests at scale.
- The author contrasts the "reasoning" era (OpenAI's o1/o3, DeepSeek R1) with the prior "scaling" era, arguing these new models require drastically more compute per query, making inference costs the bottleneck.
- Key context: DeepSeek (China) recently released a competitive reasoning model at a fraction of the cost, intensifying the focus on inference economics. The piece is written for an audience familiar with AI industry dynamics but wanting to understand where the money and engineering effort are really going now.